What Is Accountability in AI? A Working Definition

Ask what accountability in AI is and you will get one of two answers. The first is a slogan: humans must stay responsible for machines. The second is a forty-page paper. Neither helps the policymaker drafting a procurement rule or the founder writing a governance page for an investor. This piece gives the working definition in between: what accountability actually consists of, how it differs from responsibility and liability, why AI systems strain it, and how it is rebuilt in practice.

The short answer

Accountability in AI is a relation in which a named person or organisation must explain and justify how an AI system was built and used, to someone with the standing to question that explanation and impose consequences if it falls short.

That definition is borrowed from public administration, not invented for machines. In the most cited recent treatment, published in AI and Society, accountability is a relation between an agent and a forum: the agent must justify its conduct, and the forum supervises, asks questions and passes judgment. The authors call the content of that relation answerability, and they add three conditions without which it is only a press release: the agent recognises the forum's authority, the agent is exposed to interrogation, and the arrangement actually limits the agent's power. As they put it, to be accountable only to oneself means not being accountable at all.

Note what is absent from the definition: the AI system itself. The system is the thing being accounted for. It is never the one giving the account.

Accountability, responsibility and liability are not the same

The three words are used interchangeably in most writing on AI, and that is where much of the confusion starts. The OECD, in the commentary on its accountability principle, separates them carefully:

• Accountability is the expectation that an organisation can explain why decisions were taken and, after a bad outcome, act to make the next one better.
• Responsibility is the causal link between an actor and an outcome, used in both moral and legal senses.
• Liability is the adverse legal consequence of an action or omission.

The OECD asks that accountability be demonstrated, for example through documentation of key decisions across the system's lifecycle and through audits where justified. The US National Telecommunications and Information Administration takes a wider view in its AI accountability glossary, defining AI accountability as the process of holding entities answerable for the risks and harms of the systems they develop or deploy, and noting that the joint US and EU terminology treats it as an "allocated responsibility". The practical difference is small. Both readings agree that accountability is about being made to answer; liability is only one of the things that can follow.

The four parts of an accountability relation

Strip the theory down and every working accountability arrangement answers four questions.

Who answers

A named agent: the developer, the deployer, the procurement officer, the board. "The organisation" is not an answer, and neither is "the algorithm".

To whom

A forum with standing: a regulator, a court, an auditor, an ethics board, the affected public. Without a forum there is reporting, not accountability.

Against what standard

The rules the conduct is judged by: law, a published policy, a contract, a technical benchmark. The Springer authors list four goals these standards can serve, compliance, report, oversight and enforcement, and point out that policymakers stress different ones depending on whether they want to prevent harm or respond to it.

With what consequence

A correction, a withdrawal, a fine, a damages award. The NTIA calls these consequences accountability outputs, set by regulators, courts and markets, and spends most of its report on the inputs that feed them. Its glossary quotes the Ada Lovelace Institute on the point that matters most here: assessments are themselves not a form of accountability. An audit with no forum and no consequence is homework nobody marks.

Why AI strains accountability

None of this is new for human institutions. AI strains it in three specific ways.

Machines cannot be called to account

The Alan Turing Institute's teaching material on accountability states it bluntly: automated decisions are not self-justifiable. A person can be questioned about a judgment; a statistical model and the hardware it runs on cannot be responsible in the same morally relevant sense. The result is what the Institute calls an accountability gap, and the fix it proposes is to attach a clear, imputable source of human answerability to every decision an AI system makes or supports.

Many hands

An AI system is rarely the work of one actor. Data comes from one place, the base model from another, fine-tuning and deployment from a third, and the decision is taken by a fourth. Each can plausibly say the harm happened somewhere else in the chain. Accountability has to be allocated along that chain in advance, or it dissolves.

Opacity

The forum has to be able to see what it is judging. The NIST AI Risk Management Framework compresses this into one line: "Accountability presupposes transparency" (NIST AI 100-1). A system whose training data, evaluation results and deployment decisions are undocumented cannot be held to account however willing everyone is.

How accountability is built back in

The Turing material splits the remedy into two components, and they are the most useful practical lens on the whole subject.

Answerability

Answerability answers who is accountable. It requires a continuous chain of human responsibility across the whole project, with no gaps from the first design decision to the system's outputs, and competent humans able to explain both the reasoning behind results and the process that produced them in plain language.

Auditability

Auditability answers how they will be held accountable. It means traceability, documenting each stage from data collection and model selection to deployment, updating and retirement, and it means letting peers and overseers probe the running system. Google's guide to responsible AI accountability makes the same point from the engineering side: accountability typically involves sharing how models and datasets were created, trained and evaluated. The Turing Institute names the combination of the two accountability by design.

Before and after the fact

The Turing Institute's accountability guidance brief adds a timing distinction. Anticipatory, or ex ante, accountability covers decisions made during design and development; remedial, or ex post, accountability covers corrections and justifications after deployment. Its advice is to prioritise the first. Its main tool is unglamorous: a process log, a single repository of the governance decisions and documents produced along the way, so the account can be given when someone asks.

For how GAO, NIST, the OECD and the EU turn these ideas into checklists and audit procedures, see our comparison of AI accountability frameworks.

What it looks like when it works

The clearest recent illustration came from a small claims tribunal in British Columbia. Air Canada argued that its website chatbot, which had given a customer wrong information about bereavement fares, was responsible for its own statements. The tribunal rejected that argument and held the airline to what its system had said. Every part of the relation is visible: an agent (the airline), a forum (the tribunal), a standard (the duty of care to a customer) and a consequence (compensation). The attempt to make the machine the accountable party failed for exactly the reason the definition predicts. We cover that case and the wider law in who is responsible when AI gets it wrong.

At the other end of the scale, the EU AI Act attaches the heaviest consequences yet written to AI accountability: for prohibited practices, fines of up to 35 million euro or 7% of worldwide annual turnover, whichever is higher. That is the point where accountability stops being a principle and becomes a line in a budget. The difference between the binding and voluntary instruments is the subject of our piece on AI governance vs AI regulation, and the US legislative counterpart is covered in the AI Accountability Act.

Where the definition is still being argued

The definition above is settled enough to work with. What is not settled is the balance. The Springer authors close by conceding that the trade off between compliance, reporting, oversight and enforcement remains a matter of ethical, legal and political deliberation, and that is a fair description of where the field stands. Agentic systems that act over long chains of steps make the "who answers" question harder, not easier, and the documentation that auditability depends on is still mostly voluntary outside the European Union.

These are the questions on the table at EX Future Summit, which runs 18 to 20 November 2026 as a single continuous broadcast between Las Palmas and Bali, with AI Ethics and Government among its nine tracks. Online attendance is free for verified researchers and students. For the broader moral ground underneath accountability, see our explainer on ethics in artificial intelligence.

FAQ

Can an AI system itself be held accountable?

No. Accountability needs an agent that can be questioned, can justify its conduct and can bear consequences. An AI system can do none of those in the morally relevant sense, so accountability always lands on the people and organisations that build, buy or deploy it.

What is the difference between AI accountability and AI transparency?

Transparency is an input; accountability is the relation it feeds. Disclosing how a model was trained and tested lets a regulator, court or auditor judge it. Transparency with no forum to act on it is disclosure, not accountability.

Is AI accountability a legal requirement?

It depends where you operate. The OECD principle and most national frameworks are voluntary. The EU AI Act is binding law with fines attached, so organisations placing AI systems on the European market now face enforceable accountability obligations.

What is accountability by design?

It is the practice of building answerability and auditability into an AI project from the start: named owners for each stage, documentation of key decisions, logging of system behaviour, and access for reviewers, so that an account can be given later without reconstructing it from memory.