AI Ethics Program
Definition
AI Ethics Program
An artificial intelligence (AI) ethics program is the standing function that turns stated principles about acceptable AI use into decisions somebody actually makes. Principles without a decision route are decoration — that is what separates a programme from a statement.
Almost every large organisation now has AI principles. Far fewer have a named group with authority to stop a use case, a defined escalation route and a record of what has been refused.
The programme answers value questions rather than control questions. Whether a use is acceptable, whose interests it affects and what the organisation will decline to build are not answered by a technical control framework.
It therefore sits alongside governance rather than inside it. Governance decides who signs off and against which policy — while the ethics programme decides what the policy should say when the law is silent.
That silence is wider than most boards expect. Regulation covers a fraction of what organisations now build with AI, and the gap is exactly where a programme earns its cost.
Key takeaways
- The programme converts principles into a standing function with decision authority.
- It answers value questions that a control framework cannot resolve.
- A record of refused use cases is the clearest evidence it works.
- International standards give organisations a ready-made principle set to adopt.
How it works
The organisation adopts a principle set, establishes a review group with defined authority, sets triggers for when a use case must be reviewed, and records both decisions and reasoning so that precedent accumulates.
Trigger design is the practical part. Reviewing every model paralyses delivery, while reviewing none makes the group ceremonial, so most programmes trigger on impact to people rather than on technical novelty.
Ready-made principle sets exist and are widely adopted. In November 2021 UNESCO’s 193 member states adopted the first global standard on AI ethics, the Recommendation on the Ethics of Artificial Intelligence.
Intergovernmental principles came earlier. The OECD AI Principles were adopted in May 2019 and updated in May 2024, and the organisation records 47 adherents to them.
| Component | Purpose | Failure mode |
|---|---|---|
| Principle set | States what the organisation values | Copied, never applied |
| Review group | Decides contested cases | No authority to refuse |
| Triggers | Says what must be reviewed | Everything, or nothing |
| Decision record | Builds internal precedent | Verbal only |
| Escalation | Resolves deadlock | Undefined |
Examples
Programmes differ mainly in how much authority the review group actually holds when a case is contested. The four examples below show that range clearly, from a published veto to a standing condition imposed on delivery.
A bank gives its review group a veto and publishes the count of refused cases. Its responsible ai statement is measurable rather than aspirational.
A healthcare provider triggers review on any use affecting patient pathways. Every reviewed case includes an ai bias audit before a decision is recorded.
A software firm requires an explainable ai assessment for decisions affecting individuals. Where explanation is impossible, the use case is declined rather than mitigated.
A services provider embeds human in the loop review as a standing condition. The ethics programme sets the condition; delivery teams implement it.
The split of duties is deliberate. A programme that also implements its own conditions ends up reviewing its own work, and the independence that made its refusals credible disappears.
Related terms
Responsible AI vocabulary overlaps heavily in practice, and the entries below separate the values layer from the controls and the disclosure artefacts. Each answers a distinct question about how an organisation constrains its own AI work.
- AI acceptable use policy: the written rule the programme’s decisions eventually produce.
- Model card: the disclosure artefact an ethics review usually requires.
- AI washing: the marketing behaviour a credible programme has to police internally.
FAQ
How does this differ from an AI governance framework?
The framework decides who approves what against which policy. The ethics programme decides what the policy should say where law and regulation give no answer.
Who should sit on the review group?
A mix of business, legal, technical and domain specialists, plus someone representing affected users. Technical-only membership produces technical answers to value questions.
What gives the group real authority?
A documented right to refuse, exercised at least occasionally. A group that has never declined anything is treated as advisory regardless of its terms of reference.
Should principles be written from scratch?
Rarely. Adopting an established international set and adding organisation-specific commitments is faster and gives external reference points.
How is the programme measured?
By decisions made, cases refused and precedent recorded. Counting principles published or training completed measures activity rather than effect.
Does it slow delivery down?
Selectively, which is the point. Well-designed triggers mean most work proceeds untouched while a small number of high-impact cases take longer.
Read more on responsible AI and outsourcing at Outsource Accelerator.
Independent
