Social Dynamics

AI Design

List of 32 statements to obtain AI system information from stakeholders

Each guideline is marked with the "Phase" during which it can be applied. There are three phases: development (P₁), deployment (P₂), and use (P₃). Guidelines are also marked with the job "Role" that should consider them. There are three roles: designer (R_D), engineer or researcher (R_E), and manager or executive (R_M).

	Guideline	Example	Phase	Role	References
System's Use
1.	Define the purpose of using the system	Define purpose as job profile matching, assessing personality traits, managing road traffic	P_1-3	R_D,E,M	[1] [2] [3] [4]
2.	Specify technical capability of the system	Specify the capability of image recognition, question answering, sentiment analysis	P_1-3	R_D,E,M	[1]
3.	Indicate the domain, industry or sector in which the system will operate	Indicate education, employment, public service	P_1-3	R_D,E,M	[1]
4.	Identify the individual subject affected by the use of the system	List any individual, group or organization impacted by the system's decisions	P_1-3	R_D,E,M	[1] [3] [4]
5.	Identify the entity or individual responsible for using and operating the system	List any organization, corporation, public authority, agency	P_1-3	R_D,E,M	[1] [2] [4]
6.	Obtain approval from an Ethics Committee or similar body for the system's use	Confirm system compliance with sustainability goals	P_1-3	R_D,E,M	[5] [2] [4]
System's Components, Evaluations, and Data
7.	Provide basic information about each model	Specify model date, version, type, license, documentation)	P_1-3	R_D,E	[2] [3]
8.	Describe the intended and out-of-scope tasks for each model	Describe the general or specific tasks for the model, provide model warnings and disclaimers	P_1-3	R_D,E,M	[2] [3]
9.	Select relevant factors for evaluating each model's performance such as groups, instruments, and deployment environments	Consider the visual presentation of age and gender	P_1-3	R_D,E	[2] [3] [4]
10.	Select relevant metrics for evaluating the real-world impacts of each model	Consider decision thresholds, uncertainty levels, and variability rates	P_1-3	R_D,E	[2] [3] [4]
11.	Report evaluation metrics for various groups	Evaluate false positive/negative, AUC, and feature importance across protected attributes such as age, gender, and ethnicity	P_1-3	R_D,E	[5] [2] [3]
12.	Redocument each model information and contractual requirements at every system update	Update the model information when re-training the system	P₃	R_E	[5] [3]
13.	Document all system components, including all models, to enable reproducibility and scrutiny	Create UML diagrams, and specify model types, versions, hardware architecture)	P_1-3	R_D,E	[5] [2] [3]
14.	Provide mechanisms for interpretable outputs and auditing of all system components	Output feature importance and provide human-understandable explanations)	P_1-3	R_D,E	[5] [6]
15.	Document the security of all system components in consultation with experts	Guard against adversarial attacks and unauthorized access)	P_1-3	R_E,M	[5]
16.	Review the code for reliability of the system	Manage version control using software	P_1-3	R_D,E	[5] [3]
17.	Develop feedback mechanisms to update the system	Provide contact email, feedback form, and notification of new knowledge extracted)	P_1-3	R_D,E	[5] [3]
18.	Continuously monitor metrics and utilize guardrails or rollbacks to ensure the system's output stays within a desired range	Validate against concept drift and test with diverse testers	P_1-3	R_D,E	[5] [3]
19.	Ensure human control over the system, particularly for designers, developers, and end-users	Include human in the loop with the ability to inspect data, models, and training methods	P_1-3	R_D,E,M	[5] [3]
20.	Ensure safe system decommissioning	Ensure decommissioned data is either deleted or restricted to authorized personnel	P₃	R_E	[5] [3]
21.	Provide an environmental assessment of the system	Report the number of GPU hours used in training and deployment	P_1-3	R_E	[5]
22.	Provide information about the training data, including any measurement errors in input data and their associated assumptions	Describe the distribution of the data over key factors	P_1-3	R_E	[5] [7] [2] [8]
23.	Compare the quality, representativeness, and fit of training and testing datasets with the system uses	Report dataset details such as demographics, and data provenance	P_1-3	R_E	[5] [7] [2] [8]
24.	Protect sensitive variables in training/testing datasets	Protect sensitive data and use techniques such as k-anonymity and differential privacy	P_1-3	R_D,E,M	[5]
25.	Ensure compliance with agreements and legal requirements when handling data	Create data sharing and non-disclosure agreements and secure servers	P_1-3	R_D,E,M	[5] [8]
Development Teams
26.	Ensure team diversity	Consider diversity in gender, neurotypes, personality traits, and thinking styles)	P_1-3	R_D,E,M	[5] [9]
27.	Train team members on ethical values and regulations	Train on privacy regulations, ethical issues, and raising concerns	P_1-3	R_D,E,M	[5]
Risks and Mitigations
28.	Identify potential harms and risks associated with system use	Prevent privacy violation, discrimination, and adversarial attacks	P_1-3	R_D,E,M	[5] [2] [3] [4]
29.	Provide mechanism(s) for incentivizing reporting of system harms and risks	Provide contact emails and feedback form to raise concerns	P₁	R_D,E	[5] [3]
30.	Develop strategies to mitigate identified harms or risks for each system use	Use stratified sampling and safeguards against adversarial attacks during training	P_1-3	R_D,E,M	[5] [2] [3] [4]
Benefits
31.	Identify the benefits of the system's use for individuals, communities, organizations, and society	Describe how the system benefits individuals from marginalized communities	P_1-3	R_D,E,M	[9] [10]
32.	Identify the benefits of the system's use for the planet	Describe how the system provides benefits to natural and cultural environments	P_1-3	R_D,E,M	[9] [10]

References:
1. To Be High-Risk, or Not To Be—Semantic Specifications and Implications of the AI Act’s High-Risk AI Applications and Harmonised Standards. DOI
2. Model cards for model reporting.DOI
3. Co-designing checklists to understand organizational challenges and opportunities around fairness in AI.DOI
4. Closing the AI Accountability Gap: Defining an End-to-End Framework for Internal Algorithmic Auditing.DOI
5. A Method for Generating Dynamic Responsible AI Guidelines for Collaborative Action.DOI
6. One Explanation Does Not Fit All: The Promise of Interactive Explanations for Machine Learning Transparency.DOI
7. The Dataset Nutrition Label: A Framework to Drive Higher Data Quality Standards.DOI
8. Datasheets for Datasets.DOI
9. The EqualAI Algorithmic Impact Assessment Tool.URL
10. The Assessment List for Trustworthy Artificial Intelligence.URL