Definition
Model accountability means you can answer, with evidence: who owns the model, what it is allowed to do, what data it uses, how it was tested, when it changed, and who approved those changes. It is a governance property, not a technical metric.
Why it matters
- Professional responsibility: users need to justify advice supported by AI.
- Incident response: when something goes wrong, accountability enables fast root-cause analysis.
- Regulatory readiness: documentation and traceability are recurring obligations.
How it works
Owner + intended use + versioning + testing + approvals + logs -> accountabilityTypical artifacts include: a model inventory entry, evaluation results, release notes, and an audit trail of key decisions.
Practical example
If a model update changes retrieval or ranking behavior, the change is logged, validated against a test set, and approved by a named owner before deployment.
Common questions
Q: Is accountability the same as legal liability?
A: No. Accountability is about governance and evidence. Liability is a legal outcome that depends on contracts and law.
Q: Does accountability require full transparency?
A: Not necessarily. You can be accountable with well-scoped disclosures, internal documentation, and clear controls.
Related terms
- Responsible AI — broader practice
- AI Governance Framework — roles and approvals
- AI Documentation Requirements — evidence and records
- Algorithmic Transparency — explainability and disclosure
- Human Oversight — meaningful human control
References
NIST (2023), AI Risk Management Framework (AI RMF 1.0).