GB/T 47507-2026Artificial intelligence — Trustworthiness — General rules (English PDF)

5.1 Controllability

Controllability refers to the ability of humans or other external intelligent agents to effectively monitor, intervene in, and regulate the operation of an artificial intelligence system.

The nature of the system is to either suspend or halt its behavior to ensure that it performs its tasks in accordance with predetermined objectives, safety, ethical, and legal requirements. Controllability conforms to... The following requirements are made.

a) Controllability-related control points (interfaces) should be designed and implemented. These interfaces should have clearly defined functions and support verification mechanisms and good performance. Good scalability;

b) There should be a system state definition, and the state should be determined based on the stability and dwell time dimension (such as time domain) and range required by the application.

c) It should have a control point management and organization mechanism to form a control surface and map the control surface to specific control operations;

d) Before operation, a control surface reachability detection mechanism should be in place to check at least the time and space reachability status;

e) External control entry and exit should be supported, including cost estimates for entry and exit, as well as compensation control for entry and exit failures.

The case was filed, and a control point handover mechanism was provided;

f) A default control policy should be available to be used when control fails or a control policy cannot be formed;

g) Applications in government departments, critical information infrastructure, and other fields that directly affect public safety and the life and health of citizens.

When dealing with key areas, efficient and precise emergency control measures should be in place;

h) For systems that can be represented by finite states, it is advisable to identify and process all possible sets of states of the system;

i) For systems with an infinite number of states that can be represented, it is advisable to identify and process the set of states that are meaningful to the system's business lifecycle.

5.2 Security

Security emphasizes the application of artificial intelligence systems in security-related functions, especially in the face of potential failures, anomalies, or uncertainties.

The ability of a system to maintain safe operation during qualitative analysis lies in reducing the risk of damage to the artificial intelligence system through system design, development, and verification processes.

Risks arising from system malfunctions or errors. Safety must meet the following requirements.

a) It should possess the capabilities for risk identification and event prediction, event prevention mechanisms, event detection and diagnosis functions, and emergency response. measure;

b) Conduct security audits on external frameworks relied upon by artificial intelligence algorithms, such as host security, behavior monitoring, and traffic analysis;

Note: External frameworks refer to the tools, libraries, platforms, or technical standards that support the operation and development of AI algorithms.

c) Security testing should be conducted on data, models, frameworks, etc.

d) A clear mechanism for tracking software and hardware product vulnerabilities and defects should be established, including regular scanning, tiered assessment, and prioritization of remediation.

The process should be followed, and repair and reinforcement measures should be taken promptly.

e) User access control should be implemented in the management of artificial intelligence systems.

f) Technological solutions should be developed to mitigate broader socio-technical risks and prevent potential systemic harms;

g) An effective monitoring mechanism should be established, including detecting and identifying malicious use, monitoring the model's predictive capabilities in real time, and continuously monitoring the model. Unexpected performance and abilities;

h) Service provision boundaries should be set according to the actual application scenarios of users, and functions of artificial intelligence systems that may be abused should be reduced;

i) It is advisable to establish a risk level testing and evaluation system to classify and grade artificial intelligence systems based on their functions, performance, application scenarios, etc.

j) A risk management plan should be established based on the potential security risks the system may face, and kept dynamically updated to address potential security threats. threaten;

k) It is advisable to manage the use of artificial intelligence systems and clarify the application requirements of artificial intelligence technology for specific groups and scenarios;

l) The impact on the surrounding environment should be considered to reduce incidental property and personnel losses when the incident occurs.

5.3 Robustness

Robustness refers to the ability of an artificial intelligence system to maintain stable performance and accurate output when faced with maliciously crafted inputs (such as adversarial examples).

The robustness of an artificial intelligence system includes the following properties.

a) Robustness of the system input to a concrete perturbation, such as robustness in adversarial environments;

b) Robustness to changes in system input that are difficult to characterize specifically, including robustness under natural conditions such as rain, snow, fog, haze, and focus blur.

Real-world factors such as obstruction by foreign objects can cause changes to the input image samples of the system.

Note 1.A concrete disturbance refers to a disturbance whose specific characteristics (such as the form and magnitude of the disturbance) can be clearly characterized under certain assumptions. Robustness meets the following requirements.

a) Techniques such as adversarial training, credibility defense, and data processing should be employed to enhance the robustness of artificial intelligence algorithms and improve model performance. The ability to defend against adversarial attacks;

b) The impact of adversarial examples in the corresponding tasks on the operation of the artificial intelligence system should be analyzed;

c) It should be able to withstand the effects of adversarial attacks, anomalies, or "black swan events";

Note 2."Black Swan event" refers to an unexpected event that is outside of normal expectations, has an extremely low probability of occurrence, but will have a huge impact once it occurs.

d) It should resist unauthorized system input and not enter an abnormal or unsustainable state due to unauthorized system input.

5.4 Reliability

Reliability refers to the ability of an artificial intelligence system to stably, continuously, and accurately perform tasks and produce expected results under specified conditions and within a predetermined timeframe.

As a result, it exhibits consistent performance under different environments and conditions. Reliability meets the following requirements.

a) Machine learning frameworks, algorithm libraries, databases, and toolchains that meet the requirements of the application scenario should be used;

b) A system environment that meets the requirements of the application scenario should be available;

c) It should have a feedback mechanism for illegal system input;

d) Standardized testing should be conducted on system reliability, covering scenarios including load testing, boundary condition testing, and fault injection testing;

e) The impact of environmental interference data on the operation of the artificial intelligence system should be analyzed;

Note 1. Environmental interference data refers to the input data of an artificial intelligence system affected by environmental factors (such as rain, snow, fog, haze, focus blur, obstruction by foreign objects, environmental noise, etc.).

The resulting impact leads to inaccuracy and distortion of the data.

f) It should be able to withstand the impact of user interaction;

Note 2. The impact of user interaction includes. altering the system data distribution through multiple interactions, resulting in system output that does not meet information security requirements; user input containing... Biased or misleading information can cause the system to output incorrect or inaccurate answers.

g) The impact of dataset distribution shifts on the operation of the artificial intelligence system should be analyzed;

h) The impact of outlier data on the operation of the artificial intelligence system should be analyzed;

i) It is advisable to set health monitoring indicators (such as prediction deviation rate, task success rate, and processing latency) during system operation to detect problems in advance.

A decline in reliability is being detected and an early warning is being issued.

5.5 Accuracy

Accuracy refers to the property of an artificial intelligence system to produce outputs that closely match or are consistent with the expected results when performing a specific task.

Accuracy evaluation metrics include, but are not limited to, individual performance metrics (such as accuracy, precision, and recall) and overall performance metrics (such as recipient operating characteristics).

(Curve and area under the curve). Accuracy must meet the following requirements.

a) Functional correctness. In practical applications, artificial intelligence systems should be able to correctly perform various tasks according to the expected and set goals. and functions;

b) Functional Appropriateness. During the design, development, and implementation of an artificial intelligence system, its functions should accurately and appropriately meet user needs.

Furthermore, it exhibits good adaptability and stability in specific application scenarios;

c) Result Accuracy. The accuracy of the output results produced by the artificial intelligence system after processing the input data compared to the actual situation or expected results.

The accuracy ratio should meet the accuracy requirements of the business or scenario.

d) Accuracy thresholds should be defined according to different task scenarios (such as classification, prediction, and generation), and appropriate evaluation metrics should be used for verification;

e) For generative models, the generated content should be analyzed to determine whether it contains inaccurate, incomplete, or misleading information.

5.6 Toughness

Resilience refers to the property of an artificial intelligence system to return to operational status within a desired timeframe after an accident, malfunction, or external attack.

Artificial intelligence systems that meet resilience requirements possess adaptive response capabilities, collaborative protection capabilities, dynamic positioning capabilities, and versatility capabilities. The toughness must meet the following requirements.

a) In the event of functional or partial failure of the artificial intelligence system, it should be able to recover quickly and remain within the committed service level agreement. Service quality capability;

(b) The artificial intelligence system should have self-diagnostic capabilities and, upon detecting a fault, be able to automatically or quickly restore functionality with manual assistance. This ensures system resilience and consistent service quality.

c) Adaptive Response Capability. When faced with environmental changes and uncertainties, artificial intelligence systems should be able to effectively adjust their behavior and maintain [their resilience/resilience]. Performance retention;

d) Collaborative Protection Capability. When facing external interference or attacks, the artificial intelligence system should be able to work collaboratively with other systems or components to protect itself.

Together, we can resist threats and maintain overall stability and security.

e) Dynamic positioning capability. The artificial intelligence system should be able to self-position and adjust based on real-time data and environmental changes;

f) Diversity capability. Artificial intelligence systems should provide multiple redundancy mechanisms, strategy paths, or models under different faults, scenarios, and requirements.

Alternative solutions to maintain a diverse range of response and service assurance capabilities to maintain critical functions or achieve target functions.

5.7 Reproducibility

Reproducibility refers to the ability of an artificial intelligence system to produce consistent or acceptable levels of error in different operating environments under acceptable reproducibility conditions.

Results within a certain range. Reproducibility conditions generally refer to results obtained using the same data, algorithm model, parameter configuration, and operation sequence, but in different operating environments.

The context refers to different computing devices or software that are not dependent on by the business logic. Reproducibility meets the following requirements.

a) The system should have a verification mechanism to ensure that the system can reproduce the results under specific conditions.

b) User interactions should not alter the reproducibility of the AI system throughout its entire lifecycle;

Note: Reproducibility under extreme or special environments or conditions is not considered.

c) The impact of software and hardware platform dependencies on the operation of artificial intelligence systems should be analyzed, including the applicability of deep learning frameworks and operating systems. Differences, portability of hardware architecture, etc.

5.8 Generalizability

Generalizability refers to the ability of a trained artificial intelligence system to make accurate predictions when applied to new data with a distribution similar to the training data. Quality. Generalizability meets the following requirements.

a) Regularization, cross-validation, and other techniques should be used to improve the system's generalizability;

b) When accepting unknown data input, underfitting or overfitting should not occur;

c) When accepting data input with similar characteristics and distribution to the training data, key performance indicators (such as accuracy, precision, etc.) should be maintained. The recall rate is not lower than the specified threshold;

d) When accepting out-of-distribution (OOD) or weakly correlated data inputs, appropriate downgrading or rejection strategies should be adopted;

e) When the distribution of training data changes significantly, the artificial intelligence system should be retrained in a timely manner.

5.9 Real-time performance

Real-time performance refers to the ability of an artificial intelligence (AI) system to process, analyze, and respond to external requests within a given time period. The real-time performance of an AI system... The performance requirements should be determined based on the application scenario. In some scenarios (such as medical image analysis, autonomous or assisted driving, industrial production line control, etc.), the performance requirements may vary.

In this context, high real-time performance is required. Real-time performance must meet the following requirements.

a) Relevant tasks should be processed within the time delay requirements of the task requirements;

b) The task switching time requirement should be met;

Note 1. Task switching time refers to the time required for the operating system to switch from one task (process or thread) to another during task scheduling.

c) The task completion delay jitter requirement should be met.

Note 2. Latency jitter refers to the difference between the completion time and the expected completion time of different task instances or different stages within a task when an artificial intelligence system is executing a task. The differences.

5.10 Backupability

Backup refers to the process by which an artificial intelligence system, to prevent data loss due to operational errors or system malfunctions, moves all or part of its data set from a previous location.

The process of copying data from the host machine to other storage media or systems. Backup systems can use artificial intelligence with the same functionality but different algorithms/models.

The system can also use a non-AI system with similar functionality. Backupability must meet the following requirements.

a) It should implement the same business logic as the main system and meet business requirements;

(b) Logical isolation between functions should be implemented. When part of the system or part of the function malfunctions, degradation or isolation methods can be used to ensure the protection of other functions. Business functions are operating normally;

c) In the absence of a backup system, there should be corresponding procedures to ensure that professional technicians are involved in controlling the system in the event of failure;

d) A backup strategy specification should be established for the backup system, including backup cycle, backup method (full/incremental backup), and number of backup copies. Quantity, etc.;

e) It is advisable to adopt a different technical approach from the main system (e.g., backups of deep learning-based subsystems should preferably be performed using non-AI technologies).

When there are no backups using different technical approaches (to implement business logic), it is advisable to use multiple models for mutual backup.

5.11 Data Compliance

Data compliance refers to the ability of an artificial intelligence system to adhere to relevant regulations throughout the entire lifecycle of data collection, processing, storage, use, and output.

The nature of laws, regulations, ethical standards, and industry best practices. Data compliance meets the following requirements.

a) It should have security and integrity assurance functions during the transmission of control flow data, information flow data, and application data;

b) Data security management should be implemented to ensure that data complies with relevant standards and regulations on data security and personal information protection; fine-tuning and machine-based methods are recommended.

Techniques such as machine forgetting are used to delete or correct non-compliant data after model training;

c) Intellectual property rights should be tested to prevent infringement of intellectual property rights in the selection of training data and output of results;

d) Data sets should be obtained through verified and compliant channels;

e) When collecting and processing personal data, it should be limited to the specific, lawful purposes stated therein, and no further processing should be carried out after those purposes have been achieved. Step-by-step data processing;

f) Labeling rules should be established, including labeling objectives, data formats, labeling methods, quality indicators, etc.

g) A comprehensive data security framework should be established, including but not limited to technologies such as data encryption, access control, data anonymization, and data auditing. Measures to protect sensitive information and important data.

5.12 Privacy Protection

Privacy protection refers to ensuring that information about individuals or entities involved in artificial intelligence systems that they do not wish to have known by outsiders is adequately protected.

Privacy protection includes the protection of personal information and privacy, as well as the protection of trade secrets. Privacy protection must meet the following requirements.

a) Only the minimum data necessary to implement the functionality should be collected, avoiding the excessive collection of user information;

b) The work and requirements for data and model privacy protection within the lifecycle management should be clearly defined, and a privacy protection plan should be formulated;

c) The objects or subjects of privacy protection and the points of risk of privacy leakage should be clearly specified;

d) The privacy protection methods adopted should be fully evaluated, and the applicable scenarios, conditions and limitations of the privacy protection methods should be clearly defined;

e) In scenarios where users provide training data, malicious querying behavior should be prevented;

f) In scenarios involving user-system interaction, AI systems should prevent attackers from gaining access to the target's data by repeatedly querying the trained model. Copyright protection information;

g) Technical means should be adopted to protect the privacy of data and models, such as secure multi-party computation, trusted computation, differential privacy, homomorphic encryption, etc.

Federated learning and other privacy-preserving and confidential computing, and the establishment of monitoring and auditing mechanisms;

h) The balance between the availability and confidentiality of privacy-preserving data should be considered in the privacy protection methods employed;

Note: Data availability refers to the ability of users or systems to obtain, access, and use data in a timely and accurate manner. Data confidentiality refers to protecting data from... The ability to be accessed and disclosed without authorization.

i) Sensitive raw data should be processed locally in a trusted environment to reduce the risk of privacy leaks during data transmission.

5.13 Confidentiality

Confidentiality refers to the confidentiality that any natural or legal person should maintain in accordance with laws and regulations during the development, deployment, operation, and related regulatory activities of artificial intelligence systems.

The regulations stipulate that the confidentiality of information and data obtained in the course of performing tasks and activities must be ensured, and that unauthorized access, disclosure, use, or dissemination must be prevented.

The core objective is to protect sensitive information involving intellectual property, trade secrets, and national security during the implementation of regulations. Confidentiality must meet the following requirements. Require.

a) The scope of acquisition and use of collective data should be strictly limited;

b) The principle of least privilege should be adopted;

c) Collective data that is no longer used for its acquisition purpose should be deleted within the specified time limit;

d) A review and control mechanism for the output content of artificial intelligence systems should be established.

5.14 Transparency

Transparency refers to an AI system clearly defining the nature of appropriate information about the system to stakeholders. For trustworthy AI, if the model... If the parameters and overall structure of a department are observable or accessible, then its behavioral logic and decision-making basis can be studied and analyzed, i.e., artificial intelligence.

System transparency can be understood as the degree to which a model is "white-box." Transparency depends on interpretability, that is, the ability to interpret the behavior of the artificial intelligence system and its components. A useful or reasonable explanation. Transparency meets the following requirements.

a) A risk monitoring mechanism should be established, such as providing the criteria and standards on which the decisions are based and informing users of the potential risks associated with the output results. Risks and limitations;

b) Users should be informed that the objects they interact with and the sources of the services they receive are artificial intelligence systems;

c) The model capability assessment results and security measures should be disclosed;

d) Artificial intelligence systems should consider adopting relevant mechanisms to improve the predictability of their behavior;

e) The principles, capabilities, applicable scenarios, and security risks of artificial intelligence technologies and products should be appropriately disclosed in a tiered and targeted manner to ensure suitability.

To meet the needs of various stakeholders, the output content is clearly labeled.

5.15 Traceability

Traceability refers to the ability of an artificial intelligence system to trace the history, applications, and location of a specific entity. Traceability is a fundamental property of artificial intelligence systems.

Accountability is a prerequisite and the foundation for auditing and governing the application of artificial intelligence algorithms across industries, thus providing a basis for the functional safety design of artificial intelligence systems.

And to provide evidence for enhancement. Traceability meets the following requirements.

a) Tools such as logging and status monitoring should be provided;

b) The development process of the artificial intelligence system should be documented, including system design and changes, program development, data usage, model training, and user requirements. Request compliance testing, model transfer, and other relevant information;

c) The operation of the artificial intelligence system should be recorded, including operation steps, inputs, and results, so as to trace the relevant processing from the input.

And the result can also be traced back to the input or operation;

d) In artificial intelligence systems, model version control should be implemented, and the training, testing, and deployment history of each model version should be recorded in detail;

e) Persistent storage of the operation logs of the artificial intelligence system should be implemented;

f) Artificial intelligence systems that serve the public should be identified and managed using technologies such as digital certificates;

g) When personal privacy information is involved, the records and logs of the artificial intelligence system should be stored under strict privacy protection policies.

Storage and auditing should be conducted by legally authorized departments or personnel when necessary.

5.16 Explainability

Interpretability refers to the ability of an artificial intelligence system to express the nature of key factors influencing its (execution) results in a way that is understandable to humans.

Interpretability helps people understand the internal workings of complex models and important questions such as how models make specific decisions. Interpretability aligns with... The following requirements are made.

a) Different levels of explanation should be provided to facilitate understanding for different stakeholders, such as providing users with the system intent and decision results.

Provide the product provider with explanations such as algorithm models and operating environments;

b) An explanation should be provided beforehand (an internal, design-time explanation), such as statistical analysis of the training data and the size of the training dataset.

Analysis, including deductive and inductive analysis of various typical scenarios in the original training dataset;

c) It should be able to assess data quality, such as providing clear labeling benchmarks and defining data labeling methods;

d) Explainable aspects of the modeling process should be provided, such as explanations of the modeling decisions, modeling basis, and optimization/adjustment behaviors of the artificial intelligence system.

Explanation. This provides definitions for hyperparameters and setting conditions during model optimization;

e) A post-implementation explanation should be provided to stakeholders after the system is put into use, explaining the logical relationship between specific inputs and outputs. It should be understandable to stakeholders;

f) Clear descriptions should be provided for the internal structure, reasoning logic, technical interfaces, and output results of artificial intelligence systems to accurately reflect the characteristics of artificial intelligence. The process by which a system produces results;

g) It is advisable to add a user feedback mechanism to support users in asking questions or providing feedback on the provided explanations.

5.17 Fairness

Fairness refers to the fact that artificial intelligence systems respect established facts, social norms, and beliefs, ensuring that stakeholders are not subject to bias or unfair discrimination.

The nature of the treatment, behavior, or outcome affected. Fairness is often related to factors such as culture, region, and political views. Fairness meets the following criteria. Require.

a) Appropriate algorithm bias detection solutions should be provided based on the actual application scenario;

b) Exploratory analysis and prediction should be conducted on the size and balance of the datasets used to train the intelligent algorithms employed in the artificial intelligence system. Treatment to eliminate or mitigate bias;

c) The possibility of inherent prejudices and discrimination arising in the cultural, personnel, and social environments within its application areas should be reduced;

d) Provide similar quality of service to all target groups;

e) On the test dataset for detecting bias, the change in the prediction accuracy of the artificial intelligence system should not exceed the maximum change threshold range;

f) In the processes of algorithm design, model training and optimization, and service provision, methods such as training data screening and output verification should be adopted to prevent... To prevent discrimination based on ethnicity, religion, nationality, region, gender, age, occupation, or health;

g) It is advisable to identify potentially bias-related characteristics, including but not limited to gender, race, ethnicity, and special groups (such as people with disabilities).

Identify the correlation between results and features during the verification or testing process;

h) It is advisable to identify potential causes of bias, including but not limited to sample distribution, incorrect exclusion, inconsistent measurement or sampling, and inconsistencies. mark;

Note: Examples of sample distribution include whether training is performed on a dataset with significant features; examples of incorrect exclusion include failing to include non-significant features in the training dataset.

Examples of inconsistent measurements or sampling include visual samples of disabled protected animals taken only under indoor lighting conditions, and inconsistent labeling. Examples include single-person annotation, single-algorithm annotation, and no cross-validation.

i) It is advisable to provide a mechanism for reviewing biased results, or at least to provide a channel for feedback on issues.

5.18 Ethical Compliance

Ethical compliance refers to the adherence of an artificial intelligence system to ethical guidelines throughout its entire lifecycle. Ethical compliance meets the following requirements.

a) An ethical compliance assessment of artificial intelligence should be conducted, covering aspects such as legal and regulatory compliance, fairness and non-discrimination, human controllability, professional responsibility, and human rights.

Values, internationally accepted codes of conduct, environmental sustainability, worker awareness and collaboration, etc., and clearly define the corresponding ethical bottom line. and risk red lines;

(b) The goal should be to promote human well-being and reduce the possibility of algorithm designers misleading users or causing excessive reliance on algorithms for their own benefit.

The risk of algorithm abuse stemming from relying on the algorithm itself and blindly expanding its application scope;

c) Channels for feedback on algorithm ethics should be provided to identify and improve potential algorithmic biases, algorithmic bullying, and information cocoons in recommendation algorithms. question;

d) The datasets used by artificial intelligence systems should conform to their collection purpose, collecting only the minimum data necessary to achieve the function;

e) The ethical implications of artificial intelligence systems should be reviewed periodically;

f) The design and deployment of artificial intelligence systems should take into account their long-term environmental and social impacts and promote sustainable development.

5.19 Usability

Usability refers to the design, development, deployment, and application of artificial intelligence systems that fully consider various social groups and maximize their ease of use. The degree of user engagement. Usability meets the following requirements.

a) An age-friendly and accessible environment should be provided, such as a simple and easy-to-understand user interface, enhanced font readability, and voice prompts.

Supports accessibility features such as screen reader, voice recognition, and gesture control;

b) The operation process should be optimized to adapt to users' operating habits and reduce users' learning costs;

c) It is advisable to provide operation guidance, operation feedback, and confirmation prompts for important operations to guide and prompt users;

d) Clear error messages should be provided to help users understand the problem and provide solutions or suggested next steps.

5.20 Accountability

Accountability refers to the nature of artificial intelligence systems and their stakeholders being held accountable for their actions, decisions, and behaviors. Accountability requires human... AI developers and users provide transparent explanations of how their systems operate and take responsibility for problems that arise. (Ask a question) Accountability must meet the following requirements.

a) The responsible party should be clearly identified, and the responsibilities and obligations of the team and relevant personnel should be clearly defined;

(b) A mechanism for defining responsibilities should be in place, and third-party supervision should be accepted when necessary;

c) Artificial intelligence systems should establish an auditable recording mechanism for key behaviors and decision-making processes; […]