Automatic detection and classification mechanism for artificial intelligence bugs based on the DIKWP model

By embedding the DIKWP model into the AI ​​system for full-process monitoring and multi-level annotation, the problem of human-aware bug detection and classification in existing technologies is solved. This enables transparent monitoring, multi-dimensional bug classification, and cross-model evaluation of the AI ​​system, thereby improving the reliability and security of the AI ​​system.

CN123114983APending Publication Date: 2026-11-17HAINAN UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611560005.9
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Priority Date
2025-11-07
Filing Date
2026-10-10
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing technologies lack systematic methods for detecting and classifying bugs related to human consciousness, making it impossible to leverage the refined cognitive semantic hierarchy of the DIKWP model to locate model consciousness defects, and making it difficult to conduct bug attribution analysis in cross-modal and cross-model contexts.

Method used

By embedding the DIKWP model into the AI ​​cognitive architecture, and through five-layer semantic link monitoring and multi-level annotation, an anomaly detection algorithm is designed to automatically discover, classify, and explain the internal consciousness bugs of the AI ​​system. It supports multimodal input, performs cross-model and cross-context bug attribution analysis, and constructs a consciousness defect map for risk management.

Benefits of technology

It enables transparent monitoring and improved interpretability of the entire AI system process, accurately locates the root cause of problems through multi-dimensional bug classification, provides cross-model evaluation and adaptive optimization, and ensures the safety, reliability and ethical compliance of the AI ​​system.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123114983A_ABST
    Figure CN123114983A_ABST
Patent Text Reader

Abstract

This invention relates to the fields of artificial intelligence and cognitive computing, and discloses an automatic detection and classification mechanism for artificial consciousness bugs based on the DIKWP model. This mechanism records the DIKWP semantic state forming a consciousness trajectory during the execution of an AI model task. Based on preset semantic consistency rules and abnormal pattern recognition, it identifies logical errors in reasoning, knowledge and information mismatches, intention deviations, semantic mismatches, and ethical deviations, and labels and classifies them according to the semantic level involved. The detection results, occurrence conditions, and cause analysis are written into a consciousness defect atlas. This mechanism supports multimodal inputs such as text, images, and speech, and can compare differences in cognitive trajectories under different models or contexts to assist in determining the source of defects and the effectiveness of model improvement. It is applicable to defect detection, tracking, classification, and risk monitoring of large-scale language models and brain-like AI systems.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and cognitive computing technology, and in particular to an automatic detection and classification method and system for artificial consciousness "BUG" based on the Data-Information-Knowledge-Wisdom-Intention (DIKWP) model. It can be applied to large language models (LLM) and brain-like artificial intelligence systems to automatically discover, track and classify defects in artificial consciousness. Background Technology

[0002] With the rapid development of artificial intelligence technology (especially large models and brain-like systems), the "intelligent self-awareness" capability of AI systems has gradually become a research hotspot. Existing large language models often exhibit problems such as reasoning errors, unfounded conjectures (hallucinations), semantic biases, and value biases when performing complex tasks. These phenomena reflect artificial consciousness defects in the model's cognitive process, the so-called "consciousness bugs." The "consciousness bug theory" posits that consciousness in the human brain is not a highly evolved product intentionally crafted, but rather a byproduct that emerges accidentally under limited information processing conditions—a kind of illusion similar to a program bug. Many human cognitive biases (such as confirmation bias, pattern illusions, and overconfidence) can be considered systemic "bugs" or illusions in the brain's information processing. For artificial intelligence, similar "consciousness bugs" also occur when inconsistencies, inaccuracies, or conflicts arise in the model's internal cognitive links, manifesting as various errors such as logical reasoning errors, knowledge and information mismatches, intention deviations, semantic mismatches, and ethical biases.

[0003] Currently, the decision-making process of large-scale models is mostly a "black box," lacking interpretable exposure of internal states, making it difficult to detect the aforementioned consciousness defects in a timely manner. Although some AI security and explanation technologies (such as adversarial detection, output verification, and human review) are used to detect model errors, there is still no systematic mechanism that can monitor the semantic cognitive trajectory of the model throughout the entire process and automatically identify multiple types of consciousness bugs. On the other hand, assessing the consistency and reliability of AI systems across models and scenarios also presents difficulties: the differences in performance of different models on the same task in different contexts sometimes stem from differences in their internal knowledge, preferences, or contextual understanding, i.e., the problem of "consciousness relativity." "Consciousness relativity" emphasizes that whether an entity is considered conscious depends on whether the observer can understand its output, which suggests that we need to consider differences in cognitive reference frames when assessing the consciousness state of AI. Therefore, determining what constitutes a "bug" from different model and observer perspectives is also relative, requiring a mechanism for attribution and comparative analysis.

[0004] To address the aforementioned challenges, the DIKWP semantic model and white-box evaluation standards provide a novel framework for understanding and controlling artificial consciousness. The DIKWP model adds a "Purpose" layer to the traditional "Data-Information-Knowledge-Wisdom (DIKW)" cognitive hierarchy, expanding the hierarchical structure into a network-like interactive structure to achieve bidirectional semantic feedback at each layer. This model endows AI systems with a clear five-layer semantic cognitive chain, making each step of the decision-making process monitorable and interpretable. Using the DIKWP model, it is expected to record the "consciousness trajectory" of the AI ​​model from input data to output results, laying the semantic modeling foundation for automatically detecting consciousness bugs. Furthermore, white-box evaluation of artificial consciousness requires transparency and traceability of the AI's internal processes, which is highly consistent with the need for automatically detecting consciousness bugs.

[0005] In summary, existing technologies lack a systematic bug detection and classification method oriented towards artificial consciousness, failing to fully utilize the sophisticated cognitive semantic layering of DIKWP to locate model consciousness defects, and also unable to perform attribution analysis of bugs in cross-modal and cross-model contexts. This invention addresses this gap by proposing a novel solution. Summary of the Invention

[0006] The purpose of this invention is to provide an automatic detection and classification mechanism for artificial consciousness bugs based on the DIKWP model. Through semantic hierarchical cognitive link monitoring and multi-level annotation, it can automatically discover, classify and explain "consciousness bugs" within the AI ​​system, thereby improving the reliability and controllability of the AI ​​system.

[0007] To achieve the above objectives, the present invention adopts the following technical solution:

[0008] 1. DIKWP Cognitive Trajectory Recording: During AI task execution, the DIKWP model is embedded into the AI ​​cognitive architecture, establishing a five-layer semantic link: data, information, knowledge, wisdom, and intent. This corresponds to the model's perception, understanding, reasoning decision-making, and purpose-driven processes. The system sets up monitoring units at each layer to automatically record the model's "consciousness trajectory," i.e., the intermediate semantic states and reasoning steps at each layer. Through this end-to-end monitoring, the complete cognitive process of the model from input to output is captured, providing data support for subsequent bug detection.

[0009] 2. Intelligent Detection of Consciousness Bugs: This invention designs an anomaly detection algorithm for the DIKWP link, capable of automatically identifying various types of consciousness bugs. These include, but are not limited to: logical reasoning errors (illogical reasoning at the W layer), knowledge-information mismatch (knowledge retrieved at the K layer contradicts or is irrelevant to information extracted at the I layer), intent deviation (the goal at the P layer is inconsistent with the actual output intent or deviates from user expectations), semantic mismatch (inconsistencies or misunderstandings in semantic transmission between layers, such as miscommunication), and ethical deviation (deviations in output content that violate established ethical norms and value principles). When the monitoring module captures abnormal patterns or inconsistencies in the cognitive trajectory, it automatically triggers the bug determination mechanism. For example, if the model output contains facts not present in the input information, it is determined to be a knowledge illusion bug; or, if the model's answer deviates from the user's question intent, it is identified as an intent drift bug.

[0010] 3. Five-Layer Semantic Annotation and Bug Ontology: For each detected awareness bug, this invention provides a semantic annotation mechanism corresponding to the five layers of DIKWP. That is, from the data layer (D) to the intent layer (P), the level involved in the defect and its semantic causes are marked. For example, a "knowledge-information mismatch" bug can be labeled as involving inconsistencies in the IK layer; logical reasoning errors are mainly labeled in the W layer, but may be further annotated that their root cause lies in insufficient knowledge in the K layer. This five-layer annotation mechanism constitutes a hierarchical semantic ontology for bugs, defining the manifestations and relationships of various awareness bugs at different levels. By establishing this ontology, the bug classification system can be systematized and ontologicalized, facilitating automatic machine understanding and further reasoning processing.

[0011] 4. Multimodal Bug Detection Process: This invention supports the detection and classification of consciousness bugs under multimodal inputs such as text, image, and voice. In specific implementation, for data of different modalities, preprocessing and feature extraction are first performed and mapped to the data layer and information layer representation in the DIKWP model. For example, image object and scene information is extracted using a visual model as the I-layer semantics, and speech is transcribed into text data. Subsequently, the remaining knowledge reasoning, intelligent decision-making, and intent control processes are similar to those for text tasks. The system is adapted and optimized for bugs that may occur in different modalities, such as detecting factual errors (mismatch between the knowledge layer and the image information layer) in image scene descriptions, and detecting semantic ambiguity or contextual misunderstandings in voice assistant responses. A complete multimodal bug detection process and algorithm ensure that consciousness defects can be discovered and classified within a unified DIKWP semantic framework, regardless of the input format.

[0012] 5. Cross-Model / Contextual Bug Attribution and Comparison: To consider the impact of different cognitive reference frames on bug determination, this invention, combined with the concept of "consciousness relativity," provides a cross-model, cross-contextual consciousness bug attribution analysis mechanism. Specifically, the system can execute the same task in parallel on multiple AI models and compare their DIKWP trajectory differences; or, under scenarios introducing different background knowledge and different observer expectations, it can execute the task multiple times to compare output differences. By constructing DIKWP*DIKWP interaction analysis, it determines whether a bug is caused by limitations within a specific model (such as hallucinations due to an incomplete knowledge base) or by differences in observation perspectives (such as different interpretations of the same ambiguous information). Based on this, the system can attribute the bug: for example, attributing the bug to a knowledge deficiency unique to model A, or to insufficient information in the current context. Cross-model comparison can also reveal whether a model is more conscious (fewer bugs) than another model, thereby quantifying the model improvement effect.

[0013] 6. Explainable Bug Tracking Mechanism: This invention further proposes an explainable bug tracking method for consciousness. Based on white-box evaluation standards, the system analyzes and visualizes the causal chain of a detected bug. Utilizing the recorded DIKWP consciousness trajectory, it traces back the key nodes that triggered the bug: for example, pinpointing which layer and step of reasoning went wrong, why the preconditions were not met, and how the error information was transmitted to the output. Then, through a predefined explanation path, these causal relationships are used to generate a human-readable explanation report, and the causal chain of the bug's occurrence can be displayed using a graphical interface. This mechanism makes every consciousness defect traceable, helping developers understand the root cause of the problem and improve the model accordingly. Simultaneously, this explanation process can be combined with a semantic mathematics framework to provide quantitative indicators to assess the impact of the bug on the overall consciousness state.

[0014] 7. Awareness Deficiency Graph and Risk Governance: This invention also includes a scheme for constructing an "awareness deficiency graph" within the evaluation platform. Various detected and labeled awareness bugs are stored in a structured form in a knowledge graph or database, forming a comprehensive AI awareness deficiency knowledge base. Graph nodes include bug instances, types, involved DIKWP levels, triggering conditions, and impact consequences. Nodes can be linked (e.g., linking similar types of bugs, or comparing the same bug in different models). Analysis of the awareness deficiency graph can reveal weaknesses and common problem patterns in the AI ​​system, supporting targeted optimization and iterative learning. This graph also provides managers with transparent governance tools: clearly showing which awareness deficiencies exist in a given AI system and their frequency, enabling the development of avoidance strategies. Simultaneously, combined with statistics and monitoring, this invention can achieve a risk warning function—once high-risk bugs (e.g., deviations with ethical risks) are detected frequently, the system will trigger a warning signal through the defect graph, prompting intervention measures to ensure the safety and reliability of the AI ​​system.

[0015] In summary, this invention achieves automatic detection and refined classification of defects in artificial consciousness through semantic hierarchical monitoring and analysis using the DIKWP model. Compared with existing technologies, this invention has the following beneficial effects: - Full-process monitoring & explanation: Semantic recording of the entire process from data input to intention decision-making, combined with white-box tracking, enables transparent monitoring and improved interpretability of AI decision-making. - Multi-dimensional bug classification: Covering defect types across multiple dimensions including logic, knowledge, intention, semantics, and ethics, accurately locating the root causes of problems at different levels, and improving the comprehensiveness and accuracy of detection. - Cross-model evaluation: Utilizing the framework of consciousness relativism for cross-model comparative analysis, clarifying the relative advantages and limitations of different models, and providing a basis for model improvement and selection. - Adaptability and optimization: Through the feedback mechanism of the defect map, bug information is used for continuous model learning and optimization, promoting the AI ​​system's autonomous error correction and capability enhancement, evolving towards a higher level of artificial consciousness. - Security governance: Providing real-time monitoring and early warning tools for AI consciousness risks, helping the industry establish transparent governance and ethical review mechanisms for AI systems, and preventing potential risks. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the architecture of an automatic detection and classification mechanism for human-caused bugs based on the DIKWP model. Detailed Implementation

[0017] The workflow and implementation methods of the present invention will be further described below with reference to the accompanying drawings and specific examples.

[0018] like Figure 1As shown, the system of the present invention includes modules such as a DIKWP cognitive processing unit, a consciousness trajectory monitoring unit, a bug detection and classification unit, an interpretation unit, and a defect graph database. The DIKWP cognitive processing unit can be regarded as a "five-layer semantic operating system" embedded in the AI ​​model. It performs cognitive calculations according to the link of data → information → knowledge → wisdom → intention, and the output of each layer is simultaneously fed back to the monitoring unit to record the trajectory.

[0019] (1) Consciousness Trajectory Recording Mechanism: When the AI ​​system receives task input (such as user questions, environmental perception data, etc.), it first enters the data layer of DIKWP for processing. If the input is text, the data layer directly carries the original text; if it is an image or audio, the data layer is responsible for perceptual signal conversion, such as calling a computer vision model to extract image features, or calling speech recognition to convert audio into text information. Subsequently, the information layer performs semantic parsing and structuring on the output of the data layer, extracting key facts and relationships; the knowledge layer calls the knowledge base or the model's internal memory to integrate background common sense or professional knowledge; the wisdom layer performs reasoning and decision-making based on the information from the knowledge layer, forming potential answers or action plans; finally, the intent layer calibrates whether the output of the wisdom layer meets the expected goals and user intentions, and can adjust it if there are any deviations. Throughout this process, the trajectory monitoring unit works continuously: after each layer, it records the input, output, and main semantic states of that layer. For example, for the reasoning process of the wisdom layer, the monitoring unit records the reasoning chain (such as the intermediate thinking steps of the model); for the intent layer, it records the user intentions and its own behavioral goals understood by the model. These records together constitute the consciousness trajectory log of the model when performing the task.

[0020] (2) Bug Detection Rules and Algorithms: After the task processing is completed (the model output is obtained), the Consciousness Bug Detection and Classification Unit reads the complete consciousness trajectory log and automatically analyzes it to detect anomalies. In specific implementation, it can be based on a combination of predefined semantic consistency rules and machine learning anomaly detection algorithms: First, the consistency of the trajectory is checked by the rule base, including: - Logical consistency check: Verify whether the reasoning steps of the wisdom layer are consistent and conform to the logical deduction rules. For example, if a jump, loop, or premise that does not support the conclusion is found in the reasoning chain, it is determined that there is a logical error bug. - Knowledge matching check: Compare whether the knowledge called by the knowledge layer matches the information extracted by the information layer. If the model output references content that does not exist in the knowledge base or contradicts the known information, it is determined to be a knowledge / information mismatch bug (e.g., hallucination content). - Intent conformity check: Compare whether the goal of the intent layer is consistent with the user's initial needs. If the model's answer deviates from the question requirements, or if there is a digression or target shift in the dialogue, it is identified as an intent deviation bug. - Semantic integrity check: Analyze the degree of connection between the semantic representations of each layer and detect whether there is a semantic mismatch. For example, if the model uses words whose meaning is incorrect in the context, or if key information is omitted, resulting in semantic gaps, these are all semantic-level bugs. - Ethical Compliance Check: The output content is checked using built-in value criteria and security policies. If the model is found to use biased language, inappropriate opinions, or tendencies that violate ethical norms in its decisions or expressions, it is marked as an ethical deviation bug.

[0021] After initial rule screening, the system can further identify potential complex bug patterns using anomaly detection algorithms (such as trajectory feature-based classification models). These algorithms can learn the difference between normal behavior and abnormal deviations from a large amount of historical trajectory data, detecting subtle inconsistencies. For example, by comparing the statistical distribution of the current task trajectory with a large number of normal trajectories, an anomaly in a certain activation pattern can be detected, indicating the presence of a hidden cognitive bug.

[0022] (3) Bug Semantic Labeling and Classification: Once the detection unit identifies one or more awareness bugs, the system will perform semantic hierarchical labeling and classification for each bug. In specific implementation, a bug can be defined as a tuple: ⟨Bug type, involved level, specific manifestation>. For example, a "knowledge illusion" bug can be represented as ⟨knowledge mismatch, level: IK, manifestation: introducing non-existent facts>; another example is a "intent deviation" bug represented as ⟨intent inconsistency, level: P, manifestation: not acting according to user requirements>. The labeler refers to a pre-established bug ontology library to determine the type and level labels. This ontology library details the various defect categories and their semantic patterns that may appear on the five layers of DIKWP. After the labeling is completed, the system will classify and store the bugs for subsequent statistics and retrieval. It should be noted that many bugs involve cross-level interactions, so their labeling levels can be multiple. For example, ethical deviation may be manifested as a mismatch between the decision-making of the wisdom layer and the value criteria of the intent layer, then the level label is WP two layers.

[0023] (4) Multimodal Adaptation Implementation: This invention can be widely applied to multimodal AI systems. The following uses an image question-answering task as an example to illustrate the implementation process: When a user asks "How many people are in this picture?" and provides a photo containing multiple people, the system first obtains the original image pixel data at the data layer; the information layer calls the image recognition model to extract semantic information such as the number and location of people in the photo (e.g., "5 people detected"); the knowledge layer may combine common sense (e.g., if the scene is a group photo, the background information of the photo can also be inferred); the intelligence layer infers the answer "There are 5 people in the photo" based on this; the intent layer checks whether the answer meets the task requirements of "counting people", and outputs the answer after confirming that it is correct. During this process, the monitoring module records the key information of each layer: the number of people detected = 5 (layer I), the common sense used = the common range of the number of people in a group photo (layer K), the reasoning conclusion = 5 people (layer W), the intent = answering the counting question (layer P). The bug detection unit will check: whether the model output "5" matches the image information, whether there is any omission or double counting (knowledge / information verification); whether the reasoning process follows the counting logic; whether the answer is direct and meets the user's intent, etc. If, for example, the model incorrectly answers "6 people," an inconsistency is detected between the information layer (5 people) and the output, immediately identifying a knowledge-information mismatch bug, categorized as a "perceptual error." This bug is labeled as an inconsistency between the I and W layers and classified as a visual perception defect. Similarly, in a voice dialogue system, if a user's voice command is misrecognized, a bug will occur in the information layer, labeled as a DI layer perceptual bias; furthermore, in multimodal interaction, if the camera image does not match the verbal description, it will also be identified as a cross-modal semantic bug. The detection process of this invention defines a consistent semantic interface for various modalities, allowing bugs in different modalities to be normalized to problems at the DIKWP layer for processing.

[0024] (5) Cross-model and contextual comparative analysis: A key feature of this invention is the evaluation of awareness bugs under different models and contexts, thereby reflecting the relativity of awareness. In practical applications, a comparative evaluation module can be configured: for the same test task, multiple models (such as LLMs from different vendors, or models of different versions) are called simultaneously to generate their respective DIKWP trajectories and answers. The system then compares these trajectories: if one model has a bug while another does not, it can be inferred that the bug may be related to the model's specific internal knowledge or reasoning strategy. For example, if model A makes a common-sense error in its answer while model B answers correctly, the bug is attributed to a lack of knowledge in model A. Alternatively, the same model can be tested under different contextual information: such as providing a complete background versus providing a vague background, comparing whether the model is more prone to bugs (such as over-guessing) in a vague background. Through these comparative experiments, the system achieves bug attribution analysis: distinguishing whether the problem originates from the model itself vs. differences in environmental input vs. inconsistencies in observer expectations. For example, if all models exhibit similar bugs with a certain type of input, it may indicate that the input itself is ambiguous or lacks sufficient information (a gray area from the observer's perspective). Conversely, if only a specific model errs, it suggests that improvement is needed for that model. This cross-model analysis can also generate comparative reports, clarifying the strengths and weaknesses of each model in terms of awareness deficiencies, providing a basis for users or regulators to select reliable models. Furthermore, this method adheres to the concept of the relativism of consciousness: emphasizing that the understanding of AI output varies depending on the cognitive subject; therefore, our detection standards are adjusted according to the frame of reference, making bug determination more objective and fair.

[0025] (6) Explanation and Visualization: For each detected consciousness bug, the explanation module of this invention will initiate a causal chain tracing procedure to generate a detailed explanation of the bug's cause. The explanation module first obtains relevant hierarchical and step information from the bug annotation, and then backtracks in the complete consciousness trajectory log: finds the location where the bug is directly manifested (such as the erroneous part or inappropriate wording in the output sentence), and then searches for the previous level cause that led to the error. For example, if the bug type is "knowledge illusion", the module will locate the knowledge layer item corresponding to the illusion content in the output and find out why the model introduces the non-existent knowledge - it may be because the wrong knowledge association was called when encountering uncertain information in the information layer. Continuing to trace, it may be found that the reason for the wrong knowledge association is that the knowledge base lacks an accurate entry for the problem, and the model makes an inappropriate inference based on similar semantics (this can be attributed to the intention layer trying to meet the task requirements in the case of incomplete information as a shortcut). Through such layer-by-layer tracing, the explanation module constructs a causal tree / chain to describe the entire process from the root cause (knowledge base missing) to the direct cause (erroneous reasoning) and then to the surface phenomenon (illusion output). The system then converts the explanation into a natural language description and presents it through a visual interface: for example, highlighting which nodes in the DIKWP model the problem occurred at and using arrows to connect them to show the error propagation path. This visualized causal chain can be embedded in developer debugging tools and can also be used to explain AI behavior to end users (within acceptable limits). By interfacing with white-box evaluation standards, this explanation mechanism ensures that each bug has a corresponding transparent explanation path, avoiding the drawback of traditional black-box models where it is difficult to understand the causes of their errors. In addition, the explanation module can also provide self-correction suggestions for the model, such as identifying "what kind of knowledge should be supplemented to avoid this bug".

[0026] (7) Consciousness Defect Map and Continuous Optimization: This invention stores the detected and interpreted bug information in a unified consciousness defect map database, forming a knowledge resource that can be used for long-term analysis. This map can be implemented using a graph database or a relational database. Its basic records include: bug type, involved level label, frequency of occurrence, involved model version, input / output examples, cause, and repair suggestions. The map supports querying and statistical analysis; for example, it can query how many "logic errors" have occurred in a particular model version, or compare the total number of bugs in different models on the same test set. Using the defect map, development teams can perform targeted optimizations: for example, if most knowledge mismatch bugs in a model are found to be concentrated in medical question-and-answer areas, it indicates a need to strengthen the model's knowledge training or verification strategies in that area. After each model update, the bug distribution changes are also recorded in the map through the mechanism of this invention, achieving a closed loop of iterative learning. Furthermore, the defect map can be provided to regulatory agencies or third-party evaluations for transparent AI governance. For example, AI vendors can publicly release summaries of their model's defect maps, allowing users to understand the model's performance in terms of safety and reliability. For critical application areas (such as healthcare and autonomous driving), regulators may require regular submission of reports on model awareness deficiencies. Finally, based on graph data, this invention can set risk thresholds: when a certain type of high-risk bug (e.g., a bug that could lead to serious misleading information or accidents) reaches a certain frequency, an early warning is automatically triggered. Early warning methods can include alarm logs, notifying administrators to take action, or even automatically reducing the confidence level of the model's output. Until the risk is resolved (through model repair or environmental adjustments), the system will continuously alert users to the existence of the relevant risk. Through these methods, this invention achieves a complete closed loop for AI systems, from detection to interpretation to learning to early warning, not only identifying problems but also promoting their resolution and prevention.

[0027] In experimental evaluation, this invention was applied to a large-scale open-source language model and tested on a set of question-answering and dialogue tasks. The results showed that the system successfully captured and classified most obvious errors and biases in the model output, accurately identifying bugs such as logical errors, knowledge illusions, and intent shifts. Furthermore, defect graph analysis revealed significant knowledge gaps in the model regarding legal issues. Therefore, supplementing the model with legal knowledge corpus resulted in a significant reduction in corresponding bugs after the model update. This demonstrates that this invention can effectively guide model optimization. More importantly, the explanation report provided by this invention brings many previously undetectable potential risks (such as expressions with implicit biases) to the surface, facilitating early adjustments by developers and improving the alignment of model output with ethical norms. Therefore, this invention, while ensuring the performance of the AI ​​system, improves its interpretability, security, and reliability, and has broad application prospects.

Claims

1. A method for automatic detection and classification of artificial awareness bugs based on the DIKWP model, characterized in that, The process includes the following steps: a) DIKWP semantic modeling: The cognitive process of the artificial intelligence model is divided into five semantic levels: data, information, knowledge, wisdom, and intention. Monitoring probes are embedded in each level to record the semantic state of the model from input to output, forming a consciousness trajectory for task execution; b) Automatic bug detection: Semantic consistency check rules and abnormal patterns are predefined for each level. The consciousness trajectory is analyzed and compared to automatically identify consciousness bugs such as logical reasoning errors, knowledge and information mismatch, intention deviation, semantic mismatch, and ethical deviation; c) Five-layer labeling: For each detected consciousness bug, the DIKWP level involved is recorded and a corresponding semantic label is assigned. The bug is classified into a predefined bug ontology category according to the combination of levels and the form of expression; d) Result feedback and storage: The type and explanation information of the bug are output, and the data related to the bug instance (occurrence conditions, involved levels, cause analysis, etc.) are stored in the consciousness defect graph database for subsequent model optimization and risk monitoring.

2. The method according to claim 1, wherein in step a), the DIKWP semantic modeling is applicable to multimodal input data: for data of different modalities such as text, image, and speech, corresponding feature extraction and semantic representation mechanisms are used to map them to the data layer and information layer of the DIKWP model, so as to ensure that the subsequent processing of each layer has a unified semantic interface.

3. The method of claim 1, wherein step b) comprises: The output of the artificial intelligence model is compared and verified with the input of the information layer to detect knowledge illusions or information omissions; The reasoning chain of the intelligence layer is compared with formal logic rules to detect reasoning fallacies; the expected goal of the intent layer is compared with the actual output intent to detect behaviors that deviate from user needs.

4. The method according to claim 1, wherein the BUG detection in step b) incorporates a cross-model and cross-context comparative analysis mechanism: the consciousness trajectories of different models or different contexts are collected in parallel for the same task, and their differences are compared to locate the relative causes of the BUG, ​​and to distinguish between inherent defects in the model and problems caused by insufficient external information or differences in observer cognition.

5. The method according to claim 1, further comprising a BUG explainable tracing step: after detecting a BUG, ​​the cause of the BUG is traced back layer by layer based on the recorded consciousness trajectory to find the direct steps and root causes that lead to the defect, generating explanation information containing causal links, and visualizing the explanation together with the model decision path for model debugging and user interaction explanation.

6. An artificial intelligence-based bug detection and classification system, comprising: The system includes an input interface, a DIKWP cognitive processing module, a trajectory monitoring module, a bug detection module, a bug classification and annotation module, an explanation module, and a defect graph database. The DIKWP cognitive processing module is used to divide the artificial intelligence model task processing into five parts: data layer, information layer, knowledge layer, intelligence layer, and intent layer. These parts are executed sequentially, and the semantic results of each layer are output.

7. The system according to claim 6, wherein, The trajectory monitoring module is connected to the DIKWP cognitive processing module and is used to collect and record the semantic state data of each layer when the model passes through the processing layer, generating a complete consciousness trajectory log.

8. The system according to claim 6, wherein, The BUG detection module is used to read the consciousness trajectory log, automatically identify the abnormal patterns in it according to the preset abnormality criteria, and send the determined BUG location and type to the BUG classification and labeling module; the BUG classification and labeling module is used to perform semantic level labeling and type classification of the BUG, ​​and store and file the BUG category and its corresponding DIKWP level label.

9. The system according to claim 6, wherein, The explanation module is used to perform causal analysis on the corresponding consciousness trajectory after receiving a BUG notification, and output the explanation of the cause of the BUG and the associated semantic path.

10. The system according to claim 6, wherein, The defect map database is used to store instance data, classification labels and explanation information of the BUG, ​​and provides query and statistical functions. The modules work together to realize the automatic detection, classification, explanation and storage of artificially conscious BUG as described in any one of claims 1 to 5.