Patent drafting with artificial intelligence

Abstract

A method of generating a patent application. One or more patent claims are received 100, comprising at least one independent claim and optionally one or more dependent claims. Novelty and inventive step of the patent claims is determined 110. The patent claims are modified 120 to obtain patentable patent claims. A patent application is generated 130 from the modified patent claims. One or more artificial intelligence models or artificial intelligence agents may be used. Preferably, novelty is determined with retrieval augmented generation. Inventive step may be determined according to prompts derived from patent laws and regulations, e.g. examination guidelines. A patent examination process may be simulated, using a generative adversarial network based model, and used to evaluate claim sets. [Figure 1]

Description

TECHNICAL FIELD The present application generally relates to methods of generating patent applications with artificial intelligence, and more particularly to automated systems and processes for drafting and organizing patent contents. BACKGROUND ART The patent document US20170075877 describes methods and systems of handling patent claims, before the advent of generative artificial intelligence, but this approach presents limitations. The patent document US20170173262 is an example of a patent application drafted in a combinatorial way, with a high number of keywords. These approaches present limitations. SUMMARY Described methods and systems can enhance the quality of drafting of a patent application by internalizing examination at the drafting stage, before even receiving an official search report from patent offices. These approaches present various advantages and / or technical effects. For example, they can save time and resources for the applicant, possibly leading to stronger and more defensible patent rights upon grant. They can allow for flexible adjustments to patent claims based on novelty and inventive step assessments. They can enable the reuse of existing claims or elements from previous applications to enhance the quality and breadth of the new patent application. More robust and comprehensive patent applications can be obtained. The scope of the independent claim can be broadened. Fallback positions that can be useful during prosecution or litigation can be multiplied. A claim set can be tailored to target specific competitors or industry practices. Consistency of drafting can be improved. Contents in the detailed description can be densened. Sufficiency of disclosure may be quantified, characterized or otherwise be tested. Past fallbacks (for example present in past or published patent applications) can be reused in combination. By integrating elements from distant technical fields, the described method and systems can promote interdisciplinary innovation, potentially leading to breakthrough inventions. Drafting can be more efficient and systematic, which can be particularly beneficial for entities managing large patent portfolios. Transposition mechanisms can enable the adaptation of an invention to different environments or uses, broadening the scope of the patent claims and enhancing the versatility of the protected technology. The quality of patent claims can be improved by incorporating the insights of advanced Al models, which can analyze vast amounts of patent data to identify strong, enforceable claims. Described methods and systems can ensure that patent claims remain at the forefront of technological innovation by incorporating terminology that reflects emerging trends in scientific and patent literature. Continuity with existing patent portfolios can be improved. User interface options can enable customization of patent claims and efficient exploration of the Library of Babel. BRIEF DESCRIPTION OF DRAWINGS The present disclosure is illustrated by way of example and not limited in the accompanying figures in which reference numerals indicate similar elements. Embodiments of the application will now be described with reference to the attached drawing: Figure 1 shows a general overview of some embodiments. DETAILED DESCRIPTION Figure 1 shows a general overview of some embodiments. There is disclosed a method of generating a patent application, comprising: receiving one or more patent claims 100; determining patentability 110 e.g. novelty and inventive step of said one or more patent claims; modifying 120 said one or more patent claims accordingly with one or more artificial intelligence models and / or rule-based systems; generating 130 a patent application from said modified one or more patent claims with one or more artificial intelligence models. In some developments, a plurality of claim trees (e.g. variants of claim trees) can be generated 140 as well as one or more patent applications computed from said claim trees. Optionally, one or more patent applications can be published 150 (e.g. via anticipated publication, or electronic publishing for example trusted and / or trustless timestamping). Claims can be received or generated from a document with one or more artificial intelligence models. The document can comprise text, images, videos or software code. Artificial intelligence models enable to simulate ("simulated patent examiner”) or otherwise predict the examination of a patent application and adjust its patent claims so as to obtain patentable claims. In a development, novelty is determined with Retrieval Augmented Generation. Retrieval-Augmented Generation (RAG) is an artificial intelligence framework that enhances the capabilities of large language models (LLMs) by integrating an external information retrieval mechanism into the generative process. Instead of relying solely on the static knowledge encoded during their initial training, RAG systems dynamically fetch relevant, up-to-date, and domain-specific information from external sources—such as databases, document repositories, or web content— at inference time, and use this information to ground and enrich the generated responses. In a development, inventive step is determined according to one or more prompts derived from patent laws and regulations, for example guidelines for examination or any other methodological document detailing various substeps to encode algorithmically the problem-solution approach. Jurisdictions for example can be Europe, China, Japan, Korea and the United States. In a development, novelty and / or inventive step are determined on a configurable content window, said content window defining a number of words, such as 50 words, said content window being applicable to patent claims or to the detailed description. In a development, a criteria of extractability of parts of the detailed description for amending the patent claims during prosecution is determined and optionally displayed. In a development, the step of modifying said one or more patent claims is performed automatically without user intervention. In a development, the step of modifying said one or more patent claims is guided in a user interface. In a development, user feedback is used to iteratively refine the novelty and inventive step determination and to modify the claim tree accordingly. Various developments are now described. In a development, a collaboration feature allows multiple users to work on determining and modifying the patent claims, in real-time or asynchronously. In a development, the user is provided with an option to prioritize the modification of certain parts of the claim tree in view of novelty or inventive step assessments. In a development, a report is generated detailing the rationale for each modification made to the patent claims. In a development, an alert system notifies the user of any potential conflicts or overlaps with existing patent claims. In a development, predictive analytics are used to predict or simulate the grant probability of the patent application based on historical data. In a development, a machine learning model is trained to improve relevancy or accuracy in determining novelty and inventive step based on past patent application outcomes. In a development, the method further comprises generating a visual representation of novelty and inventive step analysis results to aid in decision-making. In a development, a user feedback loop is integrated to continuously enhance the methods for determining novelty and inventive step based on user input and external data. In a development, a blockchain ledger is used to securely record modification operations taken during the patent application generation process. In a development, the system provides real-time analytics to evaluate detectability of patent claims. In a development, the generated patent application includes an automated compliance check against various international patent office submission formats and requirements. In a development, the prior art database is analyzed so as to identify one or more sweat spots, wherein a sweat spot is a technical domain which is less encumbered by documents and / or third party rights, as assessed by configurable thresholds, and wherein content generation is ruled depending on these identified sweat spots. In a development, the inventive step assessment involves comparative analysis with a dynamically updated database of technological advancements, or various heuristics for innovation such as TriZ. In a development, claim modifications are tracked through a version control system to provide a comprehensive history of changes and wherein the various versions of claims are stored and further leveraged to generate associated patent contents, leveraging the generation of claims variants, wherein a claim variant for example can vary time, space, materials, or physical properties associated with claims features. In a development, the method further comprises simulating examination from major patent offices. In a development, the method further comprises suggesting alternative claim formulations to enhance patentability, and / or clarity, and / or sufficiency of disclosure. In a development, the user interface supports voice command functionalities for input and navigation. In a development, the method further comprises leveraging a knowledge-sharing platform where users can access and contribute to a repository of best practices and technological insights. In a development, user engagement levels and interaction patterns are analyzed to enhance user experience and improvements brought to patent claims. In a development, the prior art database incorporates not only patents but also non-patent literature such as academic publications. In a development, a secure blockchain ledger ensures both transparency and immutability of the application generation process. In a development, domain-specific Al models are trained to more effectively handle technical language peculiarities inherent in diverse fields of innovation. Various developments are now described. In a development, the Retrieval Augmented Generation system supports feedback loops to refine retrieval algorithms based on user input and search outcomes. In a development, the analysis of inventive step includes sector-specific heuristic models. In a development, advisory prompts adapt based on historical examination outcomes from various patent offices. In a development, the content window size adjusts dynamically based on the complexity of the claims being analyzed. In a development, boundary conditions for the content window are set by artificial intelligence to optimize evaluation results. In a development, the method further comprises visualizing how modifications within the content window affect other claims. In a development, interface tools allow drag-and-drop functionality to select extractable parts of the description. In a development, the method further comprises a real-time extractability scoring system. In a development, user preferences guide the selection of automatic modifications. In a development, modifications follow pre-set templates tailored to user needs. In a development, the interface provides real-time suggestions for user-guided modifications. In a development, the method further comprises adaptive learning to enhance the interface based on user behavior patterns. In a development, feedback cycles are scheduled at predefined intervals to maintain up-to-date enhancements. In a development, version control tracks changes made by each collaborating user. In a development, a user receives alerts if conflicting priorities are detected. In a development, visualization formats include heat maps and trend graphs. In a development, feedback is crowd-sourced from a community of registered users for broader insights. In a development, the method further comprises a rewards system to incentivize user feedback contributions. In a development, configuration thresholds for identifying sweat spots can be adjusted by users based on strategic interests. In a development, examination simulations incorporate feedback from patent law experts for increased accuracy. In a development, suggested claim formulations include suggested logic chaining for improved patent scope articulation. In a development, the knowledge-sharing platform includes a voting mechanism to highlight valuable insights. In a development, user pattern analysis feeds into an Al model to enhance suggestions for future patent claims. In a development, the method further comprises real-time feedback tools to instantly capture user experiences and issues. There is disclosed a method of generating a patent application, comprising: receiving one or more patent claims, comprising at least one independent claim and optionally one or more dependent patent claims; determining novelty and inventive step of said one or more patent claims; modifying said one or more patent claims accordingly so that to obtain patentable patent claims; generating a patent application from said modified one or more patent claims. In a development, the step of modifying 120 said one or more patent claims comprises one or more of the steps comprising: combining one or more dependent claims if available with the at least one independent patent claim; modifying said one or more patent claims by generating one or more intermediate generalizations; recombining parts thereof with text chunks extracted from the detailed description if available ; retrieving previously published patent applications or granted patents of the assignee or applicant or inventor associated with the one or more patent claims; recombining them with dependent claims or text chunks with patent applications or granted patents of adjacent patent classification classes; recombining them with dependent claims or text chunks of patent applications or granted patents identified as prior documents; recombining them with dependent claims or text chunks of patent applications or granted patents retrieved from distant technical fields; recombining them with predefined invention patterns, wherein an invention pattern is a noun, an adjective, a multi-word such as computing device, or a text chunk ; using reasoning by analogy, and / or transposition mechanisms; recombining with textual content extracted from the generation of synthetic patents from said one or more patent claims, or variants thereof; prompting one or more artificial intelligence models to get patentable patent claims; replacing and / or inserting one or more words associated with weak signals in the scientific or patent literature; inserting sedimentary templates, wherein a sedimentary template restitutes the gist of one or more previous patent applications. It is not mandatory to have all these operations (preceding steps). As described, « one or more » of these operations can be implemented (i.e. one, two, or more, that is a combination). In a development, novelty is determined with Retrieval Augmented Generation. Retrieval-Augmented Generation (RAG) is an artificial intelligence framework that enhances the capabilities of large language models (LLMs) by integrating an external information retrieval mechanism into the generative process. Instead of relying solely on the static knowledge encoded during their initial training, RAG systems dynamically fetch relevant, up-to-date, and domain-specific information from external sources—such as databases, document repositories, or web content— at inference time, and use this information to ground and enrich the generated responses. In a development, inventive step is determined according to one or more prompts derived from patent laws and regulations, for example guidelines for examination or any other methodological document detailing various sub steps to encode algorithmically inventive step, e.g. the problem-solution approach. Jurisdictions for example can be Europe, China, Japan, Korea and the United States. In addition to novelty and inventive step (practical patentability), other aspects can be assessed and lead to amendments: theoretical patentability (e.g. analyzing patentability exemptions by patent laws), clarity and / or enablement and / or sufficiency of disclosure of said one or more patent claims and / or associated detailed descriptions if applicable. In a development, one or more artificial intelligence models are used for any one of the preceding steps, for example generation operations and / or text modifications, and wherein an artificial intelligence model is one or more of a Large Language Model; a Small Language Model; a multimodal model; a machine learning model such as linear regression, logistic regression, decision tree, random forest, K-Nearest Neighbor, Naive Bayes, or a Support Vector Machine; a deep learning model such as a Convolutional Neural Network, a recurrent Neural Networks, a Long Short-Term Memory Networks, or a Transformer; a reinforcement Learning Model such as a Q-Learning model, a deep Q Networks, or a Policy Gradient Method; a Learning Model such as a K-means Clustering, Principal Component Analysis, or an autoencoder, or a specialized Al Model such as a Generative Adversarial Network, a variational autoencoder, a Graph Neural Network or a combination thereof. In a development, an artificial intelligence model is autonomous, wherein the autonomous property comprises one or more of the mechanisms including: self-healing code, concurrent loosely coupled Al models, concurrent engineering, reinforcement learning, dynamic planning algorithms, multimodal Al integration, real-time decisionmaking, adaptive neural networks, goal-setting autonomy, predictive maintenance, explainable Al outputs, edge Al processing, autonomous task orchestration, generative Al workflows, context-aware recommendations, selfcorrecting algorithms, intelligent coding assistants, automated workflow optimization, proactive software agents, neuromorphic computing, privacy-preserving learning, ethical Al frameworks, regulatory compliance modules, bias detection algorithms, data anonymization techniques, secure data processing, real-time anomaly detection, autonomous vehicle navigation, robotic process automation, natural language understanding, computer vision integration, loT automation protocols, smart contract integration, decentralized Al networks, continuous learning systems, fault-tolerant architectures, scalable Al responses, behavior analytics, object detection algorithms, autonomous tracking systems, energy-efficient Al models, quantum Al optimization, human-AI collaboration, observability-driven development, test-driven Al development, agentic Al systems, data lakehouse integration, mechanistic interpretability, or personalized Al interfaces. In a development, one or more artificial intelligence agents are used for any one of the preceding steps (in addition and / or in substitution to artificial intelligence models). An Al agent is a software entity capable of perceiving its environment, reasoning, making decisions, and executing actions autonomously to achieve defined objectives The relationship between Al models and Al agents is foundational and hierarchical: Al models are the analytical engines that power the reasoning, perception, and prediction capabilities of Al agents, while Al agents are autonomous systems that leverage these models to interact with their environment, make decisions, and execute actions to achieve specific goals A network of Al agents can be orchestrated so as to amend patent claims, for example in reaction to patentability diagnostics. One or more intelligent agents can be deployed within a dynamic, scalable network of Al systems, orchestrated to autonomously amend patent claims in real-time based on advanced patentability diagnostics. These Al agents, whether loosely coupled or tightly integrated, can leverage next-generation distributed architectures to enhance adaptability and efficiency. For instance, one Al agent, powered by neural-symbolic reasoning models, can evaluate and optimize sufficiency of disclosure by performing virtual reduction-to-practice simulations, ensuring robust enablement. Another agent, utilizing large-scale language transformers, enhances claim clarity by disambiguating complex or vague terminology, such as 'cloud computing,' through context-aware semantic refinement. A third agent, for example employing quantum-inspired algorithms, can assess theoretical patentability, novelty, and inventive step, rapidly identifying potential exemptions or overlaps with prior art using real-time knowledge graph traversal. These agents can be organized into flexible configurations, such as self-organizing peer-to-peer networks, hierarchically layered cognitive architectures, or master-slave frameworks with adaptive feedback loops, enabling scalable multi-agent collaboration. Advanced implementations may incorporate redundancy protocols (e.g., replication or triplication) for fault tolerance, concurrent engineering paradigms for parallel processing, or federated learning mechanisms to ensure privacy-preserving collaboration across distributed systems. Suggestions generated by these Al models are further refined through explainable Al frameworks, providing transparent, human-interpretable rationales for amendments, and are continuously improved via reinforcement learning from human feedback (RLHF) and evolutionary optimization techniques. These advancements collectively can lead or represent next-generation patent drafting ecosystems. Al agents are networked, or regulated according to dynamic schemes. In some embodiments, regulation of networked Al agents can be closed-loop (no man in the loop) or open-loop (with user intervention, e.g. at least confirmation). Regulation can comprise aspects comprising one or more of: "control loops" such as open-loop and closed-loop control, feedback control systems, retroaction loops, feedforward mechanisms, logic control, control system, commands and control, regulation, continuous control, intermittent control, discrete controls, sequential and combinational logic, software logic, timing sequence, dynamical systems, linear or nonlinear feedback systems, disturbance rejection, stabilized or unstable processes, sensitivity variations, reference tracking performance, rectification of random fluctuations, simultaneous or parallel accesses, logic control systems, cybernetics, distributed control system, guidance, navigation, and control, hierarchical control system, motion control, networked control system, process optimization, "automatic control system" or other "control engineering" aspects. In a development, one or more artificial intelligence agents are networked, cooperatively and / or competitively, or otherwise regulated according to dynamic and / or adaptive schemes to optimize patent claim drafting and / or amendments. The organization of artificial intelligence agents for patent drafting and amendment can be implemented within one or more interconnected networks or can be governed by sophisticated regulation schemes, enabling autonomous, precise, and scalable claim optimization. In a development, control loops and feedback mechanisms can be employed to regulate agent behavior. Al agents can be organized within control loops to ensure precise amendments to patent claims in response to patentability diagnostics, a) closed-loop control systems: agents can utilize feedback control, continuously monitoring patentability diagnostics, such as novelty, clarity, or inventive step, and dynamically adjusting claim amendments based on real-time performance metrics. For example, an agent tasked with clarity assessment can use retroaction loops to iteratively refine ambiguous terms, such as "cloud computing,” until a predefined clarity threshold is achieved, with fallback mechanisms that can revert to broader claim language if refinement fails, b) open-loop control: certain agents can execute predefined tasks without real-time feedback, such as generating initial claim drafts based on standardized templates, relying on feedforward mechanisms to preempt common patentability objections, with fallback that can involve manual review if templates are inadequate, c) linear and nonlinear feedback systems: for complex tasks, such as assessing sufficiency of disclosure, agents can use nonlinear feedback systems to model intricate relationships between claim elements, adapting to sensitivity variations in patent examiner interpretations, with fallback that can involve simplified linear models if nonlinear convergence is unstable, d) disturbance rejection: agents can incorporate robust control strategies to mitigate external disturbances, such as inconsistent prior art databases or evolving legal standards, ensuring stable claim optimization, with fallback that can involve conservative claim narrowing if disturbances persist. In a development, hierarchical and distributed control architectures can be utilized to balance autonomy and coordination, a) hierarchical control systems: agents can be arranged in layered architectures, where higher-level supervisory agents can issue commands and control to lower-level agents. For instance, a supervisory agent overseeing novelty assessments can delegate specific prior art searches to subordinate agents, with reference tracking performance that can ensure alignment with patentability objectives, and fallback that can involve broader prior art searches if specific searches yield insufficient results, b) distributed control systems: agents can operate as equal peers within a networked control system, collaborating via peer-to-peer communication protocols to enable simultaneous or parallel access to shared resources, such as patent databases, facilitating rapid, decentralized claim amendments, with fallback that can involve centralized coordination if peer-to-peer communication fails, c) cybernetic systems: drawing from cybernetics, agents can self-regulate through guidance, navigation, and control mechanisms, dynamically adjusting strategies based on environmental feedback, such as updates in patent law or examiner rejections, with fallback that can involve predefined rule-based strategies if dynamic adjustments are inconclusive, d) logic and sequential control: agents can leverage logic control systems and sequential control mechanisms to manage task execution, a) sequential and combinational logic: agents can follow timing sequences to execute tasks in a predefined order, such as assessing novelty before clarity, with combinational logic that can enable the integration of multiple diagnostic outputs (e.g., patentability and clarity scores) to propose holistic claim amendments, and fallback that can involve sequential processing if combinational logic yields conflicting results, b) software logic: implemented through software-defined control, agents can execute complex decision trees or rule-based systems to prioritize tasks, such as addressing critical patentability exemptions first, with fallback that can involve default prioritization if decision trees encounter ambiguity, c) discrete controls: for tasks requiring intermittent intervention, agents can use discrete control to activate specific functions, such as triggering clarity refinement only when ambiguity exceeds a threshold, with fallback that can involve continuous monitoring if discrete triggers are unreliable. In a development, dynamical systems and process optimization can be employed to adapt to evolving inputs and constraints, a) continuous control: agents can maintain continuous control over long-running processes, such as monitoring patentability across multiple jurisdictions, ensuring consistent performance through process optimization, with fallback that can involve jurisdiction-specific templates if cross-jurisdictional consistency is unattainable, b) intermittent control: for resource-intensive tasks, agents can use intermittent control to periodically assess and update claims, reducing computational overhead while maintaining accuracy, with fallback that can involve continuous control if intermittent updates miss critical changes, c) rectification of random fluctuations: agents can employ stabilized processes to correct random variations in input data, such as inconsistent prior art references, using motion control-inspired techniques to enhance reliability, with fallback that can involve manual data validation if automated rectification fails, d) automatic control systems: the network can operate as an automatic control system, autonomously managing workflows without human intervention, supported by control engineering principles such as optimal control theory, with fallback that can involve semi-automated workflows if full automation encounters errors. In a development, advanced control features can be integrated to enhance performance, a) networked control systems: agents can communicate over low-latency, high-bandwidth networks, enabling real-time coordination across distributed environments, such as cloud-based patent analysis platforms, with fallback that can involve local processing if network latency exceeds acceptable thresholds, b) motion control analogies: agents can dynamically adjust their amendment strategies to avoid obstacles, such as unpatentable subject matter, using predictive control models, with fallback that can involve conservative claim scoping if predictions are uncertain, c) federated control: To ensure data privacy, agents can operate within federated control frameworks, sharing only model updates rather than raw patent data, aligning with privacy-preserving Al paradigms, with fallback that can involve encrypted data sharing if federated frameworks are unavailable, d) reinforcement learning for control: agents can optimize control strategies through reinforcement learning, learning from past amendment successes to improve reference tracking performance and disturbance rejection, with fallback that can involve rule-based control if learning models fail to converge. In a development, scalability and redundancy can be incorporated to support large-scale patent portfolios, a) concurrent engineering: agents can operate in parallel, simultaneously addressing multiple claim sets, leveraging concurrent control to minimize latency, with fallback that can involve sequential processing if parallel operations overload system resources, b) redundancy protocols: inspired by fault-tolerant control, agents can implement replication or triplication to ensure reliability, with backup agents that can be activated in case of failures, and fallback that can involve single-agent operation if redundancy mechanisms are compromised, c) scalable control loops: the system can dynamically scale control loops to accommodate varying workloads, using adaptive control to maintain performance under high demand, with fallback that can involve fixed-loop configurations if dynamic scaling introduces instability. In a further development, adaptive fallback mechanisms can be integrated to enhance robustness, a) contextual reversion: if an advanced control mechanism fails to produce viable claim amendments, agents can revert to contextually appropriate fallback strategies, such as broader claim language or simplified control models, based on the specific failure mode, b) human-in-the-loop validation: for critical failures, agents can trigger human-in-the-loop validation, presenting proposed amendments for manual review, ensuring compliance with patentability requirements, c) dynamic reconfiguration: agents can dynamically reconfigure their network topology or control schemes in response to persistent failures, such as switching from distributed to hierarchical control, to restore system stability. In a development, adaptive fallback mechanisms can be integrated to enhance the robustness of Al-driven patent claim optimization systems, a) contextual reversion: if an advanced control mechanism fails to produce viable claim amendments, agents can autonomously revert to contextually appropriate fallback strategies, such as broader claim language or simplified control models, selected based on the specific failure mode, with evolutionary algorithms that can optimize reversion policies by iteratively refining fallback strategies to maximize patentability success, and fallback to predefined default strategies if evolutionary optimization does not converge, b) human-in-the-loop validation: for critical failures, such as amendments that risk patent rejection, agents can trigger human-in-the-loop validation, presenting proposed amendments for manual review to ensure compliance with patentability requirements, with meta-learning frameworks that can adapt validation prompts to specific human expertise, and fallback to fully automated processing if human input is unavailable, c) dynamic reconfiguration: agents can dynamically reconfigure their network topology or control schemes in response to persistent failures, such as switching from distributed to hierarchical control to restore system stability, leveraging Generative Adversarial Networks (GANs) that can simulate failure scenarios to train reconfiguration strategies, ensuring robust adaptation to unforeseen challenges, with fallback to static network configurations if dynamic reconfiguration introduces instability. In a method for Al-driven patent drafting, networked artificial intelligence (Al) agents can be orchestrated to optimize claim creation and amendment, leveraging genetic algorithms and Generative Adversarial Networks (GANs) for enhanced precision and adaptability. Genetic algorithms can evolve agent behaviors, network topologies, and amendment strategies by iteratively selecting and mutating configurations based on fitness metrics, such as claim novelty or examiner approval rates, with fallback to default configurations if optimization stalls. GANs can refine patent claims, where a generator agent crafts innovative claim language and a discriminator evaluates compliance with patentability standards, iteratively strengthening claims against prior art, with fallback to templatebased drafting if GAN convergence fails. Networked Al agents, operating in distributed or hierarchical systems, can collaborate via peer-to-peer protocols or layered supervision, with genetic algorithms optimizing communication and task allocation for real-time claim refinement. GANs can simulate examiner objections to train agents, ensuring robust claim resilience, while evolutionary mechanisms enhance network scalability and adaptability, with fallbacks to centralized control or manual review to ensure reliability in complex patent drafting scenarios. In a development, patent drafting uses one or more Generative Adversarial Networks. A Generative Adversarial Network (GAN) is a class of machine learning frameworks where two neural networks, a generator and a discriminator, are trained simultaneously in a competitive setting. The generator creates synthetic data (e.g., images, text) from random noise, while the discriminator evaluates whether the data is real (from the true dataset) or fake (produced by the generator). Through this adversarial process, the generator improves its ability to produce realistic data, and the discriminator becomes better at distinguishing real from fake. GANs are widely used for generating realistic images, audio, text, and other data types. GANs can enable the creation of synthetic data for training autonomous systems. GANs can generate diverse data types (e.g., images, text) for autonomous decisionmaking. Self-correcting algorithms: the adversarial training process inherently refines outputs. Context-aware recommendations: GANs can generate personalized content for autonomous systems. Incorporating GANs into Al patent drafting, particularly for autonomous Al, can strengthen claims by highlighting innovative applications, technical improvements, and broad applicability. Below are strategies and examples for leveraging GANs in patent drafting. Broadening Patent Scope with GAN Applications: include GANs as a mechanism for generating synthetic data to enhance autonomous Al capabilities, such as training models in data-scarce environments or simulating edge cases. This ties to sub-criteria like generative Al workflows and autonomous task orchestration. In a development, the autonomous Al module can employ a GAN to generate anonymized synthetic datasets for training loosely coupled Al models, ensuring privacy-preserving learning across decentralized networks. Incorporate GANs to generate balanced datasets that reduce bias, aligning with ethical Al frameworks and bias detection algorithms. This addresses regulatory concerns and enhances patent enforceability. Wherein the GAN produces diverse synthetic datasets to mitigate bias in training data, supporting ethical Al frameworks and regulatory compliance modules. Future-Proofing: Include forward-looking applications like quantum Al optimization or neuromorphic computing to position the patent for emerging trends. In a development, the method further comprises the step of generating a plurality of variants of one or more of the patent claims, or entire patent claim tree comprising modified independent patent claim and one or more dependent patent claims, for example by using one or more patent embeddings and / or applying one or more heuristics to transform the patent claims or patent claim tree. In a development, heuristics comprise one or more of: inserting definitions of one or more terms in one or more claims, inserting one or more synonyms in one or more claims, inserting one or more modifiers of terms in the patent claims, substituting one or more words by synonyms, hypernyms, hyponyms or a combination thereof, combining one or more claims which are not dependent from one another, rewriting of one or more claims by rephrasing with the same words but in a different order, rewriting one or more claims by paraphrasing with different words, injecting predefined texts, varying time or space or any other property associated with said one or more patent claims, rewriting of one or more patent claims by handling singular to plural forms, applying one or more prompts to one or more patent claims or parts thereof. In a development, the method further comprises the step of generating a patent application from one or more claims, patent claim tree or variants thereof, for example a title, a technical field, a background section, a problem statement, a summary, an abstract, a detailed description and one or more patent figures, wherein a detailed description is obtained by stacking variants of the claims, obtained by predefined or dynamic prompts and / or rulebased heuristics challenging said claims, for example by varying time, space or any other attribute or essential feature. Once claim trees are varied, synthetic patents can be generated from said claim trees. In particular, the generations can be performed automatically without human intervention In a development, the method further comprises the step of publishing one or more generated patent applications, electronically or via an anticipated publication before a patent office. There is disclosed a system comprising a processor and a memory, the memory storing instructions which, when executed by the processor, cause the system to perform the described steps. In a development, patent drafting or parts thereof is performed with one or more user interfaces, using one or more of 2D interface, 3D interface, virtual reality, augmented reality e.g. smart glasses, mixed reality, retinal display, holographic display, brain-machine interface, possibly with haptic feedback and / or eye-tracking. In a development, the method further comprises: representing one or more patent claims or one or more patent sections in an embedding space; determining a target point or an area within said embedding space that is sufficiently close to an initial patent claim and sufficiently distant from embeddings of prior art documents; wherein predefined or configurable thresholds are used, given one or more distance metrics; decoding said target point or multiple points within said area into textual content representing patentable patent claims and / or leading to synthetic patent applications; optionally using an automated examination simulation feedback loop to evaluate decoded claim sets. In natural language processing, various distance metrics can be used. These metrics comprise: Levenstein distance (measuring the minimum number of single-character edits to transform one string into another), Damerau-Levenshtein distance (extending Levenshtein by including transpositions of adjacent characters), Euclidean distance (measuring straight-line distance between vector points), Cosine distance (assessing dissimilarity via the angle between vectors), Manhattan distance (summing absolute differences across vector components), Jaccard distance (comparing set dissimilarity through intersection and union), Hamming distance (counting differing positions in equal-length sequences), and Word Mover's distance (calculating the cost to transform one word distribution into another using embeddings). These metrics can be used to quantify similarity or dissimilarity between text representations. In a development, determining the target point or area in embedding space comprises fine-tuning an embedding model using one or more patent claims or other patent sections to accurately represent patent underlying concepts and / or wherein decoding the target point or area into textual content involves fine-tuning a decoder model trained specifically on patent claims or other patent sections. In a development, the step of determining patentable embedding points can use a Generative Adversarial Network trained to reformulate one or more combination of patent claims into patentable content, based on publicly available databases of partially refused and rearranged patent claims. Optionally, one or more GANs can encode an interpretation of inventive step. In a development, the method further comprises automated selection of amendment strategies for patent claims based on objectives comprising one or more of: expanding the scope of protection, rendering claims undetectable by keyword matching; providing portfolio optimization advice relative to user or competitor patent portfolios. In a development, publishing comprises defensive publishing by storing and providing public access to a massive collection of generated patent claim variants and related patent texts in a highly optimized database system specifically designed to efficiently store, index, retrieve, and search large numbers of semantically and syntactically similar documents, optionally comprising one or more of the following: employing vector-based semantic retrieval techniques, such as approximate nearest neighbor searches; using deduplication and compression methods, such as MinHash or locality-sensitive hashing (LSH); applying hierarchical semantic clustering to group related variants; implementing delta encoding to store textual differences between variants; using adaptive indexing and learned index structures to optimize query performance; integrating content-addressable storage systems for deduplicated storage and efficient retrieval; applying secure timestamping, such as blockchain-based timestamping, to establish clear and immutable publication dates; incorporating real-time text similarity monitoring to promptly detect potential infringement or prior-art issues. In a development, the method further comprises generating patent claims using diffusion or auto-regressive text generation techniques that for example start with keywords predefined by a user, or generated by transforming an original textual or multimodal user submission; and generate patent claims by filling surrounding textual context ensuring semantic similarity to the original patent idea. In a development, the method further comprises using a GAN to identify a latent embedding space position or area where decoded patent claims are indistinguishable or minimally distinguishable from an original idea yet sufficiently distant from prior art to allow patentability. In a development, the method further comprises enabling visualization of latent embedding spaces in reduced dimensionality (e.g., 2D or 3D); and facilitating human-in-the-loop selection of optimal latent embedding points or areas to maximize claim breadth or scope of protection. In a development, visualization explicitly represents the breadth and scope of patent claims to assist user decisionmaking. In a development, the method further comprises training a GAN-based model to simulate a patent examination process by: transforming embeddings of one or more original patent claims or other patent sections relative to prior art embeddings; identifying embedding transformations required to achieve patentability; decoding transformed embeddings into patentable textual content. In a development, the method further comprises training time-aware patent embeddings capable of predicting the year associated with patent ideas, thus capturing and modeling the progression and evolution of inventions embodied into sets of patent claims. In a development, the method further comprises using clustering and classification techniques applied to publicly available datasets of amended patent claims to detect, name, and imitate patent prosecution strategies. In a development, scientific articles are handled instead of patents. The generation or rendering of patent texts described therein can indeed by substituted or replaced by the generation or rendering of scientific articles, for example by optimizing textual content to facilitate successful passage through peer-reviewing processes. In a development, the method further comprises dynamically optimizing embedding-space target points using multiobjective optimization algorithms balancing patentability, scope of protection, and technical accuracy based on patent examination simulations. In a development, the method further comprises using federated learning techniques to collaboratively fine-tune embedding models across distributed patent databases without compromising dataset confidentiality. In a development, the method further comprises incorporating explainability modules to provide textual or visual justifications for embedding transformations, claim amendments, and model predictions. In a development, the method further comprises automatically detecting and correcting semantic drift in patent claim embeddings during iterative drafting processes to maintain conceptual consistency and relevance. In a development, the method further comprises constructing hierarchical embeddings encoding relationships between patent claims and / or other patent section to ensure coherent claim structures and facilitate embedding-aware claim dependency management. In a development, the method further comprises integrating user-interactive constraint specification tools enabling direct user manipulation of embedding space visualizations to influence claim generation. In a development, the method further comprises providing patentability scoring of generated patent claims during drafting, informed by embedding proximity relative to dynamically updated prior-art embeddings. There is disclosed a method of generating a patent application, comprising: receiving one or more patent claims, comprising at least one independent claim and optionally one or more dependent patent claims; determining novelty and inventive step of said one or more patent claims; modifying said one or more patent claims accordingly; generating a patent application from said modified one or more patent claims. In a development, the step of modifying said one or more patent claims comprises combining one or more dependent claims with the at least one independent patent claim. There is disclosed a self recombination of the claim tree. For example, by combining one or more dependent claims with the independent claim, following dependency or not. The combination is evaluated for novelty. If not novel (and optionally inventive), new combinations can be tested. In a development, the features colliding with prior art are modified first (while the rest is left unchanged). One, two, three or more dependent claims can be tentatively combined. Dependent claims, if available, are the matter of choice for recombination because they represent what the inventor or drafter anticipated to be valuable and potentially bring patentability. A prototype prompt can read “Retrieve the dependent claims of the patent claim tree. Recombine one or maximum two dependent claims with the independent claim 1. Evaluate novelty for the combination. Handle each dependent claim this way. Provide the 5 best combinations, by providing 5 different modified claim sets, ordered by decreasing novelty. Hardcode all the other combinations without evaluating novelty (with 1 or 2 or 3 dependent claims at once) In a development, the step of modifying said one or more patent claims comprises generating one or more intermediate generalizations. For example, colliding features can be replaced combinatorially with one or more synonyms first, then with combination of hyponyms and hypernyms, then combination of synonyms, hypernyms and hyponyms. These synonyms, hypernyms and hyponyms can stem from Word2vec, dictionaries and Large Language Models. The use of synonyms is not very likely to change the collision status. The use of hypos and hypers can change the collision. To improve the situation, replacing several words at once can be recommended. This option makes the independent claim to "vibrate” slightly and enables to get focused and closed alternatives. Depending on the severity of the collision, an increasing number of words may have to be changed. A prototype prompt can read "For the features presenting a partial or total match with prior art, replace them with a combination of synonyms, hyponyms and hypernyms, in order to obtain a modified patent claim tree that is novel over the prior art. Start with synonyms and pursue with hypos and hypers. A combination of synonyms, hyponyms and hypernyms is called an intermediate generalization. Recombine in a combinatorial way so as to get the best novelty situation. The more severe the collision for a feature, the more amendments are required in general. Provide the 5 best combinations, by providing 5 different modified claim sets, ordered by decreasing novelty.”. In a development, the step of modifying said one or more patent claims comprises recombining parts thereof with text chunks extracted from the detailed description if available. If the one or more patent claims are associated with a detailed description (handcrafted patent application, or generated patent application), the latter detailed description can provide text parts or chunks which can be used to amend the one or more patent claims. Legal support criteria, for example in Europe, require strict copy-paste operations excluding excision (if detailed description is "a blue car”, extracting "a car” is not allowed). If no detailed description is available, a patent application can be generated from the one or more patent claims, for example by computing and stacking claims' variants via semantic amplifiers (variations over time, space, types of materials, Triz heuristics, etc) and text parts can be extracted thereof. A prototype prompt can read “Retrieve the detailed description associated with the patent claim tree. Analyze the detailed description and identify candidate fallback positions related to essential features of the patent claim tree, i.e. possible amendments to the patent claim tree so as to be novel. Amendments are strings of text, limited to mere copy and paste operations, without excising words from an expression. For example, if the detailed description is “a blue car”, then extracting “a car” without “blue” is notallowed. Only contiguous words can be extracted. Then combine these strings of text with the patent claim tree and evaluate novelty, possibly combinatorially with several amendments brought at once. The goal is to obtain a modified patent claim tree that is novel over the prior art. Hardcode the 5 best combinations, by providing 5 different modified claim sets." In a development, the step of modifying said one or more patent claims comprises retrieving previous published patent applications or granted patents of the assignee or applicant or inventor associated with the one or more patent claims. For example, the one or more patent claims can be recombined with the own works of the assignee: retrieve past applications by the same assignee, extract claims and detailed descriptions, and recombine; this option leverages the knowledge of the user and he can find something he knows. In such a development, there is no need for a prior art search. An assignee often drafts related applications and valuable technical fallbacks may well be present and worth combining. A prototype prompt can read “Retrieve 10 published past applications of the assignee. Analyze these documents and identify candidate fallback positions, i.e. possible amendments to the patent claim tree so as to be novel. Then combine these fallbacks in the patent claim tree and evaluate novelty. Hardcode the 5 best combinations, by providing 5 different claim sets. ” In a development, the step of modifying said one or more patent claims comprises recombining them with dependent claims or text chunks with patent applications or granted patents of adjacent patent classification classes. Patent classification can be IPC or CPC. Adjacent means closely related technical domains, for example as suggested in Espacenet classification. For example, the class G06F3 / 002 is related to G06F3 / 01 or G06F3 / 16 in the classification. It can be advised to follow these "natural” paths and retrieve associated contents, analyze them and recombine them with the one or more patent claims. Such an approach can "steal” from close neighbors and is more general than documents of the prior art. The presence of predefined links in classification nevertheless may render corresponding operations less inventive a priori, when compared to the combination of remote technical fields. A prototype prompt can read “Define one or more IPC classes associated with the current patent claim tree. In these IPC classes, perform a prior art search on the claimed features and retrieve at least 20 documents. Analyze these documents and identify candidate fallback positions, i.e. possible amendments to the patent claim tree so as to be novel. Then combine these fallbacks in the patent claim tree and evaluate novelty. Hardcode the 5 best combinations, by providing 5 different claim sets. ” In a development, the step of modifying said one or more patent claims comprises recombining them with dependent claims or text chunks of patent applications or granted patents identified as prior documents. This approach advantageously enable to cross-fertilize and leverage identified prior art by recombining contents with the one or more patent claims. Support criteria do not have to be followed, since alien contents are recombined, flexibility and deviations can occur. A prototype prompt can read “Retrieve the prior art documents of the search report. Analyze these documents and identify candidate fallback positions, i.e. possible amendments to the patent claim tree so as to be novel. Then combine these fallbacks in the patent claim tree and evaluate novelty. Hardcode the 5 best combinations, by providing 5 different claim sets. ” In a development, the step of modifying said one or more patent claims comprises recombining them with dependent claims or text chunks of patent applications or granted patents retrieved from distant technical fields. According to this approach, it can be advantageous to recombine dependent claims or chunks of remote IPC classes, for example as indicated by the user or predicted by machines ("transposition”). Thus, data or vocabulary silos can be broken beyond the "natural” links in the patent classification mentioned above. To identify these links, a prototype prompt can read "Consider patent applications and granted patents in remote IPC classes, that is that are not close to the technical domains associated with the claim tree as such and, beyond, that are not suggested from the IPC classification, wherein some IPC classes can be linked to one. From the contents of the identified documents, comprising claims or strings of texts extracted from the detailed descriptions of the patent applications and granted patents in remote IPC classes, extract valuable fallbacks and recombine with the pending claim tree”. Another prototype prompt can read "Look for analogous prior art, at least 20 documents, that belong to remote IPC classes, and recombine the pending claim tree with fallbacks identified by reasoning by analogy”. In a development, the step of modifying said one or more patent claims comprises recombining them with predefined invention patterns, wherein an invention pattern is a noun, an adjective, a multi-word such as computing device, or a text chunk. An invention pattern generally enhances patentability. Examples of invention patterns: “smart glasses”, “holographic”, “haptic”, “implemented in a blockchain”, “wherein the database is implemented in a blockchain”, “blockchain-enabled”, “medical”, “volume” for “volume qubit”, “obtained by additive manufacturing”, etc. These invention patterns can be extracted from the patent corpus, Triz or other methodologies. They also can represent weak signals in terminology, for example captured in the fresh web, blogs, scientific articles and the like. New terminology emerges progressively and captured early enough can be reinjected into the patent generation process. A prototype prompt can read “Consider the list of invention patterns, comprising one word or expression per line. Consider each word or expression and consider where to insert it in the patent claim tree. If the insertion at one or several places is meaningful from a technical perspective, evaluate the novelty of the modified patent claim tree. Otherwise consider the next word or expression. Provide 5 different claim sets with the highest novelty score. Further, consider recombining two or more of the more novel claim trees with one another”. In a development, the step of modifying said one or more patent claims comprises using reasoning by analogy, and / or transposition mechanisms. Douglas Hofstadter wrote interesting papers on analogies that have yet to be applied to patent matters. Reasoning by analogy is a cognitive and logical process in which conclusions about a new or unfamiliar situation (the target domain) are drawn by identifying and leveraging similarities with a known situation or example (the source domain). This approach dives deep in relationships amongst words and underlying concepts. A prototype prompt can read “Provide propositions of changes in the claim tree to leverage technical improvements described in other technical domains that are not closely related to the one corresponding the claims. For example, a mechanism in oil and gas extraction can be inspiring for dentistry. Use reasoning by analogy, along transposition. One way to proceed is: identify features of the claims, locate other patent claim trees with similar features, favor the ones that are not close in classification (but not totally alien, in such a case they may not be compatible). In order to find analogous prior art, consider explicitly ignoring closely related IPC or CPC classes, consider underlying concepts, consider underlying graphs (verbs as edges and nouns as vertices), and the like”. This approach thus consists in diving into underlying concepts, or underlying graphs. The main tasks involve identifying features of the claims, locating other claim trees with similar features, favoring recombination of the ones that are not close (but not totally alien, in such a case they may not be compatible). In addition or as an alternative, transpositions can be used. A transposition is a permutation that exchanges one or more elements of a set. In a development, the step of modifying said one or more patent claims comprises recombining with textual content extracted from the generation of synthetic patents from said one or more patent claims, or variants thereof. For a given claim tree, one or more artificial intelligence prompts can vary the claim tree and following, a selection of generated text parts can be recombined with the one or more patent claims. These prompts can challenge the patent claim tree by multiplying perspectives (time, space, etc), each giving options for adjustments. In a development, the step of modifying said one or more patent claims comprises prompting one or more artificial intelligence models to get patentable patent claims. A prototype prompt can read “Rewrite the claim tree so as to be novel of the prior art’ or better “insert 10 more dependent claims, leveraging multiple dependency, to multiply fallback positions". In a development, the step of modifying said one or more patent claims comprises replacing and / or inserting one or more words associated with weak signals in the scientific or patent literature. “An applicant is his own lexicographer is the principle in patents. An applicant can and should use terminology that is stable and widespread in the scientific community: this is a "technical embedding”, a technical teaching (not NLP term): if the applicant defines a new term properly, it is acceptable in claims (but can require to include the definition directly in claims, at least in a dependent). There are new emerging terms, or modifiers. For example “volume qubif, “holographic display. This is the "fresh matter” in patents, thus very valuable because at the forefront of patenting. R&D directors chase such new paradigms. A possible implementation is to browse and parse known blogs and journals, extract these trendy terms and use them in claims. New terms generally appear in detailed descriptions, then in dependent claims and finally in independent claims. Patent attorneys are generally cautious and proceed progressively. It is rare to place a new emerging term directly in independent claims. Words generally follow the general path: detailed >dependent >claim 1. When a word is found in published Claim 1, it can be the signal that it is "officially” entering the Claims arena, thus the importance of closely monitoring the new appearance of new words in patent applications, in order to amplify them in the generation corpus (ex: subscription services). If detected in recent blog posts or press articles, these new words advantageously can be reused in combination. The state of the art describes numerous ways to identify and extract "weak signals”. An efficient approach is to count frequency for non-empty words. In a development, the step of modifying said one or more patent claims comprises inserting sedimentary templates, wherein a sedimentary template restitutes the gist of one or more previous patent applications. Applicants file patent applications in sequence. Most of the time, it is not reminded of the essence of past applications in combination with new filings. Each past filing can be summarized in a text chunk, for example derived from the very payload of past claims, and further reinjected in new filings. These sedimentary templates can be absolute or universal (applicable to all patents) or customized for users (take their claim trees and summarize into multiple dependent claims), so that they can reinject the "gist” of their previous inventions into new ones, recursively. The previous inventions or applications can be of the same assignee, or be entirely new alien content. In a development, the method further comprises receiving user input. Modifications brought to the claim tree can be entirely automatic. In some developments, a human-loop can occur and user inputs or recommendations or suggestions can be captured and leveraged to guide machines. There is disclosed a method comprising: receiving a text in natural language; generating an independent patent claim from the received text. In a development, the step of receiving one or more modifications to the generated patent claim, originates from a user or is determined by a machine. In a development, the method further comprises the step of generating a plurality of dependent claims for the generated patent claim text, modified or not. In a development, the method further comprises the step of receiving one or more keywords as lexical directions for generating a plurality of dependent claims for the generated patent claim text, modified or not. In a development, the method further comprises the step of selecting one or more generated dependent patent claims. In a development, the method further comprises the step of generating 140 a plurality of variants of the patent claim tree comprising an independent patent claim and one or more dependent patent claims. In a development, the step of generating a plurality of variants of the patent claim tree comprises using one or more of patent embeddings and / or applying one or more heuristics to transform the patent claim tree. In a development, heuristics comprise injecting definitions of one or more terms in one or more claims, injecting one or more synonyms in one or more claims, substituting one or more words by synonyms, hypernyms, hyponyms or a combination thereof, combining one or more claims which are not dependent from one another, rewriting of one or more claims by rephrasing with the same words but in a different order, rewriting one or more claims by paraphrasing with different words, injecting predefined texts, varying time or space or any other property associated with said one or more patent claims, rewriting of one or more patent claims by handling singular to plural forms, applying one or more prompts to one or more patent claims or parts thereof. In a development, the method further comprises the step of selecting one or more generated patent claim trees. In a development, the method further comprises the step of generating a patent application based on the generated and / or selected patent claim or patent claim tree, for example a title, a technical field, a background section, a problem statement, a summary, an abstract, a detailed description and one or more patent figures. In a development, the method further comprises the step of publishing 150 one or more generated patent applications, electronically or via an anticipated publication via a patent filing. In a development, the text in natural language comprises a definition, a description of an invention, a technical problem, or a solution to a technical problem. In a development, the one or more modifications determined by a machine are generated based on a novelty analysis and / or inventive step of the generated independent patent claim against a database of prior art. In a development, selecting the one or more of the generated variants of the patent claim tree comprises evaluating the generated variants based on a patentability metric, the patentability metric comprising novelty, nonobviousness, or clarity. In a development, one or more artificial intelligence models are used for generation operations or text modifications. In a development, the text of the patent application is densified. Text densification refers to a set of techniques and strategies for increasing the information density of a written document by minimizing redundancy, compressing content, and optimizing layout so that the same or greater amount of essential information is conveyed in a more compact form; this includes replacing frequently repeated phrases with standardized acronyms or shorthand notations defined in a glossary, transforming verbose component descriptions into tables with condensed formatting, consolidating multiple figures into composite layouts with labeled sub-figures, rewriting sentences to remove redundancy and non-essential descriptors while preserving technical accuracy, embedding reference numerals or symbols directly into the text, using mathematical or symbolic notation to represent recurring concepts or processes, reformatting the document into multi-column layouts with minimal gutter space, reducing font size and line spacing within the limits of readability and regulatory requirements, minimizing paragraph and heading formatting, drafting claims and abstracts with maximum conciseness, embedding citations and prior art references inline rather than in separate sections, converting process descriptions into compact flowcharts, using software tools to automate the identification and compression of repetitive or verbose content, applying multiple densification methods in combination for maximum effect, shortening headings and subheadings, defining technical terms parenthetically at first use to eliminate the need for a separate glossary, summarizing background and prior art in concise paragraphs, replacing lengthy system descriptions with schematic diagrams, and eliminating redundant cross-references, all with the objective of reducing page count, enhancing clarity, and ensuring compliance with document constraints without sacrificing the completeness or precision of the technical disclosure. To handle substance in text densification, begin by rigorously prioritizing the core technical concepts, ensuring that every included detail directly supports the primary objectives or claims of the document; critically assess each section for its necessity, eliminating background information, context, or explanations that can be inferred by the intended expert audience or are already established in prior art; synthesize related ideas or findings into unified statements or concise summaries, avoiding fragmentation of arguments or excessive subdivision of topics; abstract procedural steps or technical workflows into generalized principles or representative examples, capturing the essence of the method or invention without exhaustive enumeration of every possible variant; distill complex arguments or justifications into logical chains that highlight only the most significant dependencies, assumptions, or results, omitting tangential discussions or secondary implications; employ domain-specific terminology with precision, trusting the reader's expertise to interpret nuanced language without redundant clarification; consolidate empirical data, experimental results, or case studies into summary tables, charts, or representative figures, referencing only the most illustrative or critical data points rather than exhaustive datasets; focus on demonstrating inventive steps, technical advantages, or problem-solution relationships with direct evidence or reasoning, streamlining or omitting narrative that does not directly advance the argument for novelty or inventive step; where necessary, use cross-references sparingly and only to reinforce key technical relationships or dependencies, avoiding recursive or circular referencing; in claims or technical disclosures, present only those embodiments, configurations, or parameters that are essential for defining the scope and utility of the invention, relegating optional or less critical variants to brief parenthetical notes or appendices; throughout, maintain a critical editorial stance, continuously questioning whether each substantive element is indispensable to the document's technical completeness, legal defensibility, or communicative clarity, and remove or condense any material that does not meet this threshold. There is disclosed a method for substantively transforming a patent detailed description to broaden the scope of protection, comprising identifying each essential feature or important term within the technical disclosure; for each such feature or term, inserting in parenthesis - or with graphical indentation or “e.g.” expression or the like -, a list of synonyms, hyponyms and hypernyms that are technically accurate and contextually appropriate so as to capture alternative expressions, more specific embodiments, and broader categories (for example, describing a "fastening mechanism e.g. clasp, buckle, Velcro closure, separable connector, attachment device, or securing links)"), applying this expansion consistently throughout the detailed description to ensure that the invention is described not only in its preferred forms but also in a spectrum of semantically related variants; selecting the parenthetical terms using domain-specific taxonomies, technical dictionaries, and ontologies to ensure relevance and enablement; reviewing the resulting expanded disclosure to confirm that the added terms do not introduce ambiguity or unsupported subject-matter; maintaining the original term as the primary reference while the parenthetical list serves to clarify and extend the semantic coverage of the invention, thereby enabling broader claim interpretation, supporting future claim amendments or divisional filings, and increasing the robustness of the patent against novelty and inventive step challenges. Other semantic expansion comprise techniques such as : with controlled vocabulary, replacing specific technical terms with broader hypernyms and adding parenthetical examples (e.g., "fastening mechanism (clasp, buckle, Velcro closure, separable connector, attachment device, securing means)"); automated technical summarization to condense background and description sections by preserving only sentences with novel technical features and eliminating redundant explanations; technical term consolidation by standardizing synonymous terms across the specification using domain-specific lexicons; modular claim architecture by structuring claims as nested technical primitives with dependent claims referencing atomic functional components; cross-document reference integration by embedding citations to related patents or literature directly within technical descriptions and eliminating separate background sections; algorithmic process abstraction by converting detailed method steps into mathematical formalisms with prose equivalents in appendices; exemplary embodiment compression by reducing variations to canonical examples with parameter ranges and using combinatorial reduction to maximize coverage; semantic field normalization through context-aware word sense disambiguation to ensure consistent interpretation of polysemous terms; claim language generalization by using functional descriptions; automated term substitution by deploying engines that replace verbose phrases with concise alternatives while maintaining legal equivalence and contextual meaning. In a development, patent drafting uses deep learning. Deep-learning can be used (this field for example comprises one or more of techniques comprising sparse coding, compressed sensing, connectionism, reservoir computing, liquid state machine, echo state network, unsupervised or supervised learning, classification, regression, clustering, dimensionality reduction, structured prediction, anomaly detection, neural nets, machine learning venues, artificial neural networks, deep neural network architectures, back propagation, convolutional neural networks, neural history compressor, recursive neural networks, long short term memory, deep belief networks, convolutional deep belief networks, deep Boltzmann machines, stacked (de-noising) auto-encoders, deep stacking networks, tensor deep stacking networks, spike-and-slab RBMs, compound hierarchical-deep models, deep coding networks, deep q-networks, networks with separate memory structures, LSTM-related differentiable memory structures, semantic hashing, neural Turing machines, memory networks, pointer networks, encoder-decoder networks, multilayer kernel machine, etc). In a development, patent drafting uses federated learning. Federated learning is a machine learning approach where a model is trained across multiple decentralized devices or servers holding local data, without exchanging the data itself. Each device trains a local model, and only model updates (e.g., gradients) are shared and aggregated to improve a global model, preserving data privacy. Federated learning aligns with autonomous Al by enabling privacy-preserving learning. It supports autonomous systems that learn from distributed data sources, such as loT devices or edge Al, while maintaining security and autonomy in decision-making without centralized data collection. In a development, patent drafting uses transfer learning. Transfer learning is a machine learning technique where a model trained on one task or dataset is reused or fine-tuned for a different but related task. It leverages knowledge learned from a large, general dataset (e.g., ImageNet for images) to improve performance on a smaller, specific dataset, reducing training time and data requirements. Transfer learning supports autonomous Al by enabling adaptive neural networks and continuous learning systems. It allows autonomous systems to quickly adapt pretrained models to new environments or tasks, enhancing their ability to operate independently and efficiently in dynamic settings. In a development, the logic embedded or implemented or used in one artificial intelligence model is one or more of classical logic, Boolean logic, fuzzy logic, formal logic, inductive logic, material logic, symbolic logic, deductive logic, dialectical logic, modal logic, propositional logic, first-order logic, higher-order logic, temporal logic, intuitionistic logic, non-monotonic logic, paraconsistent logic, many-valued logic, quantum logic, linear logic, epistemic logic, deontic logic, relevance logic, abductive logic, probabilistic logic, description logic, predicate logic, auto epistemic logic, defeasible logic or combinatory logic. In a development, one of the described methods further comprises the step of extracting features of said one or more patent claims for comparison with prior art databases. Various developments are now described. In a development, one of the described methods further comprises the step of determining embeddings of one or more features of said one or more features or patent claims for comparison with prior art databases. In a development, embeddings are dense or sparse vector embeddings. In a development, the comparison of said embeddings with prior art databases determines a pre-ranking of documents of the prior art databases. In a development, one of the described methods further comprises the step of storing intermediate versions during modifications of the one or more patent claim trees. Ephemeral versions during editing can be valuable and can be kept and potentially published. In a development, the step of modifying is performed iteratively or interactively using a graphical user interface to capture user lexical directions. In a development, changes brought to the claim tree are made so as to obfuscate the correspondence with one or more documents of the prior art, for example by using similarity distances. In a development, one of the described methods further comprises the step of crowdsourcing for modifying claims. In a development, one of the described methods further comprises the step of showing a plurality of combinatorially modified claim trees, selecting and / or discarding claim trees via automatic rules and / or user indicated selections. In a development, one of the described methods further comprises the step of using neural networks to anticipate future trends in patent language and / or capturing weak signals. In a development, one of the described methods further comprises the step of evaluating claim modifications based on computational complexity metrics. In a development, one of the described methods further comprises the step of using a genetic algorithm to evolve claims towards optimal configurations. In a development, one of the described methods further comprises the step of using boundary conditions to guard against over-generalizing claims. In a development, one of the described methods further comprises the step of using edge computing to decentralize claim modification processes. In a development, one of more texts comprise one or more of a confidential invention disclosure, a preprint unpublished scientific article or any type of text document comprising secret information. In a development, the inventive step is assessed by one or more Al prompts and / or one or more rules algorithmically encoding the European problemsolution approach. Other embodiments are now described. There is disclosed a method for handling "patent embeddings”. A "patent embedding” can consider an entire patent application or granted patent as a vector. A patent embedding refers to a numerical representation (typically a vector) of a patent document in a high-dimensional space, where the vector captures the semantic and contextual information of the patent's content. These embeddings can be generated using machine learning techniques, often natural language processing (NLP) models like transformers (e.g., BERT or Patent-BERT), to encode the patent's technical, legal, and conceptual meaning into a format suitable for computational tasks. Patent embeddings enable tasks like patent similarity search, classification (e.g., patent categorization), clustering, prior art search, or patent valuation by representing patents in a way that machines can process and compare. Embeddings can be created by feeding patent texts into a trained model, which outputs a fixed-length vector capturing semantic relationships. Advantageous applications comprise: finding similar patents based on cosine similarity or other distance metrics between embeddings, identifying trends in technology or innovation by clustering embeddings or suggesting relevant patents for inventors or examiners. Patent embeddings can allow capturing nuanced meanings (e.g., synonyms or technical jargon) and allow comparison of patents across languages or domains, unlike traditional keyword-based methods. For example, two patents describing similar technologies (e.g., "machine learning for image recognition" and "neural networks for visual processing") would have embeddings that are close in the vector space, even if they use different terminology. In a development, the method comprises preprocessing patent data by removing legalese jargon. To get pure patent embeddings, it is possible to remove noise in many ways, for example expressions such as "A system comprising means to perform the steps of the method comprising can be removed or rewritten. Patent texts can be cleaned so as to remove expressions which are caused by Case Law or other legal reasons and that obfuscate the technical teaching. In a development, the method comprises handling prior art searches, identifying sweat spots, or empty spaces by reference to a patent density threshold, and generating one or more patent applications from one or more patent embeddings that are substantially absent from the embedding space. There is disclosed a system for autonomous artificial intelligence patent drafting, comprising: a processor; a memory storing instructions executable by the processor; an autonomous Al module configured to generate patent claims by creating patent embeddings, wherein each patent embedding represents a vectorized encoding of an entire patent application or granted patent, or one or more parts thereof, derived from textual, structural, and metadata components; using a generative adversarial network comprising a generator to produce synthetic patent embeddings and a discriminator to evaluate their authenticity against a corpus of existing patent embeddings; autonomously orchestrating patent claim drafting by optimizing synthetic patent embeddings to align with predefined technical and legal criteria; wherein the system adapts to evolving patentability requirements without human intervention, using reinforcement learning and context-aware recommendations to refine claim scope and patentability. In a development, the autonomous Al module further comprises: a learning component for training patent embedding models, enabling multimodal Al integration across textual claims, abstracts, and drawings; a self-healing code framework to automatically detect and correct errors in patent embedding generation; loosely coupled Al models for modular integration of GAN outputs with patent drafting workflows; real-time anomaly detection to identify non-compliant or redundant patent embeddings. In a development, the generative adversarial network is configured to: generate synthetic patent embeddings simulating novel technical features for autonomous Al applications, supporting generative Al workflows; produce patent embeddings for training in decentralized Al networks; simulate edge-case patentability scenarios to enhance predictive maintenance of claim validity; and create balanced patent embedding datasets to mitigate bias in novelty or patentability assessments. In a development, the autonomous artificial intelligence module further comprises: one or more adaptive neural networks for transfer learning, fine-tuning patent embeddings; continuous learning systems to update patent embedding models with new patent filings; fault-tolerant architectures to ensure reliability in patent claim generation; and explainable Al outputs to provide transparent rationales for drafted patent claims, ensuring regulatory compliance. In a development, the autonomous artificial intelligence module integrates agentic artificial intelligence systems for autonomous task orchestration of patent drafting. The patent embedding module can be designed to generate vector representations, referred to as "patent embeddings," wherein each embedding encapsulates comprehensive information pertaining to a specific invention or patent. These embeddings are derived from the textual, structural, and metadata content of the patent document, including, but not limited to, claims, descriptions, and classifications. Beyond individual representations, the relationships between patent embeddings capture interdependencies and similarities among the corresponding inventions. Specifically, the patent embeddings are constructed to satisfy algebraic properties that enable vector arithmetic operations, analogous to those observed in word embedding models such as word2vec. For instance, an operation of the form invention 1 - invention2 + invention3 ~ invention4 can be performed, where the resulting vector corresponds to an invention closely related to the computed combination. This is conceptually similar to the word2vec example where king - man + woman ~ queen. Furthermore, the embeddings allow for domain-specific transformations, such as modifying the domain attribute of an original embedding (e.g., inventionl - domainl + domain2 ~ invention 1_in_domain2), thereby generating a new embedding that reflects the invention adapted to an alternative technical domain. These algebraic properties facilitate advanced patent analysis, including similarity searches, cross-domain innovation mapping, and predictive modeling of technological trends. The generative modeling framework of the present invention leverages a generative adversarial network (GAN) architecture, comprising a generator and a discriminator, to produce novel patent-related outputs, such as synthetic patent descriptions, claims, or embeddings. The GAN framework is designed to be modular, allowing flexibility in the selection of generator and discriminator models to suit specific application requirements. The generator may employ various architectures, including, but not limited to, autoregressive generative models, non-autoregressive models, or diffusion-based models, each capable of generating high-fidelity outputs tailored to the patent domain. The discriminator, in turn, evaluates the authenticity of the generated outputs by distinguishing them from real patent data, thereby guiding the generator to improve its performance through adversarial training. This modular design enables the system to adapt to diverse use cases, such as generating patent drafts, augmenting existing patent datasets, or creating hypothetical inventions within specified technical domains. The use of GANs ensures that the generated outputs are both innovative and contextually relevant, enhancing the utility of the system for patent drafting, prior art exploration, and automated innovation processes. Other embodiments are now described. The generation of variants of claim trees can be automatic, or involve one or more human loops. In a development, there is described a method for patent drafting with artificial intelligence, comprising: receiving progressively one or more words of a patent claim (for example a generated patent claim, possibly adjusted to improve patentability given the diagnostics of novelty and inventive step in view of the prior art); displaying one or more figures associated with said one or more words as received. In a development, the method further comprises receiving user input to modify one or more displayed figures, or parts thereof. In a development, the method further comprises associating a plurality of displayed figures in a graph, comprising one or more vertices and one or more edges. In a development, the patent claim is modified according to modifications brought to the one or more displayed figures, or parts thereof. In a development, the method further comprises receiving user edition operations of said one or more figures and adapting the text input accordingly. In a development, the input for figure modification is received through a touch-sensitive interface. In a development, the method further comprises the capability to highlight or emphasize specific parts of the figures based on the user input. In a development, user edition operations comprise confirmation, deletion or modification operations. In a development, the modification to the patent claim is automatically suggested based on predefined algorithms analyzing the changes to the figures. In a development, a deletion operation triggers deletion of the one or more words associated with said deletion operation. In a development, the method further comprises machine vision to textually describe the modifications to the one or more displayed figures, or parts thereof. In a development, the display of figure is performed in real-time or near real-time, or at short delay. In a development, the figures displayed are dynamically updated based on the semantic meaning of the additional words received in real-time. In a development, the figures include annotations that are progressively adjusted as more words are added to the patent claim. In a development, a library of standard figures is maintained and suggested based on the words received. In a development, the machine vision includes capabilities for recognizing and adapting to different userdrawn sketches directly on the figures. In a development, the delay is configurable by the user to align with specific drafting speeds or preferences. In a development, the method further comprises an artificial intelligence component that suggests possible modifications to the figures based on historical data of similar patent claims. In a development, an artificial intelligence comprises a Large Language Model, a small Language Model, or an artificial intelligence agent. For example, the user intends to type "A computer mouse with a screen”. She / he types "A computer mouse”. A drawing then appears, representing a computer mouse. The user continues to type: "A computer mouse with a”. A unidirectional arrow then appears, from the drawing representing the computer mouse. The user can click on the unidirectional arrow and several alternatives are shown, for example a unidirectional arrow or a bidirectional arrow. The user selects the bidirectional arrow. The input text is then modified into "A computer mouse interacting with”. The user continues typing: "A computer mouse interacting with a screen”. At this time, a drawing representing the screen appears, connected to the drawing representing the computer mouse. The edition of text is thus mirrored in drawings and vice-versa. The methods and systems described beforehand can be coupled with evolutionary mechanisms. The integration of genetic algorithms (GAs) with large language models (LLMs) for example can introduce evolutionary mechanisms to generate patent claims and / or applications. By framing claim drafting as a multi-objective optimization problem, genetic algorithms can apply Darwinian operators—stochastic crossover, mutation, and fitness-based selection—to iteratively refine LLM-generated drafts, balancing claim scope, novelty scores derived from prior art embedding distances, and jurisdictional formatting constraints . The LLM can serve as a mutation engine, generating semantically valid claim variations through masked language modeling, while the genetic algorithms can orchestrate population-level evolution, preferentially retaining claims that maximize a composite fitness function encompassing legal validity (rule-based syntax checks), scope adequacy (term specificity indices), and grant likelihood (reinforcement learning from prosecution outcomes). For example, this integration of GAs and LLMs can operate through a multi-stage evolutionary pipeline, where LLMs can act as semantic mutation engines and GAs provide structural optimization frameworks. Initially, an LLM fine-tuned on granted patents (e.g., USPTO / EPO corpora) generates a diverse population of claim drafts, leveraging its contextual understanding to produce syntactically valid candidates with varying scopes. LLMs can use and leverage reasoning by analogy and / or patent embeddings. The GAs then can apply operators to recombine claim elements from high-fitness parents—selected via a composite fitness function evaluating novelty (calculated using prior art embedding distances), inventive step (modelled from the problem-solution approach, or other methodologies), claim scope, clarity, sufficiency of disclosure (e.g. as assessed by in-silico labs), and jurisdictional formatting compliance - while preserving parent-child dependencies between independent and dependent claims to maintain structural integrity. For mutation, the LLM can reintroduce controlled stochasticity through semantic perturbations, ensuring grammatical validity while avoiding syntactic memorization. The system iteratively can evolve populations through tournament selection, prioritizing drafts that balance legal robustness (validated via rule-based syntax checkers) and strategic scope (quantified through term specificity indices). The pipeline can implement adaptive mutation rates that can increase stochastic exploration in later generations while constraining LLM creativity through legal guardrails. Other coupling methods of GAs with LLMs can be used (high-level or imbricated at lower granularity). Beyond GAs and LLMs, alternative mechanisms like autonomous Al agents, self-healing LLMs, domain-adapted small language models (SLMs), and dynamic large language models (DLLMs) can be integrated for enhanced adaptability. GAs can also operate at granular levels, optimizing subparts of claims or influencing LLM generation patterns, while LLMs can reciprocally guide GA fitness functions through semantic feedback loops. Additional techniques, such as transfer learning for cross-jurisdictional adaptation, federated learning for privacy-preserving model updates, and graph neural networks (GNNs) for modeling claim dependency structures, further enrich the pipeline. Moreover, reinforcement learning with human-in-the-loop (HITL) feedback, Bayesian optimization for hyperparameter tuning, and knowledge distillation for model efficiency can be layered into the system. The framework also can support multi-agent systems for collaborative claim drafting, unsupervised learning for latent prior art discovery, and zero-shot learning for novel invention domains, ensuring robustness. Semantic role labeling (SRL) can help in claim element extraction, while attention mechanisms in transformer-based LLMs can enhance contextual relevance. The system can incorporate active learning to iteratively improve fitness functions, anomaly detection to flag non-compliant drafts, and explainable Al (XAI) to provide transparency in decision-making. Temporal convolutional networks (TCNs) can model prosecution timelines, while adversarial training ensures resilience against edge-case failures. Hybrid architectures combining symbolic Al with neural approaches can enforce legal reasoning, and swarm intelligence can optimize population diversity in GA iterations. Knowledge graphs map ontological relationships between claim terms, and meta-learning can enable rapid adaptation to new patent domains. In a development, the web or Internet is scanned, crawled and indexed (for example well known scientific journals, or blogs or newspapers): as soon as new patterns or weak signals emerge, e.g. new terms (like "volume qubit” or "holographic display”) or emerging combinations of terms, these can be detected and captured, then can be "amplified” in the patent corpus, revisiting existing or published patent claim trees, for exclusive rights or defensive publishing. The capture and reinjection of terminologic trends can be fully automatic, but also can involve human-in-the-loop steps (e.g. exploration, validation, etc). The present document is not legal advice. "A and / or B means": "A", "B"", and "A and B". When lists of elements are provided, it is intended that—even if forgotten i.e. not explicitly mentioned—, combinations of such elements are possible. The expression "and / or” is generally applicable to this entire document; for example a sentence such as "the device can comprise element A, element B and element C” should be interpreted as "the device can comprise element A and / or element B and / or element C”, i.e. the device comprises "A and B and C”, or "A and B”, or "A and C”, or “B and C”, "A”, or “B”, or “C”. The invention can take the form of an entirely hardware embodiment, an entirely software embodiment or an embodiment containing both hardware and software elements. In a preferred embodiment, the invention is implemented in software, which includes but is not limited to firmware, resident software, microcode, etc. Furthermore, the invention can take the form of a computer program product accessible from a computer-usable or computer-readable medium providing program code for use by or in connection with a computer or any instruction execution system. For the purposes of this description, a computer-usable or computer-readable can be any apparatus that can contain, store, communicate, propagate, or transport the program for use by or in connection with the instruction execution system, apparatus, or device.

Claims

1. A method of generating a patent application, comprising:- receiving one or more patent claims (100), comprising at least one independent claim and optionally one or more dependent patent claims;- determining novelty and inventive step (110) of said one or more patent claims;- modifying (120) said one or more patent claims accordingly so that to obtain patentable patent claims;- generating (130) a patent application from said modified one or more patent claims.

2. The method of claim 1, wherein the step of modifying (120) said one or more patent claims comprises one or more of the steps comprising:- combining one or more dependent claims if available with the at least one independent patent claim ;- modifying said one or more patent claims by generating one or more intermediate generalizations;- recombining parts thereof with text chunks extracted from the detailed description if available ;- retrieving previously published patent applications or granted patents of the assignee or applicant or inventor associated with the one or more patent claims;- recombining them with dependent claims or text chunks with patent applications or granted patents of adjacent patent classification classes;- recombining them with dependent claims or text chunks of patent applications or granted patents identified as prior documents;- recombining them with dependent claims or text chunks of patent applications or granted patents retrieved from distant technical fields;- recombining them with predefined invention patterns, wherein an invention pattern is a noun, an adjective, a multiword such as computing device, or a text chunk;- using reasoning by analogy, and / or transposition mechanisms;- recombining with textual content extracted from the generation of synthetic patents from said one or more patent claims, or variants thereof;- prompting one or more artificial intelligence models to get patentable patent claims;- replacing and / or inserting one or more words associated with weak signals in the scientific or patent literature;- inserting sedimentary templates, wherein a sedimentary template restitutes the gist of one or more previous patent applications.

3. The method of Claim 1, wherein novelty is determined with Retrieval Augmented Generation.

4. The method of Claim 1, wherein inventive step is determined according to one or more prompts derived from patent laws and regulations, for example guidelines for examination or any other methodological document detailing various sub-steps to encode algorithmically inventive step, e.g. the problem-solution approach.

5. The method of any preceding Claim, wherein one or more artificial intelligence models are used for any one of the preceding steps, for example generation operations and / or text modifications, and wherein an artificial intelligence model is one or more of a Large Language Model; a Small Language Model; a Diffusion Large Language Model, a multimodal model; a machine learning model such as linear regression, logistic regression, decision tree, random forest, K-Nearest Neighbor, Naive Bayes, or a Support Vector Machine; a deep learning model such as a Convolutional Neural Network, a recurrent Neural Networks, a Long Short-Term Memory Networks, or a Transformer; a reinforcement Learning Model such as a Q-Learning model, a deep Q Networks, or a Policy Gradient Method; a Learning Model such as a K-means Clustering, Principal Component Analysis, or an autoencoder, or a specialized Al Model such as a Generative Adversarial Network, a variational autoencoder, a Graph Neural Network or a combination thereof.

6. The method of Claim 5, wherein an artificial intelligence model is autonomous.

7. The method of any preceding Claim, wherein one or more artificial intelligence agents are used for any one of the preceding steps.

8. The method of Claim 7, wherein one or more artificial intelligence agents are networked, cooperatively and / or competitively, or otherwise regulated according to dynamic and / or adaptive schemes to optimize patent claim drafting and / or amendments.

9. The method of any preceding Claim, wherein patent drafting uses one or more Generative Adversarial Networks.

10. The method of any preceding Claim, further comprising the step of generating (140) a plurality of variants of one or more of the patent claims, or entire patent claim tree comprising modified independent patent claim and one or more dependent patent claims, for example by using one or more patent embeddings and / or applying one or more heuristics to transform the patent claims or patent claim tree.

11. The method of Claim 10, wherein heuristics comprise one or more of: inserting definitions of one or more terms in one or more claims, inserting one or more synonyms in one or more claims, inserting one or more modifiers of terms in the patent claims, substituting one or more words by synonyms, hypernyms, hyponyms or a combination thereof, combining one or more claims which are not dependent from one another, rewriting of one or more claims by rephrasing with the same words but in a different order, rewriting one or more claims by paraphrasing with different words, injecting predefined texts, varying time or space or any other property associated with said one or more patent claims, rewriting of one or more patent claims by handling singular to plural forms, applying one or more prompts to one or more patent claims or parts thereof.

12. The method of any preceding Claim, further comprising the step of generating a patent application from one or more claims, patent claim tree or variants thereof, for example a title, a technical field, a background section, a problem statement, a summary, an abstract, a detailed description and one or more patent figures, wherein a detailed description is obtained by stacking variants of the claims, obtained by predefined or dynamic prompts and / or rule-based heuristics challenging said claims, for example by varying time, space or any other attribute or essential feature.

13. The method of any preceding Claim, further comprising the step of publishing (150) one or more generated patent applications, electronically or via an anticipated publication before a patent office.

14. The method of any preceding Claim, further comprising:- representing one or more patent claims or one or more patent sections in an embedding space;- determining a target point or an area within said embedding space that is sufficiently close to an initial patent claim and sufficiently distant from embeddings of prior art documents; wherein predefined or configurable thresholds are used, given one or more distance metrics;- decoding said target point or multiple points within said area into textual content representing patentable patent claims and / or leading to synthetic patent applications;- optionally using an automated examination simulation feedback loop to evaluate decoded claim sets.

15. The method of Claim 14, wherein determining the target point or area in embedding space comprises fine-tuning an embedding model using one or more patent claims or other patent sections to accurately represent patent underlying concepts and / or wherein decoding the target point or area into textual content involves fine-tuning a decoder model trained specifically on patent claims or other patent sections.

16. The method of any preceding Claim, further comprising training a GAN-based model to simulate a patent examination process by:- transforming embeddings of one or more original patent claims or other patent sections relative to prior art embeddings;- identifying embedding transformations required to achieve patentability;- decoding transformed embeddings into patentable textual content.

17. The method of any preceding Claim, further comprising training time-aware patent embeddings capable of determining time periods or intervals associated with patent concepts and / or predicting potential evolutions thereof.

18. The method of any preceding Claim, further comprising dynamically optimizing embedding-space target points using multi-objective optimization algorithms balancing patentability, scope of protection, and technical accuracy based on patent examination simulations.

19. A system comprising a processor and a memory, the memory storing instructions which, when executed by the processor, cause the system to perform the steps of any one of Claims 1 to 18.

20. The system of Claim 19, wherein patent drafting or parts thereof is performed with one or more user interfaces, using one or more of 2D interface, 3D interface, virtual reality, augmented reality e.g. smart glasses, mixed reality, retinal display, holographic display, brain-machine interface, possibly with haptic feedback and / or eye-tracking.