I AM J RESPONSIBLE AI PRINCIPLES

1. Principle 1 — Start with a Legitimate Human Need

Define the problem, intended beneficiaries, expected benefit, and responsible owner before selecting an AI solution.

• Confirm that AI is suitable for the problem.
• Document intended uses, users, environments, and foreseeable misuse.
• Avoid deploying AI merely because it is novel or available.

2. Principle 2 — Assess Risk Before Deployment

Perform risk assessment early and update it as the system, data, context, or deployment changes.

• Consider safety, privacy, security, bias, manipulation, misinformation, accessibility, human-rights, operational, environmental, and reputational risks.
• Classify risk using the severity and likelihood of harm, affected population, scale, and reversibility.
• Require additional approval and controls for high-impact or vulnerable-population uses.

3. Principle 3 — Use Appropriate Data

Use data that is lawfully obtained, relevant, sufficiently representative for the intended purpose, and protected according to its sensitivity.

• Document important sources, permissions, quality limitations, exclusions, transformations, and retention rules.
• Reduce unnecessary personal data and sensitive attributes.
• Evaluate whether historical data may encode exclusion, discrimination, or harmful institutional practices.

4. Principle 4 — Build for Safety and Security

Design for predictable behavior, safe defaults, abuse resistance, access control, and graceful failure.

• Use threat modeling and adversarial testing appropriate to the system.
• Protect credentials, models, prompts, user data, logs, and connected tools.
• Limit autonomy, permissions, and external actions to what is necessary.
• Provide fallback behavior when models, connectivity, integrations, or external data are unavailable.

5. Principle 5 — Evaluate in Context

Evaluation should reflect the languages, devices, connectivity, environments, user abilities, and real-world conditions in which the system will operate.

• Measure task performance and material failure modes, not only aggregate accuracy.
• Evaluate hallucination, unsafe behavior, prompt injection, bias, privacy leakage, reliability, and accessibility where relevant.
• Do not generalize evaluation results beyond the tested scope without justification.

6. Principle 6 — Preserve Meaningful Human Oversight

Determine where human review, confirmation, interruption, escalation, or override is necessary.

• High-impact decisions must not rely solely on unverified AI output.
• Reviewers should have sufficient time, authority, information, and competence to challenge the system.
• Interfaces should avoid automation bias and false impressions of certainty.

7. Principle 7 — Be Transparent About AI

Provide clear, audience-appropriate information about AI involvement, intended purpose, important limitations, data practices, and human responsibility.

• Distinguish factual retrieval, generated content, estimates, recommendations, and automated actions.
• Label or otherwise disclose synthetic content where context requires it.
• Avoid claims that cannot be supported by evidence.

8. Principle 8 — Promote Fair Access and Inclusive Performance

Design for diverse users and avoid creating unnecessary barriers based on language, disability, location, connectivity, culture, income, or technical literacy.

• Test important differences in performance where feasible and lawful.
• Provide alternatives or accommodations when known limitations could materially disadvantage users.
• Seek feedback from affected groups and domain experts.

9. Principle 9 — Protect Privacy and Confidentiality

Apply privacy-by-design and data-minimization practices.

• Limit access to personal and confidential data.
• Use retention periods aligned with defined purposes and legal requirements.
• Provide appropriate notice, choice, consent, correction, and deletion mechanisms.
• Avoid exposing personal information through outputs, logs, analytics, or integrations.

10. Principle 10 — Maintain Accountability and Traceability

Assign an accountable owner for each material AI system or deployment and maintain records proportionate to risk.

• Record key decisions, approvals, tests, known limitations, changes, incidents, and remediation.
• Maintain version awareness for models, prompts, policies, data sources, and integrations.
• Enable investigation without creating excessive or unjustified surveillance.

11. Principle 11 — Monitor, Learn, and Correct

Monitor deployed systems for material drift, failures, misuse, security events, user complaints, and changes in context.

• Establish thresholds for review, rollback, restriction, or shutdown.
• Prioritize remediation according to risk and impact.
• Communicate material incidents and corrective actions to affected stakeholders when appropriate.

12. Principle 12 — Govern Partners and Deployments

Responsible AI duties extend to vendors, foundation-model providers, integrators, resellers, research partners, customers, and field deployment partners.

• Conduct proportionate due diligence.
• Define responsibilities and prohibited uses contractually.
• Provide deployment guidance, training, and escalation channels.
• Review material third-party changes that can alter risk.

13. Decision Rule

When evidence is incomplete, choose the course that protects people, preserves reversibility, limits exposure, and allows further testing. Commercial urgency does not override safety, law, human rights, or responsible governance.

14. Contact and Governance Questions

Questions, concerns, research inquiries, partnership proposals, or reports related to this document may be submitted through the I am J contact page.