Course Content
AI Safety & Guardrails
5 sections · 50 lessons
How should organizations handle accountability for AI decisions?
What you need to know
A sample RACI for an AI system
| Activity | Responsible | Accountable | Consulted |
|---|---|---|---|
| Training / evaluation data | Data team | System owner | Privacy, legal |
| Model and prompt changes | ML engineers | System owner | Safety reviewer |
| Launch decision | Product | System owner | Legal, risk |
| Monitoring and incidents | On-call | System owner | Support |
| User appeals | Operations | Business owner | Legal |
The building blocks
- An owner per system with a name, not a team.
- Sign-off with evidence: eval results, fairness slices, red-team findings, known limits. Signing means reading and accepting the remaining risk.
- Traceability: every consequential output links to model version, prompt version, data snapshot, guardrail decisions and the human approver. You cannot answer for a decision you cannot reconstruct.
- Human decision-makers for high-impact decisions.
- Recourse: an explanation and an appeal to someone with power to overturn. GDPR gives people rights to human intervention and to contest solely automated decisions with significant effects.
- Vendor contracts that say who is responsible for model failures.
Authority and incentives
If the accountable owner cannot delay a launch or fund a fix, accountability is only a name on a document — someone positioned to absorb blame, not prevent harm.
A real-life example
An HR screening assistant at a large IT services firm ranks 30,000 applicants a month. After a bias complaint, an internal review asks, "Who approved using this for campus hiring?" Nobody can answer: the data-science team built it for lateral hiring, and the campus team started using it without review.
The fix: the head of talent acquisition becomes the accountable owner; any new use of the tool requires a documented approval with a fresh bias evaluation on that population; every ranking stores model version, prompt version and the recruiter who acted on it; and candidates can request a human review through the careers portal. Six months later, a similar request from the internship team goes through review and is approved with an adjusted rubric.
Follow-up questions to expect
- "Can accountability be shared by a committee?" — Committees can advise, but one person must be accountable, or no one is.
- "What about vendor models?" — You can transfer some financial risk by contract, not accountability to your users and regulators.
- "How does automation bias affect accountability?" — If reviewers approve whatever the model suggests, the "human decision" is nominal. Measure override rates and review times to check oversight is real.