An artificial intelligence content generation compliance management method, device and medium

Abstract

The application discloses a compliance management method and device for artificial intelligence content generation and a medium, and relates to the field of artificial intelligence.The method comprises the following steps: receiving a content generation request sent by a client; acquiring tenant compliance configuration information corresponding to the content generation request; performing pre-compliance verification based on generation task data and the tenant compliance configuration information to obtain a pre-verification result; executing content generation in the case that the pre-verification result represents compliance, and monitoring intermediate content in real time during the generation process, terminating or completing the generation according to a monitoring result to obtain target generation content; performing post-compliance auditing on the target generation content to obtain a compliance level; and determining and executing a disposal strategy based on the compliance level to obtain a disposal result.The application realizes effective compliance management of AI-generated content from demand submission to result output through a whole-link management architecture composed of pre-compliance verification, real-time monitoring during the generation process and post-compliance auditing.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a method, device and medium for compliance management of artificial intelligence content generation. Background Technology

[0002] With the rapid development of AI-generated content technology, AI-generated content such as text, images, audio, and video is being used more and more widely in various application scenarios. However, compliance management has also become a key challenge for platform operations. Due to the diverse forms and complex semantics of generated content, coupled with the high degree of openness and uncertainty of user input, how to effectively manage the compliance of AI-generated content has become a technical challenge of widespread concern in the industry.

[0003] In existing technologies, mainstream AI content generation platforms typically employ two technical approaches for compliance control. One approach involves machine review or manual sampling of the generated results after content generation. This is achieved by using keyword matching or classification models to identify whether the generated text, images, and other outputs contain illegal information. The other approach focuses on basic keyword filtering of the input text when a user submits a request, preventing obviously illegal user requests from entering the generation process. However, both of these approaches have significant limitations in practical applications. First, keyword matching-based filtering is difficult to effectively identify implicit violations disguised through semantic spoofing or variant rewriting, allowing some illegal requests to bypass initial screening and enter the generation stage. Furthermore, post-generation review cannot intervene in a timely manner during content generation. Once illegal content is generated and then intercepted, it not only consumes significant computing resources but also makes it difficult to trace the complete generation chain of illegal outputs afterward due to the lack of record-keeping, complicating the determination of responsibility and rectification of compliance issues. Furthermore, the aforementioned control links are fragmented and lack coordination throughout the entire generation cycle. This makes the platform's control effectiveness highly dependent on human experience and repeated adjustments when facing complex and ever-changing violations, resulting in low efficiency and difficulty in meeting the large-scale control needs of actual operations. Summary of the Invention

[0004] This invention provides a method, device, and medium for compliance management of AI-generated content, aiming to solve the technical problem of how to provide a solution for effective compliance management of AI-generated content.

[0005] In a first aspect, the present invention provides a compliance management method for artificial intelligence content generation, comprising: Receive a content generation request sent by a client. The content generation request includes generation task data and generation parameters. The generation task data includes generation instruction data and reference material data. Obtain the tenant compliance configuration information corresponding to the content generation request. The tenant compliance configuration information is used to define the compliance control rules of the tenant to which the client belongs. Based on the generated task data and the tenant compliance configuration information, a pre-compliance verification is performed on the generated instruction data, the reference material data, and the generated parameters to obtain the pre-verification result. If the pre-verification result indicates compliance, content generation is performed based on the generation task data. During the content generation process, the currently generated intermediate content is monitored in real time. If the intermediate content contains illegal content, the content generation is terminated. If the intermediate content does not contain illegal content, the content generation is completed, and the target generated content is obtained. Perform a post-event compliance audit on the target generated content to obtain a compliance level; Based on the compliance level, a handling strategy is determined and executed to obtain the handling result.

[0006] Optionally, the step of performing pre-compliance verification on the generation instruction data, the reference material data, and the generation parameters based on the generation task data and the tenant compliance configuration information to obtain the pre-verification result includes: Based on the tenant-defined sensitive content thesaurus in the tenant compliance configuration information, keyword detection is performed on the generated instruction data to obtain keyword detection results; The generated instruction data is input into a semantic recognition model, and the implicit violations in the generated instruction data are identified based on the output of the semantic recognition model to obtain semantic detection results; If the generated instruction data corresponds to a preset industry, then an industry-specific compliance verification is performed based on the industry compliance template in the tenant compliance configuration information to obtain an industry-specific verification result; Perform a material compliance check on the reference material data to obtain the material check result; Perform compliance verification on the generated parameters to obtain the parameter verification results; Specifically, if the generated instruction data falls within a preset whitelist range, the keyword detection result and the semantic detection result are considered compliant; if the generated parameters fall within a whitelist range, the parameter verification result is considered compliant; if the generated instruction data does not correspond to a preset industry, the industry-specific verification result is considered compliant. Based on the keyword detection results, the semantic detection results, the industry-specific verification results, the material verification results, and the parameter verification results, the pre-verification results are determined.

[0007] Optionally, the step of monitoring the currently generated intermediate content in real time during the content generation process, and terminating content generation if the intermediate content contains illegal content, and completing content generation to obtain the target generated content if the intermediate content does not contain illegal content, includes: During the content generation process, the currently generated intermediate content is obtained; Perform real-time compliance checks on the intermediate content to determine whether it contains any illegal content. If the intermediate content contains illegal content, then content generation will be terminated; If the intermediate content does not contain any illegal content, the content generation process continues until it is complete, and the target generated content is obtained.

[0008] Optionally, the post-processing compliance audit of the target generated content to obtain a compliance level includes: The target generated content is subjected to a preliminary machine review to obtain the preliminary machine review result. If the machine's initial review result indicates that the review is passed, the target generated content will be delivered to the client, and the compliance level of the target generated content will be determined as compliant. If the machine's initial review result indicates that the review is unsuccessful, the target generated content will be transferred to manual review. Compliance management personnel will make a final judgment based on the complete generation information of the target generated content and the machine's initial review result to obtain the manual review result. If the manual review result indicates a violation, the compliance level of the target generated content will be determined as medium-risk or high-risk based on the severity of the violation. If the manual review result indicates compliance, the interception of the target generated content will be lifted, the target generated content will be delivered to the client, and the compliance level of the target generated content will be determined as compliant. The results of the manual review are fed back to the machine review model to optimize the review accuracy of the machine review model; The initial machine review includes: if the target generated content is text data, keyword detection and semantic compliance detection are performed, and the text data is checked for newly added illegal content; if the target generated content is image data or video data, the presence of pornographic, vulgar, violent, terrorist, politically sensitive, infringing portrait, and inappropriate content is detected, as well as hidden violations and borderline content are detected; if the target generated content is audio data, the audio data is converted into text and then a text review is performed, and illegal content and inappropriate speech in the audio data are detected.

[0009] Optionally, obtaining the tenant compliance configuration information corresponding to the content generation request includes: Based on the content, generate the tenant identifier and tenant level to which the request belongs, and query the tenant compliance configuration information from the pre-stored tenant compliance configuration library; The tenant compliance configuration information includes the tenant's custom sensitive content thesaurus, industry compliance templates, whitelist rules, and control severity level; The control levels include high-level control and standard-level control. High-level control corresponds to tenants in the government and financial sectors, while standard-level control corresponds to ordinary tenants.

[0010] Optionally, determining and executing a handling strategy based on the compliance level to obtain a handling result includes: If the compliance level is compliant, the target generated content will be delivered to the client normally. If the compliance level is low risk, then a risk warning message is generated and the target generated content is delivered to the client along with the risk warning message; If the compliance level is medium-risk or high-risk, the target generated content is intercepted. The submitting account information in the content generation request, the number of violations by the account corresponding to the content generation request, and the severity of the violation of the target generated content are obtained. Based on the severity and the number of violations, the violation is classified into Level 1, Level 2, Level 3, and Level 4 violations. For Level 1 violations, the target generated content is intercepted, the violation is recorded, and a violation reminder is generated. For Level 2 violations, the target generated content is intercepted, the points for this generation task are frozen, a written warning is generated, and some high-risk functions of the client are restricted. For Level 3 violations, the target generated content is intercepted, account points are frozen, the account generation function is temporarily banned, and a compliance commitment letter signing requirement is generated. For Level 4 violations, the account is permanently banned, all points are frozen, and evidence of the violation is retained.

[0011] Optionally, after classifying violations into Level 1, Level 2, Level 3, and Level 4 violations based on their severity and frequency, the method further includes: Based on the submitted account information, a violation account profile is created for the account corresponding to the content request generated by the submitted account information. The violation account profile includes the number of violations, the type of violation, and the severity of the violation. Based on the profile of the violating account, the control level of the account corresponding to the content generation request is dynamically adjusted. The dynamic adjustment includes increasing the frequency of review and restricting high-risk functions. If multiple accounts of the tenant corresponding to the content generation request commit batch violations or serious violations within a preset time window, then overall control will be implemented on the tenant. The overall control includes raising the tenant's control level, restricting the tenant's functions, freezing the tenant's points, and notifying the tenant administrator.

[0012] Optionally, the method further includes: Record a full-link compliance log, which includes the submitter identifier of the client, the tenant identifier corresponding to the content generation request, the submission time, the generation instruction data, the reference material data, the generation parameters, the model identifier of the content generation model used to perform content generation, the pre-verification result, the monitoring record of the content generation process, the review record of the post-compliance review, the violation content, the handling measures, and the handling result; The end-to-end compliance logs are processed to be immutable and permanently retained, resulting in solidified end-to-end compliance logs; In response to an audit query request, the system retrieves the solidified end-to-end compliance logs according to the query criteria, generates a compliance audit report based on the search results, and outputs the compliance audit report to the client. In response to a source tracing request for specific illegal content, based on the task identifier corresponding to the specific illegal content, the entire process record of the generation of the specific illegal content is traced from the solidified full-link compliance log, and the traced entire process record of the generation is exported as evidence. Based on the solidified full-chain compliance logs, violation data is statistically analyzed, a compliance analysis report is generated based on the violation data, high-risk points are identified based on the compliance analysis report, and the control strategies corresponding to the high-risk points are adjusted.

[0013] Secondly, the present invention also provides a computer device including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the above-described method.

[0014] Thirdly, the present invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, can implement the above-described method.

[0015] This invention provides a method, device, and medium for compliance management of AI-generated content. By receiving content generation requests and obtaining tenant compliance configuration information, it performs pre-process compliance checks on generation instruction data, reference material data, and generation parameters, screening user-submitted content and parameters at the source to prevent non-compliant requests from entering the generation stage. Furthermore, it monitors intermediate content in real time during content generation, completing generation and obtaining the target content only when no non-compliant content is detected, terminating the task before non-compliant content is fully generated, further reducing computational power consumption. Finally, it performs post-process compliance review of the target generated content and determines a disposal strategy based on the compliance level, achieving closed-loop compliance interception of the final output. These three stages—pre-process, during-process, and post-process—form a full-link management system from request submission to result output, effectively solving the problem of effective compliance management of AI-generated content. Attached Figure Description

[0016] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 A flowchart illustrating a compliance control method for artificial intelligence content generation provided in an embodiment of the present invention; Figure 2 This is a schematic block diagram of a computer device provided in an embodiment of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] It should be understood that, when used in this specification and the appended claims, the terms "comprising" and "including" indicate the presence of the described features, integrals, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.

[0020] It should also be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0021] It should also be further understood that the term "and / or" as used in this specification and the appended claims refers to any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0022] As used in this specification and the appended claims, the term "if" may be interpreted, depending on the context, as "when," "once," "in response to determination," or "in response to detection." Similarly, the phrase "if determined" or "if [described condition or event] is detected" may be interpreted, depending on the context, as "once determined," "in response to determination," "once [described condition or event] is detected," or "in response to detection of [described condition or event]."

[0023] Please see Figure 1 This invention provides a compliance control method for AI-generated content. This method implements compliance control across the entire content generation process in three stages: pre-generation, during-generation, and post-generation. It constructs a proactive defense and closed-loop management mechanism from request submission to result output, thereby solving problems in existing solutions such as compliance control only covering post-generation review, inability to intercept non-compliant requests at the source, and inability to trace compliance records throughout the entire process. Specifically, the method includes the following steps: S1, Receive a content generation request sent by the client. The content generation request includes generation task data and generation parameters. The generation task data includes generation instruction data and reference material data.

[0024] In practice, the server first receives a content generation request from the client. The client is the user's terminal device. The user inputs text information, such as prompts, spoken text, or script content, through the client's human-computer interaction interface to guide the AI ​​model's generation direction. This information is collectively referred to as generation instruction data. Furthermore, the user can also upload source files such as images, videos, audio, or documents through the client; these source files are collectively referred to as reference source data. Furthermore, the user can configure various parameters on the client to control the content generation process, such as specific character names or style parameters for the character generation model; these parameters are collectively referred to as generation parameters. Furthermore, the client packages the aforementioned generation instruction data, reference source data, and generation parameters into a content generation request and sends it to the server over the network. Furthermore, after receiving the content generation request, the server extracts generation task data and generation parameters from it. The generation task data includes generation instruction data and reference source data, which are temporarily stored in local memory to prepare for subsequent compliance verification processes.

[0025] S2, obtain the tenant compliance configuration information corresponding to the content generation request. The tenant compliance configuration information is used to define the compliance control rules of the tenant to which the client belongs.

[0026] In practice, after receiving a content generation request, the server further obtains the tenant compliance configuration information corresponding to that request. This tenant compliance configuration information is a set of compliance rules pre-configured independently for each tenant, defining the compliance control standards for the tenant's account on the platform. Furthermore, the server queries the database for the tenant's corresponding compliance configuration information based on the tenant identifier carried in the content generation request. This configuration information includes at least the tenant's self-maintained sensitive content thesaurus, the compliance template corresponding to their industry, a whitelist allowing exemptions from specific verification rules, and the level of control strictness. Finally, the server uses the obtained tenant compliance configuration information as the basis for subsequent compliance verification.

[0027] In some preferred embodiments, obtaining the tenant compliance configuration information corresponding to the content generation request includes: querying the tenant compliance configuration information from a pre-stored tenant compliance configuration library based on the tenant identifier and tenant level to which the content generation request belongs; the tenant compliance configuration information includes a tenant-defined sensitive content thesaurus, industry compliance templates, whitelist rules, and control strictness level; the control strictness level includes high-level control and standard-level control, the high-level control corresponding to tenants in the government and financial industries, and the standard-level control corresponding to ordinary tenants.

[0028] In practice, the server retrieves the compliance configuration information of the corresponding tenant from the tenant compliance configuration library by querying a combination of tenant identifier and tenant level.

[0029] First, upon receiving a content generation request, the server extracts the tenant identifier from the request. This tenant identifier is a unique identifier for each tenant, typically a tenant ID assigned by the platform. Next, the server retrieves the tenant level corresponding to that tenant from the tenant management database. This tenant level is pre-set by the platform based on factors such as the tenant's reputation, historical compliance performance, and business scale; for example, a high-level tenant or a standard-level tenant. Finally, the server uses the tenant identifier and tenant level as search criteria to query the corresponding compliance configuration information for that tenant from a pre-built tenant compliance configuration database. This tenant compliance configuration database is a relational database or key-value store, storing each tenant's compliance configuration records using the tenant identifier as the primary key.

[0030] Furthermore, the server retrieves tenant compliance configuration information from the query results. This tenant compliance configuration information includes a tenant-defined sensitive content thesaurus, industry compliance templates, whitelist rules, and a control severity level. The tenant-defined sensitive content thesaurus is a set of sensitive keywords maintained by the tenant based on its own business characteristics; its content can be dynamically added, deleted, or modified by the tenant administrator. The industry compliance templates are a set of compliance rules preset by the platform for different industries; tenants can directly select the corresponding template based on their industry. The whitelist rules are a list of keywords, materials, or accounts configured by the tenant that can be exempted from specific compliance checks. The control severity level is used to control the intensity of compliance control over the tenant.

[0031] Furthermore, the server determines the compliance control strategy based on the level of control stringency. The stringency levels include high-level control and standard-level control. High-level control applies to tenants in the government and financial sectors, which have extremely high compliance requirements and must undergo compliance checks according to the strictest standards. Standard-level control applies to ordinary tenants, who undergo checks according to standard compliance standards. The server automatically matches the corresponding compliance control strategy based on the tenant's level of control stringency and executes that strategy in subsequent pre-verification, in-process monitoring, and post-audit stages.

[0032] This embodiment achieves efficient acquisition and dynamic matching of tenant compliance configuration information through a combined query method using tenant identifiers and tenant levels. The tenant compliance configuration information includes multiple configuration items such as a tenant-defined sensitive content thesaurus, industry compliance templates, whitelist rules, and control severity levels, enabling different tenants to personalize their configurations based on their own business characteristics and compliance needs. Furthermore, the control severity level is automatically matched to either a high-level or standard-level control based on the tenant's industry, ensuring that tenants in sectors with high compliance requirements, such as government and finance, receive stricter control protection, while ordinary tenants can enjoy a more flexible user experience while meeting basic compliance requirements. This differentiated configuration mechanism based on tenant level and industry attributes effectively solves the problem of traditional solutions' one-size-fits-all compliance control rules, which fail to meet personalized needs.

[0033] S3. Based on the generated task data and the tenant compliance configuration information, perform pre-compliance verification on the generated instruction data, the reference material data, and the generated parameters to obtain the pre-verification result.

[0034] In practice, the server performs pre-compliance verification based on the aforementioned generated task data, generated parameters, and tenant compliance configuration information. This pre-compliance verification is a comprehensive compliance review conducted before the user formally submits the content generation task, covering all information input by the user, including generation instruction data, reference material data, and generation parameters. Furthermore, through this review, the server can determine whether the content generation request submitted by the user has any compliance risks and generate a pre-compliance verification result accordingly. The pre-compliance verification result indicates whether the request meets compliance requirements before entering the content generation process.

[0035] In some preferred embodiments, the step of performing pre-compliance verification on the generation instruction data, the reference material data, and the generation parameters based on the generation task data and the tenant compliance configuration information to obtain pre-verification results includes: performing keyword detection on the generation instruction data based on the tenant-defined sensitive content thesaurus in the tenant compliance configuration information to obtain keyword detection results; inputting the generation instruction data into a semantic recognition model, identifying hidden violations in the generation instruction data based on the output of the semantic recognition model to obtain semantic detection results; and if the generation instruction data corresponds to a preset industry, performing industry-specific compliance verification based on the industry compliance template in the tenant compliance configuration information. Compliance verification is performed to obtain industry-specific verification results; material compliance verification is performed on the reference material data to obtain material verification results; compliance verification is performed on the generated parameters to obtain parameter verification results; wherein, if the generated instruction data belongs to a preset whitelist range, the keyword detection results and the semantic detection results are considered compliant; if the generated parameters belong to a whitelist range, the parameter verification results are considered compliant; if the generated instruction data does not correspond to a preset industry, the industry-specific verification results are considered compliant; based on the keyword detection results, the semantic detection results, the industry-specific verification results, the material verification results, and the parameter verification results, the pre-verification result is determined.

[0036] In practice, the server will conduct compliance reviews of the content and parameters submitted by users from five dimensions, and at the same time, it will combine a whitelist mechanism to flexibly exempt certain verification items for specific users.

[0037] First, the server performs keyword detection on the generated instruction data based on the tenant's custom sensitive content thesaurus in the tenant compliance configuration information. This tenant-customized sensitive content thesaurus is a dynamic set of sensitive keywords maintained by the tenant administrator according to their business needs. This thesaurus includes at least categories such as politically sensitive words, pornographic and vulgar words, violent and terrorist words, and illegal and non-compliant words. Further, the server performs word-by-word matching between the text content in the generated instruction data and the sensitive keywords in the tenant-customized sensitive content thesaurus. Matching methods can use string matching algorithms or regular expression matching algorithms. If a match is found, the matched keyword and its position in the text are recorded, forming a keyword detection result. This result is used to characterize whether there are obvious non-compliant keywords in the generated instruction data.

[0038] Furthermore, the server inputs the generated instruction data into the semantic recognition model. Based on the output of the semantic recognition model, it identifies implicit violations in the generated instruction data and obtains a semantic detection result. The implicit violations refer to violations expressed through variations, homophones, metaphors, or other obscure means. The semantic recognition model is a pre-trained language model trained on a large corpus of compliant text. Its core function is to perform deep semantic understanding of the input text and identify the aforementioned implicit violations. The server inputs the generated instruction data into this model, which outputs a semantic violation score. This score reflects the probability that the generated instruction data contains implicit violations. The server obtains a semantic detection result based on this score. This result characterizes whether the generated instruction data contains implicit violations that cannot be detected through keyword matching. Keyword detection focuses on explicit violation identification at the literal level, while the semantic recognition model focuses on implicit violation identification at the semantic level. The two complement each other, jointly improving the coverage and accuracy of pre-compliance verification.

[0039] Further, the server then determines whether the generated instruction data corresponds to a preset industry. These preset industries include at least the financial, education, government, and e-commerce industries. If the generated instruction data corresponds to a preset industry, the server reads the corresponding compliance template from the tenant's compliance configuration information. The industry compliance template is a pre-set set of exclusive compliance rules for a specific industry; for example, the financial industry's compliance template includes rules prohibiting illegal marketing and false advertising, while the education industry's compliance template includes rules prohibiting illegal teaching content. The server performs industry-specific compliance verification on the generated instruction data according to the rules in the template, obtaining an industry-specific verification result. This result indicates whether the generated instruction data contains industry-specific violations. If the generated instruction data does not correspond to a preset industry, the server does not perform industry-specific compliance verification and directly considers this step as passed.

[0040] Furthermore, the server performs compliance checks on the reference material data. This reference material data may include various types of files such as images, videos, audio, and documents. The server performs corresponding checks for each type. For image or video materials, the server uses image recognition algorithms to analyze the content and detect the presence of illegal content, sensitive figures, or infringing portraits. For audio materials, the server first uses speech recognition technology to convert the audio into text, then performs text compliance checks on the converted text, and simultaneously uses audio feature analysis technology to detect the presence of illegal background noise or sound effects. For document materials, the server first parses the document content, extracts the full text, and then performs text compliance checks on the entire text. The results of all these checks are then aggregated to form the material verification results.

[0041] Furthermore, the server performs compliance checks on the generated parameters. These parameters are a set configured by the user to control the AI ​​model's generation behavior. The server performs compliance checks on these parameters, focusing on whether they contain parameters related to sensitive or infringing individuals. For example, in a person generation model, if the user configures unauthorized celebrity portrait parameters or the names of sensitive political figures, the server determines that the generated parameter is in violation and records it as a parameter verification result.

[0042] Furthermore, the server utilizes whitelist rules to handle exemptions for certain verification items. The whitelist includes whitelisted keywords, whitelisted materials, and whitelisted accounts. If the generated instruction data falls within the whitelist, the keyword detection and semantic detection results are considered compliant, meaning keyword detection and semantic recognition are exempted for the generated instruction data. If the generated parameters fall within the whitelist, the parameter verification results are considered compliant, meaning compliance verification is exempted for the generated parameters. If the generated instruction data does not correspond to a preset industry, the industry-specific verification results are considered compliant, meaning industry-specific verification is not performed. This whitelist mechanism allows compliant specific users to avoid unnecessary duplicate checks, improving the user experience without compromising overall compliance requirements.

[0043] Furthermore, the server ultimately determines the pre-verification result based on keyword detection results, semantic detection results, industry-specific verification results, material verification results, and parameter verification results. Specifically, if all of the above verification results indicate compliance, the server determines the pre-verification result to be compliant; if any one of the verification results indicates non-compliance, the server determines the pre-verification result to be non-compliant. This pre-verification result is directly used to decide whether to allow entry into the content generation stage.

[0044] This embodiment constructs a comprehensive screening mechanism from explicit to implicit violations by performing dual verification of generated instruction data using keyword detection and semantic recognition. Keyword detection can quickly locate obvious violations at the literal level, while the semantic recognition model can deeply identify implicit violations after semantic camouflage or variant rewriting, overcoming the limitations of simply relying on keyword matching. Furthermore, industry-specific verification provides differentiated management rules for tenants in different industries through industry compliance templates, specifically meeting the compliance needs of industries such as finance and education, and improving the industry adaptability of compliance management. Furthermore, material compliance verification provides a unified compliance review process for multiple types of materials such as images, videos, audio, and documents, ensuring comprehensive coverage of the compliance of multimodal input files. Furthermore, the whitelist exemption mechanism allows compliant specific users to be exempt from unnecessary duplicate checks, optimizing the user experience without compromising compliance requirements.

[0045] It should be noted that, in one embodiment, the semantic recognition model employs a pre-trained language model based on the Transformer architecture. Its core function is to perform deep semantic understanding of the input text and identify the aforementioned implicit violations. During the training phase, the semantic recognition model uses a compliant sample dataset labeled with implicit violations as training samples, employs the cross-entropy loss function as the optimization objective, and iteratively optimizes the model's parameters through a backpropagation algorithm to ensure that its output accurately reflects the probability score of the input text containing implicit violations. The compliant sample dataset at least includes implicit violation text constructed through variant rewriting, homophonic substitution, and metaphorical expressions.

[0046] S4, if the pre-verification result indicates compliance, content generation is performed based on the generation task data, and the intermediate content being generated is monitored in real time during the content generation process. If the intermediate content contains illegal content, the content generation is terminated; if the intermediate content does not contain illegal content, the content generation is completed, and the target generated content is obtained.

[0047] In practice, assuming compliance is indicated by the pre-verification results, the server generates content based on the generated task data. This step is only triggered after the user-submitted content and parameters pass the pre-compliant verification. The server inputs the generated task data into the AI ​​content generation model to initiate the content generation process. During content generation, the server monitors the currently generated intermediate content in real time. The intermediate content refers to the incomplete content fragments output by the AI ​​model during the generation process. Furthermore, the server performs real-time compliance checks on each generated intermediate content to determine whether it contains any illegal content. If illegal content is detected, the server immediately issues a termination command, forcibly stopping the current content generation task to prevent the complete generation of illegal content. If no illegal content is detected during the continuous generation process, the server waits for the AI ​​model to complete the generation of all content and uses the final generated content as the target generated content. This target generated content is the complete content result output by the AI ​​model and will be provided to the subsequent post-compliant review stage.

[0048] In some preferred embodiments, the step of monitoring the currently generated intermediate content in real time during content generation, and terminating content generation if the intermediate content contains illegal content, and completing content generation to obtain the target generated content if the intermediate content does not contain illegal content, includes: during content generation, acquiring the currently generated intermediate content; performing real-time compliance detection on the intermediate content to determine whether the intermediate content contains illegal content; terminating content generation if the intermediate content contains illegal content; and continuing content generation until it is completed to obtain the target generated content if the intermediate content does not contain illegal content.

[0049] In practice, the server continuously monitors the content fragments currently output by the AI ​​model during the content generation process and dynamically decides whether to continue generating based on the detection results, thereby blocking the illegal content before it is fully generated.

[0050] First, after the AI ​​content generation model starts content generation, the server enters real-time monitoring mode. In this mode, the AI ​​model generates one content fragment per iteration. Upon completion of the current iteration, the server immediately reads this fragment from the AI ​​model's output cache, using it as intermediate content. This intermediate content is a portion of the content output by the AI ​​model during the generation process; it does not yet constitute the complete final output, but it is sufficient to reflect the compliance of the generation direction. For example, when the AI ​​model is generating text, each generated segment of text constitutes intermediate content; when the AI ​​model is generating an image, the pixel rendering result of each iteration is the intermediate content. The server continuously reads the currently generated intermediate content from the AI ​​model's output cache.

[0051] Furthermore, the server performs real-time compliance checks on the intermediate content. The real-time compliance checks are essentially the same as those used in pre-compliance verification, including keyword detection and semantic recognition. However, due to the high real-time requirements, the detection efficiency needs to be optimized accordingly. The server quickly matches the intermediate content against a pre-defined real-time sensitive word library, and simultaneously uses a lightweight semantic detection model to quickly scan the intermediate content. After the checks, the server obtains a judgment result: whether the intermediate content contains illegal content. This detection process is repeated after each new intermediate content is generated, ensuring that there are no regulatory blind spots throughout the entire generation process.

[0052] Furthermore, if the detection results indicate that the intermediate content contains illegal content, the server immediately sends a termination command to the AI ​​content generation model, forcibly stopping the current generation task. Termination operations include stopping subsequent iterative calculations, releasing the computing resources occupied by the model, and clearing unfinished output buffers. After terminating the generation task, that content generation task will no longer output any content, preventing illegal content from spreading in any form.

[0053] Furthermore, if the detection results indicate that the intermediate content does not contain any illegal content, the server allows the AI ​​content generation model to continue subsequent generation iterations while continuing the aforementioned acquisition and detection loop. That is, each segment of intermediate content is detected once until the AI ​​model outputs a completion flag. Once the AI ​​model has completed the generation of all content, the server uses the final generated content as the target generated content and outputs it to the subsequent post-compliance review stage.

[0054] This embodiment achieves proactive intervention during the content generation process through real-time monitoring and immediate termination mechanisms. Compared to the passive mode of traditional solutions that can only conduct reviews after the content is fully generated, this embodiment can terminate the process before the illegal content is fully generated, reducing the spread of illegal content and the consumption of computing power from the source.

[0055] This embodiment achieves proactive intervention before non-compliant content is fully generated by acquiring the intermediate content currently generated by the AI ​​model in real time during the content generation process and performing real-time compliance checks on the intermediate content. Compared to the passive mode of traditional solutions that only implement review after the content is fully generated, this embodiment can dynamically decide whether to continue generation during the generation process. Once non-compliant content is detected in the intermediate content, the generation task is immediately terminated, thus blocking the spread of non-compliant content before it is fully generated, reducing the scope of non-compliant content spread and computational consumption. Furthermore, for post-event accountability, this embodiment retains monitoring records of the intermediate content during the generation process, providing a complete chain of evidence for subsequent compliance audits and violation tracing.

[0056] S5, perform a post-compliance audit on the target generated content to obtain the compliance level.

[0057] In practice, the server then performs a post-implementation compliance audit on the generated content. This audit is a final compliance check on the completed content. The server inputs the generated content into a machine audit model, which automatically performs multi-dimensional detection and determines the compliance level of the generated content as one of four categories: compliant, low-risk, medium-risk, or high-risk, based on the detection results. Furthermore, the compliance level is determined based on the actual compliance status of the generated content. For example, it is classified as compliant if fully compliant, low-risk if there are minor potential issues but do not meet the violation standards, and medium-risk or high-risk if there are clear violations, depending on the severity of the violation.

[0058] In some preferred embodiments, the post-compliance review of the target generated content to obtain a compliance level includes: performing a machine-based initial review of the target generated content to obtain a machine-based initial review result; if the machine-based initial review result indicates that the review is passed, the target generated content is delivered to the client, and the compliance level of the target generated content is determined to be compliant; if the machine-based initial review result indicates that the review is failed, the target generated content is transferred to manual review, whereby compliance management personnel make a final judgment based on the complete generation information of the target generated content and the machine-based initial review result to obtain a manual review result; if the manual review result indicates a violation, the compliance level of the target generated content is determined to be a medium-risk level or a high-risk level according to the severity of the violation; if the manual review result indicates compliance, the review is lifted. The interception of the target generated content involves delivering the target generated content to the client and determining its compliance level as compliant. The manual review results are then fed back to the machine review model to optimize its accuracy. The initial machine review includes: if the target generated content is text data, keyword detection and semantic compliance detection are performed, and the text data is checked for any newly added illegal content; if the target generated content is image or video data, it is detected for pornographic, vulgar, violent, terrorist, politically sensitive, infringing portrait, and misleading content, and hidden violations and borderline content are detected; if the target generated content is audio data, the audio data is converted to text and then undergoes text review, and illegal content and misleading speech are detected in the audio data.

[0059] In practice, the server divides post-review into two levels: machine initial review and manual review, and establishes a feedback optimization mechanism from the results of manual review to the machine review model.

[0060] First, the server performs a preliminary machine review of the target generated content. This preliminary review is automatically executed by a machine review model, which is a machine learning model trained on a large corpus of compliant language and human review samples. Its input is the target generated content, and its output is a pass / fail result. The specific review dimensions of the preliminary machine review vary depending on the data type of the target generated content. Further, if the target generated content is text data, the preliminary machine review will perform keyword detection and semantic compliance detection, while also verifying whether there is any new illegal content in the text data that was not detected in the pre-verification stage. Further, if the target generated content is image or video data, the preliminary machine review will detect whether there is any pornographic, vulgar, violent, terrorist, politically sensitive, infringing portrait, or inappropriate content, while using image recognition algorithms to detect hidden violations and borderline content. Further, if the target generated content is audio data, the preliminary machine review first converts the audio data into text using speech recognition technology, then performs text review, while simultaneously detecting illegal content and inappropriate speech in the audio data. Furthermore, after the machine's initial review is completed, the server receives the initial review result, which represents the machine review model's preliminary judgment on whether the target generated content is compliant.

[0061] Furthermore, if the machine's initial review result indicates that the review is passed, the server considers that the target generated content has met the compliance requirements at the machine review level and directly delivers the target generated content to the client. At the same time, the server determines the compliance level of the target generated content to be compliant.

[0062] Furthermore, if the initial machine review indicates failure, the server will transfer the target generated content to manual review. The manual review channel will present the target generated content, its complete generation information (including generation instruction data, reference material data, generation parameters, and the model used), and the initial machine review result to the compliance management personnel. The compliance management personnel will then review all the above information and, based on their understanding of compliance rules and practical experience, make a final judgment on the compliance of the target generated content. After the judgment is completed, the compliance management personnel will enter the judgment result on the operation interface to obtain the manual review result. If the manual review result indicates a violation, the compliance management personnel will conduct further assessment based on the severity of the violation: if the violation is minor, such as involving only low-level sensitive words but without substantial impact, the compliance level of the target generated content will be determined as medium risk; if the violation is serious, such as involving politically sensitive or infringing content, the compliance level of the target generated content will be determined as high risk. Furthermore, if the manual review result indicates compliance, it means that the machine's initial review may have made a mistake. The server will then remove the interception of the target generated content, deliver the target generated content to the client, and determine the compliance level of the target generated content as compliant.

[0063] Furthermore, the server compares the results of manual review with the initial machine review, records cases of misjudgment in the initial machine review, and feeds these cases back to the machine review model as training data. By adjusting the parameters of the machine review model or through iterative training, the accuracy of the model's review is improved, making the judgment results of the initial machine review increasingly closer to the true standards of manual review.

[0064] This embodiment organically combines the efficiency advantages of machine review with the accuracy advantages of human review through a two-tiered review mechanism of machine initial review and human review. Machine initial review can quickly filter a large amount of target-generated content at a high processing speed, reducing the review burden on compliance management personnel. Furthermore, human review provides refined processing for complex cases that are difficult for machines to accurately judge, ensuring the accuracy and reliability of the review results. Moreover, the closed-loop feedback mechanism between the two allows the results of human review to optimize the machine review model, continuously improving the accuracy of machine initial review. Over long-term operation, this gradually reduces reliance on human review, forming an intelligent improvement chain from data accumulation to model iteration.

[0065] It should be noted that, in one embodiment, the machine review model includes a text encoder for processing text data and an image encoder for processing image / video data. The machine review model is iteratively trained based on a large number of multi-type compliant samples and human review feedback data. During the training process, positive and negative sample pairs are constructed using the results of human review, and the model parameters are optimized through a comparative learning strategy to gradually improve the model's recognition accuracy for illegal content such as pornography, vulgarity, violence, terrorism, and infringing portraits.

[0066] S6. Based on the compliance level, determine the handling strategy and execute the handling strategy to obtain the handling result.

[0067] In practice, the server-side ultimately determines and executes a handling strategy based on the aforementioned compliance level, resulting in a handling outcome. The handling strategy comprises different pre-defined processing methods according to different compliance levels. For example, content at the compliance level is delivered normally, low-risk content is delivered with a risk warning, and medium- or high-risk content is intercepted and further tiered handling is performed. Furthermore, after executing the above strategies, the server-side obtains the handling outcome, which is the final output of the entire compliance control process.

[0068] This embodiment achieves closed-loop control over the entire AI content generation process through three stages: pre-process verification, in-process monitoring, and post-process review. Pre-process verification prevents unauthorized requests from entering the generation stage before content generation, avoiding unnecessary computational waste; in-process monitoring intervenes in real time during content generation, preventing the complete generation of unauthorized content; and post-process review conducts end-point checks on the final output, ensuring that the content delivered to users is fully compliant. These three stages work together to form a complete compliance control chain from request submission to result output.

[0069] This embodiment constructs a three-stage, end-to-end control architecture encompassing pre-compliance verification, real-time monitoring during generation, and post-compliance review. This expands compliance control from the single post-generation review stage of traditional solutions to the entire process before requirement submission, during generation, and after result output. Pre-compliance verification performs multi-layered compliance screening on user-generated command data, reference material data, and generation parameters before the task enters the AI ​​model, preventing non-compliant content from entering the generation stage and avoiding unnecessary computational waste. Furthermore, during-generation monitoring detects intermediate content during the AI ​​model's output process in real time, terminating the task before non-compliant content is fully generated, further reducing computational waste. Finally, post-compliance review conducts multi-dimensional compliance checks on the final generated content, ensuring that the content delivered to the user is fully compliant. These three stages form a collaborative, closed-loop control system, significantly improving the compliance control efficiency and resource utilization of the AI ​​content generation platform compared to existing solutions that only cover post-compliance review.

[0070] In some preferred embodiments, determining and executing a handling strategy based on the compliance level to obtain a handling result includes: if the compliance level is compliant, then delivering the target generated content normally to the client; if the compliance level is low-risk, then generating a risk warning message and delivering the target generated content to the client along with the risk warning message; if the compliance level is medium-risk or high-risk, then intercepting the target generated content, obtaining the submitting account information in the content generation request, the number of violations by the account corresponding to the content generation request, and the severity of the violation of the target generated content. Based on the severity and frequency of the violation, violations are categorized into Level 1, Level 2, Level 3, and Level 4 violations. For Level 1 violations, the generated content is intercepted, the violation is recorded, and a violation alert is generated. For Level 2 violations, the generated content is intercepted, the points for this generation task are frozen, a written warning is generated, and some high-risk functions of the client are restricted. For Level 3 violations, the generated content is intercepted, account points are frozen, the account generation function is temporarily banned, and a compliance commitment letter signing requirement is generated. For Level 4 violations, the account is permanently banned, all points are frozen, and evidence of the violation is retained.

[0071] In practice, the server will implement corresponding handling methods according to different compliance levels, and will implement graded handling based on the severity and number of violations at medium-risk or high-risk levels.

[0072] Specifically, if the compliance level is compliant, the server will deliver the generated content to the client normally, and the user can receive and use the content without any additional processing. The compliance level represents the lowest risk level among the generated content, indicating that the content is fully compliant.

[0073] Furthermore, if the compliance level is low-risk, the server generates a risk warning message and attaches it to the target generated content before delivering it to the client. The risk warning message is explanatory text about the low-risk content, such as informing the user that the content may pose minor compliance risks in certain scenarios and advising the user to use it with caution. Users can receive the risk warning while viewing the content and decide for themselves whether to use it.

[0074] Furthermore, if the compliance level is medium-risk or high-risk, the server first intercepts the target content, blocking its delivery to the client. Then, the server obtains the submitting account information from the content generation request; this submitting account information is the unique identifier of the user account that initiated the request. The server queries the account's historical violation records based on the submitted account information to obtain the number of violations. Simultaneously, the server assesses the severity of the violation based on the violation type and scope of impact of the target content. The server classifies violations into four levels—Level 1, Level 2, Level 3, and Level 4—according to a preset grading rule, using both severity and frequency as two dimensions. The criteria for each level are as follows: Level 1 violation is a minor violation, such as a first offense with minimal impact; Level 2 violation is a general violation, such as multiple violations or moderately impactful content; Level 3 violation is a serious violation, such as multiple violations with significant impact; and Level 4 violation is an extremely serious violation, such as repeated offenses involving politically sensitive or infringing content.

[0075] Furthermore, for Level 1 violations, the server intercepts the target content, preventing it from entering the client's output stream. At the same time, it records the violation in the compliance log and generates a violation reminder message to notify the user that their submitted content has been blocked, and informs them of the violation type and suggested modifications.

[0076] Furthermore, for Level 2 violations, in addition to intercepting the target generated content and recording the violation, the server also freezes the points consumed by this generation task, generates a written warning message and sends it to the user, and restricts some high-risk functions of the user account, such as restricting the user from using certain specific models or parameter configurations.

[0077] Furthermore, for Level 3 violations, the server will not only block the content but also freeze the user's account points and temporarily ban the account generation function. This will prevent the user from continuing to use the platform to generate content for a certain period of time, and require the user to sign a compliance commitment letter promising not to engage in similar violations again.

[0078] Furthermore, for Level 4 violations, the server will permanently ban the violating account, preventing it from using the platform services again, and freeze all of the account's points. At the same time, the complete record of this violation will be saved as evidence of the violation, which can be used for subsequent legal action or regulatory response.

[0079] This embodiment achieves a complete handling chain from normal delivery to risk alerts and tiered blocking by implementing differentiated handling strategies for different compliance levels. The tiered handling mechanism classifies violations into four levels, from minor to extremely serious, and matches them with progressively stricter handling strategies, ensuring that the severity of penalties matches the degree of violation and reflecting the principle of proportionality. Furthermore, measures such as freezing points, temporary bans, and permanent bans effectively prevent further violations, while written warnings and the requirement to sign compliance commitments strengthen users' awareness of and willingness to comply with compliance rules. Moreover, the complete preservation of evidence of violations provides strong support for subsequent regulatory responses and legal accountability, enhancing the platform's risk response capabilities.

[0080] In some preferred embodiments, after classifying violations into Level 1, Level 2, Level 3, and Level 4 violations based on the severity and frequency of the violations, the method further includes: establishing a violation account profile for the account corresponding to the content generation request based on the submitting account information, the violation account profile including the number of violations, violation type, and violation severity; dynamically adjusting the control level of the account corresponding to the content generation request based on the violation account profile, the dynamic adjustment including increasing the review frequency and restricting high-risk functions; if multiple accounts of the tenant corresponding to the content generation request commit batch violations or serious violations within a preset time window, then implementing overall control over the tenant, the overall control including raising the tenant's control level, restricting tenant functions, freezing tenant points, and notifying the tenant administrator.

[0081] In practice, the server establishes profiles of violating accounts to achieve long-term monitoring and key management of high-risk accounts, and implements systematic handling of batch violations by the entire tenant through tenant-level overall control.

[0082] First, after completing the tiered handling, the server establishes an independent violation account profile data structure for each account, indexed by the submitted account information. This violation account profile includes the number of violations, violation type, and violation severity. The number of violations is the total number of times the account has historically initiated requests to generate inappropriate content. The violation type is the categorization of violations the account has historically committed, such as politically sensitive content, pornography, vulgarity, violence, or terrorism. The violation severity is the severity rating of each violation historically, such as ratings from level one to four. This profile is updated in real-time after each violation occurs to ensure it always reflects the account's latest violation status. The server stores this profile in a compliance database for subsequent management decisions.

[0083] Furthermore, the server periodically evaluates the data in the profiles of violating accounts. When an account's number of violations exceeds a preset threshold, or the type of violation involves a high-risk category, or the severity of the violation accumulates to a specific level, the server automatically upgrades the account's control level. Upgrading the control level manifests as increased review frequency, such as transferring all content generation tasks submitted by the account to manual review, and restricting high-risk functions, such as restricting the account from using certain models or parameter configurations that may generate sensitive content. The purpose of this dynamic adjustment is to implement continuous and focused monitoring of high-risk accounts to prevent repeated violations.

[0084] Furthermore, the server monitors the tenant's overall compliance metrics. If multiple accounts of the same tenant commit batch violations or serious violations within a preset time window, it indicates that there may be systemic compliance issues within that tenant. In response, the server implements overall control measures for the tenant, including: escalating the tenant's control level to a higher level; restricting tenant functionality, such as limiting certain model call limits or parameter configuration capabilities; freezing tenant points, preventing the tenant from using platform services during the rectification period; and notifying the tenant administrator, requiring the administrator to investigate and rectify the tenant's accounts and submit a rectification report to the platform. These overall control measures effectively curb systemic violation risks at the tenant level.

[0085] This embodiment achieves long-term tracking and dynamic control of high-risk accounts through account profiling, enabling the platform to continuously and intensively monitor accounts with repeated violations. The dynamic control mechanism adjusts the control level of accounts in real time based on their violation history, ensuring that high-risk accounts receive the level of supervision commensurate with their risk level, effectively reducing the probability of repeated violations. Furthermore, tenant-level overall control addresses batch violations by multiple accounts within a tenant. By upgrading the overall control of the tenant, tenant administrators proactively intervene to investigate and rectify the issues, elevating the problem from the account level to the tenant management level, fundamentally reducing the risk of systemic violations.

[0086] In some preferred embodiments, the method further includes: recording a full-link compliance log, the full-link compliance log including the client's submitter identifier, the tenant identifier corresponding to the content generation request, the submission time, the generation instruction data, the reference material data, the generation parameters, the model identifier of the content generation model used to perform content generation, the pre-verification result, the monitoring record of the content generation process, the audit record of the post-compliance audit, the violation content, the handling measures, and the handling result; performing tamper-proof and permanent retention processing on the full-link compliance log to obtain a solidified full-link compliance log; responding to an audit query request, according to... The system retrieves the solidified end-to-end compliance logs based on query criteria, generates a compliance audit report based on the search results, and outputs the compliance audit report to the client. In response to a source tracing request for specific violations, the system traces the entire generation process of the specific violation from the solidified end-to-end compliance logs based on the task identifier corresponding to the specific violation, and exports the traced generation process record as evidence. Based on the solidified end-to-end compliance logs, the system compiles violation data, generates a compliance analysis report based on the violation data, identifies high-risk points based on the compliance analysis report, and adjusts the control strategies corresponding to the high-risk points.

[0087] In practice, the server-side constructs a traceable, tamper-proof, and auditable compliance data foundation by fully recording the entire content generation process.

[0088] First, the server collects compliance data in real time at each stage of the content generation process and records it as a full-link compliance log. The data recorded in this log includes: the client's submitter identifier, i.e., the user identity identifier that initiated the content generation request; the tenant identifier corresponding to the content generation request, i.e., the unique identifier of the tenant to which the user belongs; the submission time, i.e., the time the content generation request was initiated; generation instruction data, reference material data, and generation parameters, i.e., all input information submitted by the user for content generation; the model identifier of the content generation model used to execute the content generation, i.e., the unique identifier of the AI ​​model used to generate the target content; the pre-verification result, i.e., the compliance judgment result obtained by the server during the pre-compliance verification stage; the monitoring record of the content generation process, i.e., all records of real-time detection of intermediate content during the content generation process; the review record of the post-compliance review, i.e., all judgment process records of the server during the post-compliance review stage, including machine initial review and manual review; the violating content, i.e., the specific fragments or descriptions of the detected violating content; the handling measures, i.e., various handling methods taken against the violation, such as interception, freezing points, temporary bans, etc.; and the handling result, i.e., the final result after implementing the handling measures. The aforementioned log data is collected, formatted, and stored in the log database by the server-side log collection component.

[0089] Furthermore, the server performs tamper-proof and permanent retention processing on the end-to-end compliance logs, resulting in solidified end-to-end compliance logs. The specific implementation methods for tamper-proof and permanent retention can employ blockchain technology, digital signature technology, or write-protected storage technology. For example, the server can generate a digital digest for each log record and write this digest to a blockchain or distributed ledger, ensuring that any modification to the log will be detected. Alternatively, the server can write the logs to read-only storage media and ensure permanent data preservation through redundant backups. The resulting solidified end-to-end compliance logs possess the technical characteristics of being tamper-proof and permanently retained, serving as authoritative evidence for compliance audits and regulatory verification.

[0090] Furthermore, the server provides an audit query interface. Compliance managers can initiate audit query requests through the client, which include query conditions. These conditions can include one or more of the following dimensions: time (e.g., specifying start and end times), tenant (e.g., specifying a specific tenant ID), account (e.g., specifying a specific account ID), violation type (e.g., specifying a certain type of violation), and handling result (e.g., specifying a specific handling result). The server searches the solidified end-to-end compliance logs based on the query conditions, filtering out all log records that meet the criteria. The server then structures and summarizes the search results, generating a compliance audit report, which is output to the client. This report includes compliance statistics for the query period, a list of specific violation cases, and detailed records.

[0091] Furthermore, the server also provides a source tracing interface. Compliance managers can initiate source tracing requests for specific violations through the client. These requests include a task identifier corresponding to the specific violation. This task identifier is generated when the content generation request is initiated and persists throughout the entire content generation process. Using the task identifier as an index, the server locates the complete record of the task from the solidified end-to-end compliance logs, including all operation records, review records, and the target generated content from the content generation request to the final handling result. The server compiles these records into a booklet and exports them as evidence. This evidence is tamper-proof and can be used to address regulatory audits or legal prosecution.

[0092] Furthermore, the server regularly performs data mining and statistical analysis on the solidified end-to-end compliance logs. The statistically analyzed violation data includes: violation type distribution (the proportion of each type of violation in all violation events); violation account distribution (the frequency of violations for each account); approval rate (the percentage of content generation tasks that pass post-compliance review); and interception rate (the percentage of content generation tasks that are blocked). Based on these statistical results, the server generates a compliance analysis report, which visually displays the platform's compliance status and risk distribution. Compliance management personnel or the server's automated analysis module use this report to identify high-risk points, such as frequent occurrences of a certain type of violation, a surge in the number of violations by a particular account, or an abnormally low approval rate for a particular tenant. The server adjusts its corresponding control strategies for high-risk points, such as including the violation in the priority matching list for keyword detection, increasing the control level of the account, or updating the tenant's sensitive content keyword database, thereby continuously optimizing the effectiveness of compliance control at the operational level.

[0093] This embodiment systematically collects and stores data from the entire process of content generation, from request submission to final disposal, through complete end-to-end compliance logging, giving the platform full compliance traceability capabilities. Furthermore, the immutability and permanent retention of log data endows it with the credibility of regulatory verification evidence, eliminating the possibility of post-event modification or deletion. Furthermore, multi-dimensional audit query functions enable compliance managers to quickly locate compliance data for specific time periods, tenants, or accounts, improving audit efficiency. Furthermore, the source tracing and evidence collection function ensures that accountability investigations for specific violations are well-documented, allowing for one-click export of complete process records, meeting the stringent requirements of regulatory compliance audits. Furthermore, compliance analysis reports, through statistical and visual presentation of violation data at the platform and tenant levels, help compliance managers identify systemic risks, providing data support for the dynamic optimization of control strategies, forming a complete closed loop from data recording to risk identification and strategy optimization.

[0094] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0095] Please see Figure 2 , Figure 2 This is a schematic block diagram of a computer device provided in an embodiment of this application. The computer device 500 can be a server, wherein the server can be a standalone server or a server cluster composed of multiple servers.

[0096] The computer device 500 includes a processor 502, a memory, and a network interface 505 connected via a system bus 501. The memory may include a non-volatile storage medium 503 and internal memory 504.

[0097] The non-volatile storage medium 503 may store an operating system 5031 and a computer program 5032. When the computer program 5032 is executed, it enables the processor 502 to execute a compliance control method for artificial intelligence content generation.

[0098] The processor 502 provides computing and control capabilities to support the operation of the entire computer device 500.

[0099] The internal memory 504 provides an environment for the operation of the computer program 5032 in the non-volatile storage medium 503. When the computer program 5032 is executed by the processor 502, the processor 502 can execute a compliance control method for artificial intelligence content generation.

[0100] The network interface 505 is used for network communication with other devices. Those skilled in the art will understand that the above structure is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device 500 to which the present application is applied. A specific computer device 500 may include more or fewer components than shown in the figures, or combine certain components, or have different component arrangements.

[0101] The processor 502 is used to run a computer program 5032 stored in a memory to implement the steps of a compliance management method for artificial intelligence content generation provided in any of the above method embodiments.

[0102] It should be understood that in the embodiments of this application, the processor 502 may be a central processing unit (CPU), or it may be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor.

[0103] It will be understood by those skilled in the art that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program may be stored in a storage medium, which is a computer-readable storage medium. The computer program is executed by at least one processor in the computer system to implement the process steps of the embodiments of the above methods.

[0104] Therefore, the present invention also provides a storage medium. This storage medium can be a computer-readable storage medium. The storage medium stores a computer program. When executed by a processor, the computer program causes the processor to perform the steps of the compliance control method for artificial intelligence content generation provided in any of the above method embodiments.

[0105] The storage medium is a physical, non-transient storage medium, such as a USB flash drive, external hard drive, read-only memory (ROM), magnetic disk, or optical disk, or any other physical storage medium capable of storing program code. The computer-readable storage medium can be non-volatile or volatile.

[0106] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0107] In the several embodiments provided by this invention, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative. For example, the division of each unit is merely a logical functional division, and there may be other division methods in actual implementation. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed.

[0108] The steps in the method of this invention can be adjusted, merged, or reduced in order according to actual needs. The units in the device of this invention can be merged, divided, or reduced according to actual needs. Furthermore, the functional units in the various embodiments of this invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0109] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, a terminal, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention.

[0110] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0111] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, this invention is also intended to include these modifications and variations as long as they fall within the scope of the claims of this invention and their equivalents.

[0112] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A compliance management method for artificial intelligence content generation, characterized in that, include: Receive a content generation request sent by a client. The content generation request includes generation task data and generation parameters. The generation task data includes generation instruction data and reference material data. Obtain the tenant compliance configuration information corresponding to the content generation request. The tenant compliance configuration information is used to define the compliance control rules of the tenant to which the client belongs. Based on the generated task data and the tenant compliance configuration information, a pre-compliance verification is performed on the generated instruction data, the reference material data, and the generated parameters to obtain the pre-verification result. If the pre-verification result indicates compliance, content generation is performed based on the generation task data. During the content generation process, the currently generated intermediate content is monitored in real time. If the intermediate content contains illegal content, the content generation is terminated. If the intermediate content does not contain illegal content, the content generation is completed, and the target generated content is obtained. Perform a post-event compliance audit on the target generated content to obtain a compliance level; Based on the compliance level, a handling strategy is determined and executed to obtain the handling result.

2. The compliance management method for AI content generation according to claim 1, characterized in that, Based on the generated task data and the tenant compliance configuration information, the pre-compliance verification is performed on the generated instruction data, the reference material data, and the generated parameters to obtain the pre-verification result, including: Based on the tenant-defined sensitive content thesaurus in the tenant compliance configuration information, keyword detection is performed on the generated instruction data to obtain keyword detection results; The generated instruction data is input into a semantic recognition model, and the implicit violations in the generated instruction data are identified based on the output of the semantic recognition model to obtain semantic detection results; If the generated instruction data corresponds to a preset industry, then an industry-specific compliance verification is performed based on the industry compliance template in the tenant compliance configuration information to obtain an industry-specific verification result; Perform a material compliance check on the reference material data to obtain the material check result; Perform compliance verification on the generated parameters to obtain the parameter verification results; Specifically, if the generated instruction data falls within a preset whitelist range, the keyword detection result and the semantic detection result are considered compliant; if the generated parameters fall within a whitelist range, the parameter verification result is considered compliant; if the generated instruction data does not correspond to a preset industry, the industry-specific verification result is considered compliant. Based on the keyword detection results, the semantic detection results, the industry-specific verification results, the material verification results, and the parameter verification results, the pre-verification results are determined.

3. The compliance management method for AI content generation according to claim 1, characterized in that, The process of monitoring the currently generated intermediate content in real time during content generation, terminating content generation if the intermediate content contains illegal content, and completing content generation to obtain the target generated content if the intermediate content does not contain illegal content, includes: During the content generation process, the currently generated intermediate content is obtained; Perform real-time compliance checks on the intermediate content to determine whether it contains any illegal content. If the intermediate content contains illegal content, then content generation will be terminated; If the intermediate content does not contain any illegal content, the content generation process continues until it is complete, and the target generated content is obtained.

4. The compliance management method for AI content generation according to claim 1, characterized in that, The post-processing compliance review of the generated content to obtain a compliance level includes: The target generated content is subjected to a preliminary machine review to obtain the preliminary machine review result. If the machine's initial review result indicates that the review is passed, the target generated content will be delivered to the client, and the compliance level of the target generated content will be determined as compliant. If the machine's initial review result indicates that the review is unsuccessful, the target generated content will be transferred to manual review. Compliance management personnel will make a final judgment based on the complete generation information of the target generated content and the machine's initial review result to obtain the manual review result. If the manual review result indicates a violation, the compliance level of the target generated content will be determined as medium-risk or high-risk based on the severity of the violation. If the manual review result indicates compliance, the interception of the target generated content will be lifted, the target generated content will be delivered to the client, and the compliance level of the target generated content will be determined as compliant. The results of the manual review are fed back to the machine review model to optimize the review accuracy of the machine review model; The initial machine review includes: if the target generated content is text data, keyword detection and semantic compliance detection are performed, and the text data is checked for newly added illegal content; if the target generated content is image data or video data, the presence of pornographic, vulgar, violent, terrorist, politically sensitive, infringing portrait, and inappropriate content is detected, as well as hidden violations and borderline content are detected; if the target generated content is audio data, the audio data is converted into text and then a text review is performed, and illegal content and inappropriate speech in the audio data are detected.

5. The compliance management method for AI content generation according to claim 1, characterized in that, The step of obtaining the tenant compliance configuration information corresponding to the content generation request includes: Based on the content, generate the tenant identifier and tenant level to which the request belongs, and query the tenant compliance configuration information from the pre-stored tenant compliance configuration library; The tenant compliance configuration information includes the tenant's custom sensitive content thesaurus, industry compliance templates, whitelist rules, and control severity level; The control levels include high-level control and standard-level control. High-level control corresponds to tenants in the government and financial sectors, while standard-level control corresponds to ordinary tenants.

6. The compliance management method for AI content generation according to claim 1, characterized in that, The process of determining and executing a handling strategy based on the compliance level to obtain a handling result includes: If the compliance level is compliant, the target generated content will be delivered to the client normally. If the compliance level is low risk, then a risk warning message is generated and the target generated content is delivered to the client along with the risk warning message; If the compliance level is medium-risk or high-risk, the target generated content is intercepted. The submitting account information in the content generation request, the number of violations by the account corresponding to the content generation request, and the severity of the violation of the target generated content are obtained. Based on the severity and the number of violations, the violation is classified into Level 1, Level 2, Level 3, and Level 4 violations. For Level 1 violations, the target generated content is intercepted, the violation is recorded, and a violation reminder is generated. For Level 2 violations, the target generated content is intercepted, the points for this generation task are frozen, a written warning is generated, and some high-risk functions of the client are restricted. For Level 3 violations, the target generated content is intercepted, account points are frozen, the account generation function is temporarily banned, and a compliance commitment letter signing requirement is generated. For Level 4 violations, the account is permanently banned, all points are frozen, and evidence of the violation is retained.

7. The compliance management method for AI content generation according to claim 6, characterized in that, After classifying violations into Level 1, Level 2, Level 3, and Level 4 violations based on their severity and frequency, the method further includes: Based on the submitted account information, a violation account profile is created for the account corresponding to the content request generated by the submitted account information. The violation account profile includes the number of violations, the type of violation, and the severity of the violation. Based on the profile of the violating account, the control level of the account corresponding to the content generation request is dynamically adjusted. The dynamic adjustment includes increasing the frequency of review and restricting high-risk functions. If multiple accounts of the tenant corresponding to the content generation request commit batch violations or serious violations within a preset time window, then overall control will be implemented on the tenant. The overall control includes raising the tenant's control level, restricting the tenant's functions, freezing the tenant's points, and notifying the tenant administrator.

8. The compliance management method for AI content generation according to claim 1, characterized in that, The method further includes: Record a full-link compliance log, which includes the submitter identifier of the client, the tenant identifier corresponding to the content generation request, the submission time, the generation instruction data, the reference material data, the generation parameters, the model identifier of the content generation model used to perform content generation, the pre-verification result, the monitoring record of the content generation process, the review record of the post-compliance review, the violation content, the handling measures, and the handling result; The end-to-end compliance logs are processed to be immutable and permanently retained, resulting in solidified end-to-end compliance logs; In response to an audit query request, the system retrieves the solidified end-to-end compliance logs according to the query criteria, generates a compliance audit report based on the search results, and outputs the compliance audit report to the client. In response to a source tracing request for specific illegal content, based on the task identifier corresponding to the specific illegal content, the entire process record of the generation of the specific illegal content is traced from the solidified full-link compliance log, and the traced entire process record is exported as evidence. Based on the solidified end-to-end compliance logs, violation data is statistically analyzed, a compliance analysis report is generated based on the violation data, high-risk points are identified based on the compliance analysis report, and the control strategies corresponding to the high-risk points are adjusted.

9. A computer device, characterized in that, The computer device includes a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the method as described in any one of claims 1-8.

10. A computer-readable storage medium, characterized in that, The storage medium stores a computer program that, when executed by a processor, can implement the method as described in any one of claims 1-8.

Images