A method and system for fair adjustment of dissemination of generative artificial intelligence based on the principle of distributive justice

Abstract

This application relates to a method and system for regulating fairness in generative artificial intelligence content dissemination based on the principle of distributive justice. The method includes: collecting data such as user identifiers, content category tags, exposure volume, interaction volume, language features, and geographic tags from a generative artificial intelligence content dissemination platform; performing deduplication, noise reduction, and normalization on the collected data to form a standardized sample set; calculating the group average exposure value and the overall average exposure value based on the sample exposure data to generate an equal and free weight factor, and compensating for low-exposure samples with a differential compensation coefficient; randomly desensitizing sensitive attributes and detecting model bias through a veil of ignorance simulation mechanism; calculating the equality index, compensation index, and diversity entropy index to obtain a comprehensive justice index; and merging this index with the basic recommendation score to generate a final recommendation score, achieving content ranking and recommendation. This application improves the fairness and diversity of content dissemination by introducing the principle of distributive justice.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence algorithm governance and fair regulation of information dissemination, and in particular to a generative artificial intelligence dissemination fair regulation method and system based on the principle of distributive justice. Background Technology

[0002] With the rapid development of artificial intelligence (AI) technology, generative AI has gradually become an important technological means in the field of information production and dissemination. Generative AI trains deep learning models on large-scale data to automatically generate content in various forms, including text, images, audio, and video, thereby significantly improving content production efficiency. Currently, this technology is widely used in scenarios such as news writing, short video creation, intelligent customer service, Q&A, content marketing, and social media information generation. Leveraging generative AI technology, platforms can generate large amounts of content in a short time and automatically distribute it through recommendation algorithms, thus achieving efficient information dissemination.

[0003] In existing technologies, generative artificial intelligence is typically combined with content recommendation systems to form a complete content dissemination system. Specifically, platforms first generate a large amount of content using generative models or content creation tools, and then sort and distribute this content using recommendation algorithms. Recommendation algorithms generally use information such as user historical behavior data, click records, browsing time, interaction frequency, and interest tags to personalize content matching, thereby improving the efficiency of users obtaining information. Under this technological model, platforms can achieve accurate recommendations based on user interests, enabling users to quickly obtain information content that meets their needs. At the same time, this model can also effectively improve the platform's user activity and commercial value, and therefore has been widely used in various internet platforms.

[0004] However, with the increasing application of generative artificial intelligence and recommendation algorithms in the field of content dissemination, issues regarding fairness in dissemination have gradually emerged. Firstly, in the current internet platform ecosystem, data and computing resources are often concentrated in the hands of a few large platform companies. These platforms, leveraging their massive user base and technological advantages, can continuously accumulate vast amounts of user behavior data and utilize their powerful computing capabilities to train complex deep learning models. In contrast, small and medium-sized content creators and ordinary users typically lack access to the same level of technical resources, resulting in relatively fewer opportunities for their content to be exposed in recommendation systems. In the long run, this concentration of resources may lead to significant resource monopolies in the content dissemination field, thereby affecting the fairness of the dissemination ecosystem.

[0005] Secondly, algorithmic bias is also a significant problem in existing generative AI dissemination systems. Recommendation algorithms typically rely on historical data for training, which often originates from mainstream cultural contexts or language environments with a large number of speakers. Because mainstream language constitutes a high proportion of the training data, content from less populated cultural groups or less commonly spoken languages ​​has a relatively low representation. Therefore, the model is more likely to prioritize mainstream content during the recommendation process, while recommending niche content with a lower probability. This uneven data distribution can lead to implicit biases in the algorithm during dissemination, thereby affecting the opportunities for expression for different cultural groups in cyberspace.

[0006] Furthermore, in existing recommendation systems, platforms typically use metrics such as click-through rate, viewing time, and interaction rate as primary optimization goals to maximize user engagement and platform revenue. While this optimization strategy can boost platform activity in the short term, it can also easily create a so-called "Matthew effect." Specifically, content that has already gained high exposure is more likely to continue receiving recommendations, while new creators or niche content struggle to get sufficient exposure. Over time, this phenomenon may lead to a gradual concentration of attention resources and economic benefits in a few top-tier content creators, thereby exacerbating the uneven distribution of dissemination resources.

[0007] While existing research on algorithmic fairness has made some progress, most solutions still focus on mathematical optimization. For example, some studies reduce prediction bias between different groups by introducing fairness constraints during model training or adjusting the loss function; others reduce the impact of imbalanced data distribution on model results by resampling or reweighting training data. However, these methods typically only consider statistical fairness and lack a systematic consideration of social and ethical principles. In other words, existing technologies rely more on mathematical models when designing fair algorithms and less on incorporating social fairness concepts or ethical principles into the algorithmic decision-making framework, thus failing to fundamentally solve the fairness problem in the propagation process.

[0008] In the field of social philosophy, the distributive justice theory proposed by American philosopher John Rawls provides an important theoretical foundation for solving the problem of resource allocation. This theory mainly includes the principles of equality and liberty, the difference principle, and the veil of ignorance. The principle of equality and liberty emphasizes that members of society should enjoy basic and equal rights of freedom; the difference principle argues that when inequality exists in resource allocation, the interests of the most vulnerable groups should be prioritized; and the veil of ignorance assumes that people do not know their future social status when formulating social rules, thus enabling the creation of more equitable institutional arrangements. These principles have a significant impact on public policy and social system design, but they have not yet been fully utilized in existing artificial intelligence algorithms.

[0009] Therefore, in generative AI dissemination systems, how to integrate the aforementioned ethical principles into algorithm design, enabling algorithms to not only achieve efficient information distribution but also ensure fairness in dissemination, has become a crucial issue that urgently needs to be addressed in the field of AI governance. Especially in the process of content recommendation and dissemination, if reasonable algorithmic mechanisms can be used to dynamically adjust the exposure opportunities for different groups and detect and correct potential algorithmic biases, it is hoped that dissemination efficiency can be improved while simultaneously mitigating the problem of uneven distribution of dissemination resources.

[0010] To address the aforementioned technologies, it is necessary to propose a new technical solution that embeds the principle of distributive justice into a generative AI dissemination system. By constructing a fair weighting calculation mechanism, a sensitive attribute desensitization mechanism, and a comprehensive fair evaluation mechanism, the content recommendation results can be dynamically adjusted, thereby achieving a fairer and more reasonable allocation of dissemination resources while maintaining algorithm efficiency. Summary of the Invention

[0011] In order to construct a fair weighting calculation mechanism, a sensitive attribute desensitization mechanism, and a comprehensive fairness evaluation mechanism, this application provides a generative artificial intelligence propagation fairness adjustment method and system based on the principle of distributive justice.

[0012] This application provides a generative artificial intelligence propagation fairness adjustment method based on the principle of distributive justice, which adopts the following technical solution: Firstly, a generative artificial intelligence-based method for propagating fairness regulation based on the principle of distributive justice includes the following steps: S1. Data Collection: Obtain content dissemination-related data from the generative AI content dissemination platform. The data includes user identifiers, content category tags, exposure volume, interaction volume, language features, and geographic tags. S2. Feature preprocessing: The collected data is deduplicated, denoised and standardized, and the numerical features are normalized according to the normalization formula to obtain a standardized sample set; S3. Fair weight calculation: Calculate the group average exposure value and the overall average exposure value based on the exposure data in the sample set, and calculate the equal free weight factor based on the average exposure value. At the same time, introduce the difference compensation coefficient to generate difference compensation weight for low exposure samples. S4. Veil of Ignorance Simulation: The sensitive attribute set in the sample data is randomly anonymized to generate a neutral data subset. The model is then trained using the standardized sample set and the neutral data subset, respectively. The difference between the predictions of the two models is calculated. When the difference between the predictions exceeds the preset fairness tolerance threshold, the data is adjusted and retrained. S5. Justice Assessment: Calculate the equality index, compensation index, and diversity entropy index, and sum the three types of indicators according to the preset weight coefficients to obtain the comprehensive justice index; S6. Dynamic readjustment: When the comprehensive justice index is lower than the preset justice threshold, the sample recommendation weights are iteratively updated according to the weight adjustment formula; S7. Recommendation Output: The dynamically readjusted comprehensive recommendation weights are merged with the basic recommendation scores generated by the original recommendation model to obtain the final recommendation score, and the content is sorted and output according to the final recommendation score.

[0013] By adopting the above technical solutions, the multi-dimensional data in the platform's dissemination process is uniformly organized and standardized through data collection and feature preprocessing steps, thus providing a reliable data foundation for subsequent algorithm calculations. Through the fair weight calculation step, equal and free weight factors and differential compensation coefficients are introduced to appropriately compensate content samples with low exposure, making the allocation of dissemination resources more balanced. Through the veil of ignorance simulation step, sensitive attributes are randomly desensitized and model bias detection is performed, effectively identifying and reducing potential biases in the recommendation model regarding language, region, or identity. Through the justice evaluation step, equality indicators, compensation indicators, and diversity entropy indicators are constructed, and a comprehensive justice index is calculated to achieve a quantitative assessment of the overall fairness of the system. When the comprehensive justice index falls below a preset threshold, the recommendation weights are iteratively updated through a dynamic readjustment mechanism, and the adjusted comprehensive weights are merged with the output of the basic recommendation model to generate the final recommendation score and complete the content ranking output.

[0014] Optionally, in the feature preprocessing step, the normalization of numerical features adopts the min-max normalization method: Where x is the original feature value, xmin is the minimum value of the feature in the sample set, and xmax is the maximum value of the feature in the sample set.

[0015] By employing the aforementioned technical solution, the min-max normalization method is used to standardize numerical features during feature preprocessing. This maps raw data with different dimensions and numerical ranges to the same numerical interval, effectively eliminating the impact of scale differences between different features and improving data consistency and comparability. Simultaneously, this processing method enhances the computational stability and convergence efficiency of subsequent algorithm models, enabling the models to more accurately reflect the true relationships between features during training and inference, thereby improving the reliability of overall analysis and recommendation results.

[0016] Optionally, the equal and free weighting factor is calculated by comparing the deviation between the average exposure value of each group and the overall average exposure value. When the average exposure value of a group is lower than the overall average exposure value, the recommendation weight is increased for the corresponding sample.

[0017] By employing the aforementioned technical solution, an equal and free weighting factor is calculated by comparing the deviation between the average exposure value of each group and the overall average exposure value. When the average exposure value of a certain group is lower than the overall average level, the recommendation weight of the corresponding sample of that group is appropriately increased, thereby compensating for content with fewer exposure opportunities during the recommendation process. This mechanism can effectively alleviate the imbalance caused by the excessive concentration of dissemination resources in a few high-exposure groups, promote a more balanced display opportunity for content from different groups on the platform, and improve the fairness and overall rationality of content dissemination.

[0018] Optionally, the difference compensation weight is calculated by introducing a difference compensation coefficient α, the value of which is in the range of 0 < α ≤ 1, and is used to control the weight increase of the low exposure group.

[0019] By adopting the above technical solution, a differential compensation coefficient is introduced to adjust the recommendation weight of low-exposure groups. By limiting the compensation coefficient to the range of 0 < α ≤ 1, the magnitude of weight increase can be reasonably controlled, allowing the algorithm to compensate for the exposure opportunities of less prominent content while avoiding excessive amplification of weight changes. This mechanism can moderately modify the allocation of dissemination resources while ensuring the stability of the recommendation system, increasing the probability of low-exposure group content being recommended, and further promoting the fairness and overall balance of content dissemination on the platform.

[0020] Optionally, in the veil of ignorance simulation step, the set of sensitive attributes includes regional tags, language features, and user identity-related attributes, and the sensitive attributes are desensitized by randomly setting them to empty or replacing them.

[0021] By adopting the above technical solutions, sensitive information such as regional labels, language features, and user identity-related attributes are randomly emptyed or replaced to desensitize the training data. This reduces the direct impact of identity information on recommendation results during model training and decision-making, simulates the algorithm decision-making environment under the condition of "veil of ignorance," effectively reduces the model's dependence on specific group attributes, reduces potential algorithmic bias, and improves the fairness and objectivity of the recommendation system among different groups.

[0022] Optionally, the predicted difference is obtained by calculating the absolute difference between the output of the model trained on the standardized sample set and the output of the model trained on the neutral data subset. When the difference is less than a preset fairness tolerance threshold, it is determined that the model does not have significant bias.

[0023] By adopting the above technical solution, the absolute difference between the output of the model trained on the standardized sample set and the output of the model trained on the neutral data subset is used to quantitatively evaluate the possible bias of the recommendation model under sensitive attribute conditions. When the difference is less than the preset fairness tolerance threshold, it is determined that the model does not have significant bias, ensuring the fairness of the algorithm's decision-making. It can promptly detect and suppress potential biases during the model training stage, improve the objectivity and stability of the recommendation results, and further enhance the system's ability to fairly propagate among different groups.

[0024] Optionally, the comprehensive justice index is calculated by weighting the equality index, the compensation index, and the diversity entropy index: Wherein, β1, β2, and β3 are weighting coefficients and satisfy β1 + β2 + β3 = 1; The diversity entropy index is calculated using Shannon entropy: Where pi is the proportion of the i-th type of language feature in the sample.

[0025] By adopting the above technical solution, the equality index, compensation index, and diversity entropy index are weighted and integrated to construct a comprehensive justice index, which quantitatively evaluates the overall fairness of the content dissemination system. Among them, the diversity index is calculated using Shannon entropy to reflect the degree of balanced distribution of different language or cultural content in the dissemination process. This comprehensive evaluation mechanism can monitor the system's operating status from multiple dimensions such as resource allocation, fair compensation, and content diversity, providing a basis for subsequent weight adjustment, thereby improving the fairness and diversity of the dissemination process.

[0026] Optionally, in the dynamic readjustment step, the recommendation weights are updated using the following formula: Where Wt is the current weight, Wt+1 is the updated weight, γ is the learning rate parameter, and θ is the justice threshold; The final recommendation score is calculated by combining the basic recommendation score with the comprehensive weight, and the recommendations are ranked and recommended based on the final recommendation score.

[0027] By adopting the above technical solution, a dynamic weight update formula is introduced to iteratively adjust the recommendation weights. When the comprehensive justice index is lower than the preset justice threshold, the weights are appropriately corrected according to the learning rate parameter to achieve dynamic adjustment of the recommendation results. At the same time, the updated comprehensive weights are integrated with the basic recommendation scores to generate the final recommendation score and output it in a sorted manner, so that the system can continuously optimize the fairness of propagation and the balance of resource allocation while ensuring recommendation efficiency.

[0028] Secondly, a generative artificial intelligence-based system for regulating fairness based on the principle of distributive justice is characterized by comprising: The data acquisition module is used to obtain content dissemination-related data from the generative AI content dissemination platform; The feature preprocessing module is used to perform deduplication, denoising, and normalization on the data to generate a standardized sample set; The fair weight calculation module is used to calculate the equal free weight factor and the differential compensation weight based on the exposure volume; The Veil of Ignorance simulation module is used to randomly desensitize sensitive attributes and train a neutral data model to detect algorithm bias. The justice assessment module is used to calculate the equality index, the compensation index, and the diversity entropy index, and to generate a comprehensive justice index. The dynamic readjustment module is used to iteratively update the recommendation weights when the comprehensive justice index falls below a preset threshold. The recommendation output module is used to generate the final recommendation score based on the updated recommendation weights and output the ranking results.

[0029] By adopting the above technical solutions, a system structure is constructed consisting of a data acquisition module, a feature preprocessing module, a fair weight calculation module, an ignorance veil simulation module, a justice assessment module, a dynamic readjustment module, and a recommendation output module. This structure enables fair regulation of the entire process of generative AI content dissemination. The data acquisition and feature preprocessing module standardizes the platform's dissemination data, providing a reliable data foundation for subsequent calculations. The fair weight calculation module introduces equal and free weight factors and differential compensation weights to reasonably adjust the weights of content with different exposure levels. The ignorance veil simulation module randomly desensitizes sensitive attributes and trains a control model to detect and suppress algorithmic bias. The justice assessment module generates a comprehensive justice index by calculating equality, compensation, and diversity entropy indicators, evaluating the fairness of dissemination from multiple dimensions. When the assessment result is lower than a preset threshold, the dynamic readjustment module iteratively updates the recommendation weights. Finally, the recommendation output module integrates the updated weights and the basic recommendation score to generate the ranking result.

[0030] Optionally, the system is deployed on a hybrid computing node based on CPU and GPU, and data communication between modules is achieved through REST interfaces or message queues.

[0031] By adopting the above technical solutions, the system is deployed on a hybrid computing node of CPU and GPU, which enables the coordinated acceleration of data processing and model calculation tasks, improves the overall computing efficiency of the system, and realizes data communication and task transfer between modules through REST interfaces or message queues, so that the system has good scalability and stability, thereby ensuring that the system can run stably and efficiently in large-scale data processing scenarios.

[0032] In summary, this application includes at least one of the following beneficial technical effects: This application introduces the principle of distributive justice into a generative artificial intelligence dissemination system, constructs a fair weight calculation mechanism, a sensitive attribute desensitization mechanism, and a comprehensive justice evaluation mechanism, enabling the algorithm to dynamically adjust content dissemination resources during operation, thereby improving the fairness and rationality of platform content allocation while ensuring recommendation efficiency. This application calculates an equal and free weighting factor by comparing the average exposure value of the group with the overall average exposure value, and combines it with a differential compensation coefficient to appropriately increase the weight of low-exposure content, thereby alleviating the problem of excessive concentration of dissemination resources on a few high-exposure contents and increasing the display opportunities of different groups' content on the platform. This application introduces a "veil of ignorance" simulation mechanism to randomly desensitize sensitive attributes such as regional labels, language features, and user identities, and detects algorithm deviations by comparing the output results of the training model. This allows for the identification and reduction of potential algorithmic discrimination or bias during the model training phase, thereby improving the fairness and objectivity of the recommendation system among different groups. This application constructs a comprehensive justice index composed of equality index, compensation index, and diversity entropy index to quantitatively evaluate the fairness of system propagation, and combines a dynamic weight update mechanism to iteratively adjust the recommendation weights, enabling the system to continuously monitor and optimize the propagation structure during operation, thereby improving overall fairness and diversity. This application integrates the fair adjustment mechanism with the output of existing recommendation models to achieve the fair adjustment function without changing the main structure of the original recommendation algorithm. This results in good compatibility and scalability, making it easy to deploy and apply in existing generative artificial intelligence content platforms. This application achieves efficient collaboration and elastic scaling between system modules by adopting a hybrid CPU and GPU computing node and a REST interface or message queue communication architecture, enabling the system to maintain stable and efficient operation when facing large-scale data processing and high-concurrency recommendation scenarios. Attached Figure Description

[0033] Figure 1 This is a schematic diagram of the overall system architecture of an embodiment of this application.

[0034] Figure 2 This is a flowchart of the weighting and justice index calculation process.

[0035] Figure 3 This is a flowchart of a simulation training exercise using the veil of ignorance.

[0036] Figure 4 It is a system feedback closed-loop diagram. Detailed Implementation

[0037] The following is in conjunction with the appendix Figure 1-4 This application will be described in further detail.

[0038] This application discloses a generative artificial intelligence system for promoting fairness based on the principle of distributive justice. (Refer to...) Figure 1 , 1. The overall architecture includes: A data acquisition module for obtaining content dissemination-related data from generative AI content dissemination platforms; A feature preprocessing module for performing deduplication, denoising, and normalization on the data to generate a standardized sample set; The fair weight calculation module is used to calculate the equal free weight factor and the differential compensation weight based on the exposure volume; A veil of ignorance simulation module used to randomly desensitize sensitive attributes and train neutral data models to detect algorithmic biases; A justice assessment module used to calculate equality indicators, compensation indicators, and diversity entropy indicators, and to generate a comprehensive justice index; Used to iteratively update the recommendation weights when the comprehensive justice index falls below a preset threshold; The dynamic readjustment module is a recommendation output module used to generate the final recommendation score and output the ranking result based on the updated recommendation weights. The system is deployed on a hybrid computing node based on CPU and GPU, and data communication between modules is achieved through REST interfaces or message queues; The modules communicate with each other through a data bus interface (API or MQ) to form a closed loop of input → calculation → evaluation → adjustment → output.

[0039] Operating environment: Hardware: CPU / GPU hybrid computing node Software: Linux operating system, Python language environment, TensorFlow / PyTorch framework Network: HTTP REST interface or message queue (MQ) 2. Working principle and process, refer to Figures 2-4 ; (1) Data acquisition and feature preprocessing (①-②) Data collection: Samples are obtained from the AI ​​platform's real-time / historical logs, including the following fields: in: User ID, For category labels, For exposure, For interaction volume, As a linguistic feature, For regional tags.

[0040] Data is deduplicated, noise-reduced, and normalized according to a formula: Form a standardized sample set S′.

[0041] The table below describes the relevant fields and variables; (2) Weight calculation (③) Equal and Freedom Weighting: First calculate the average exposure value for each group. and overall average exposure value : A value close to 1 indicates that the group's exposure is close to the overall average, while a value close to 0 indicates a large gap.

[0042] This option increases the recommendation weight for low-exposure samples.

[0043] The table below describes the relevant fields and variables.

[0044] (3) The Veil of Ignorance Simulation (④) Desensitization: This involves processing the set of sensitive attributes. Randomly empty or replace to generate a neutral subset .

[0045] Deviation detection: separate training and The model calculates the predicted difference in output: in This is the fairness tolerance threshold. If If no significant bias is found, the data is adjusted and the training is repeated.

[0046] The table below describes the relevant fields and variables.

[0047] (4) Comprehensive Justice Index Assessment (⑤) Define three types of sub-indicators: Equality: Compensation degree: Diversity entropy: in For the first Proportion of language-like features.

[0048] Overall Justice Index: parameter With threshold Compare: like : Output result.

[0049] like : Enter dynamic readjustment.

[0050] The table below describes the relevant fields and variables.

[0051] (5) Dynamic readjustment and recommended output (⑥-⑦) Automatic weight adjustment: This is the learning rate.

[0052] The output is sorted and stored on the platform. Recorded in the log for auditing purposes.

[0053] 3. Implementation Results Example In a simulated dataset (500,000 samples), with α=0.8, β1=0.4, β2=0.4, β3=0.2, θ=0.8, γ=0.05, and δ=0.02, after three iterations: The average exposure of vulnerable groups increased by 12%. J increased from 0.73 to 0.86 Click-through rate decreased by <1.5% The table below describes the relevant fields and variables.

[0054] Technical effects of the technical solution in this application: 1. The system automatically adjusts the algorithm weights to achieve the goal of equalizing propagation.

[0055] 2. Identity desensitization significantly reduces the risk of bias based on gender, region, language, etc.

[0056] 3. The Justice Index quantifies fairness and provides a basis for verification in regulation.

[0057] 4. Modular design ensures compatibility with existing recommendation engines, resulting in low deployment costs.

[0058] 5. Log and audit interfaces enhance platform governance capabilities.

[0059] The key technical points of this application's technical solution are: dual-weight algorithm, veil of ignorance random desensitization, and justice index feedback closed loop. The technical protection points of this application's technical solution are: the weight formulas, the index calculation method, the dynamic readjustment function, and the seven-module collaborative architecture.

[0060] Alternative solutions to the technical solution of this application: 1. In scenarios with limited computing power, simplify the justice index to F1+F2; 2. Use batch processing adjustment instead of real-time dynamic readjustment; 3. The desensitization process can be replaced by virtual group simulation.

[0061] The implementation principle of a generative artificial intelligence propagation fairness adjustment method and system based on the principle of distributive justice in this application embodiment is as follows: ......

[0062] The above are all preferred embodiments of this application, and are not intended to limit the scope of protection of this application. Therefore, all equivalent changes made in accordance with the structure, shape and principle of this application should be covered within the scope of protection of this application.

Claims

1. A generative artificial intelligence propagation fairness adjustment method based on the principle of distributive justice, characterized in that, Includes the following steps: S1. Data Collection: Obtain content dissemination-related data from the generative AI content dissemination platform. The data includes user identifiers, content category tags, exposure volume, interaction volume, language features, and geographic tags. S2. Feature preprocessing: The collected data is deduplicated, denoised and standardized, and the numerical features are normalized according to the normalization formula to obtain a standardized sample set; S3. Fair weight calculation: Calculate the group average exposure value and the overall average exposure value based on the exposure data in the sample set, and calculate the equal free weight factor based on the average exposure value. At the same time, introduce the difference compensation coefficient to generate difference compensation weight for low exposure samples. S4. Veil of Ignorance Simulation: The sensitive attribute set in the sample data is randomly anonymized to generate a neutral data subset. The model is then trained using the standardized sample set and the neutral data subset, respectively. The difference between the predictions of the two models is calculated. When the difference between the predictions exceeds the preset fairness tolerance threshold, the data is adjusted and retrained. S5. Justice Assessment: Calculate the equality index, compensation index, and diversity entropy index, and sum the three types of indicators according to the preset weight coefficients to obtain the comprehensive justice index; S6. Dynamic readjustment: When the comprehensive justice index is lower than the preset justice threshold, the sample recommendation weights are iteratively updated according to the weight adjustment formula; S7. Recommendation Output: The dynamically readjusted comprehensive recommendation weights are merged with the basic recommendation scores generated by the original recommendation model to obtain the final recommendation score, and the content is sorted and output according to the final recommendation score.

2. The method according to claim 1, characterized in that: In the feature preprocessing step, the normalization of numerical features adopts the minimum-maximum normalization method: Where x is the original feature value, xmin is the minimum value of the feature in the sample set, and xmax is the maximum value of the feature in the sample set.

3. The method according to claim 1, characterized in that: The equal and free weighting factor is calculated by comparing the deviation between the average exposure value of each group and the overall average exposure value. When the average exposure value of a group is lower than the overall average exposure value, the recommendation weight is increased for the corresponding sample.

4. The method according to claim 1, characterized in that: The differential compensation weight is calculated by introducing a differential compensation coefficient α, which has a value range of 0 < α ≤ 1, and is used to control the weight increase of the low-exposure group.

5. The method according to claim 1, characterized in that: In the veil of ignorance simulation step, the set of sensitive attributes includes regional tags, language features, and user identity-related attributes. The sensitive attributes are desensitized by randomly setting them to empty or replacing them.

6. The method according to claim 1, characterized in that: The predicted difference is obtained by calculating the absolute difference between the output of the model trained on the standardized sample set and the output of the model trained on the neutral data subset. When the difference is less than the preset fair tolerance threshold, it is determined that the model does not have significant bias.

7. The method according to claim 1, characterized in that: The comprehensive justice index is calculated by weighting the equality index, the compensation index, and the diversity entropy index. Wherein, β1, β2, and β3 are weighting coefficients and satisfy β1 + β2 + β3 = 1; The diversity entropy index is calculated using Shannon entropy: Where pi is the proportion of the i-th type of language feature in the sample.

8. The method according to claim 1, characterized in that: In the dynamic readjustment step, the recommendation weights are updated using the following formula: Where Wt is the current weight, Wt+1 is the updated weight, γ is the learning rate parameter, and θ is the justice threshold; The final recommendation score is calculated by combining the basic recommendation score with the comprehensive weight, and the recommendations are ranked and recommended based on the final recommendation score.

9. A generative artificial intelligence propagation fairness adjustment system based on the principle of distributive justice, characterized in that, include: The data acquisition module is used to obtain content dissemination-related data from the generative AI content dissemination platform; The feature preprocessing module is used to perform deduplication, denoising, and normalization on the data to generate a standardized sample set; The fair weight calculation module is used to calculate the equal free weight factor and the differential compensation weight based on the exposure volume; The Veil of Ignorance simulation module is used to randomly desensitize sensitive attributes and train a neutral data model to detect algorithm bias. The justice assessment module is used to calculate the equality index, the compensation index, and the diversity entropy index, and to generate a comprehensive justice index. The dynamic readjustment module is used to iteratively update the recommendation weights when the comprehensive justice index falls below a preset threshold. The recommendation output module is used to generate the final recommendation score based on the updated recommendation weights and output the ranking results.

10. The system according to claim 11, characterized in that: The system is deployed on a hybrid computing node based on CPU and GPU, and data communication between modules is achieved through REST interfaces or message queues.

Images