What we actually mean when we say an AI is accountable

Oxford's ordinary-language philosophers, J.L. Austin foremost among them, had an unglamorous but devastating habit: before arguing about a big word, take it apart and ask what it actually does. Half of philosophy's disputes, they suspected, were people using one word to mean four different things and talking past each other.

Accountable AI is exactly that kind of word. It appears in every governance policy and vendor deck, and it currently means everything and therefore nothing. Let us do the Oxford thing and take it apart, because the distinctions turn out to be the difference between a slogan and a system you can actually build.

Four things people mean by "accountable"

When someone demands accountable AI, they usually mean one or more of these, without saying which:

1. Explainable — we can understand why the system produced this output.
2. Attributable — we can say what produced it: which workflow, which model, which data, at what time.
3. Answerable — some person can be called on to justify it and explain the decision to run it.
4. Liable — some person or entity bears the consequences when it causes harm.

These come apart cleanly. An output can be attributable but not explainable (we know the model produced it; we cannot fully say why). It can be explainable but leave no one answerable. Conflating them is why "we need accountable AI" so often produces a dashboard nobody is responsible for reading.

The responsibility gap

Andreas Matthias named the sharpest problem in 2004: the responsibility gap. Traditional responsibility assumes the operator could have known and controlled what the machine would do. But learning systems act in ways their creators did not specifically foresee and cannot fully control. If nobody could have predicted the behavior, the old rule threatens to assign responsibility to no one — a gap.

This is not academic. When an agent sends a wrong-toned collections email to a major customer, "the model did it" is a genuine attempt to fall into the gap. The whole point of AI governance is to make sure that gap never opens — that there is always a nameable human on the answerable and liable side, by design.

Machines can act without being moral agents

Luciano Floridi and J.W. Sanders offered the key move: distinguish an entity that is a source of action (a "moral agent" in a minimal, functional sense) from a bearer of responsibility. An AI can be accountable-as-source — we correctly attribute the action to it — while a human remains responsible. We do not need to decide whether the AI "really" has intentions to hold a person answerable for deploying it. Attribution to the system and responsibility for the person are two different ledgers, and both can be filled.

Approval is a speech act

Austin's most famous idea was the performative: some utterances do not describe the world, they change it. Saying "I do" at a wedding, or "I name this ship," does something. So does "I approve."

When a manager approves an AI-drafted quote, that is not a description of the quote's quality — it is an act that transfers answerability onto the approver and authorizes the action. This is why the record of who approved what, and when is not paperwork. It is the trace of the exact moment a human stepped into the loop and took on the answerable role. Remove the record and the speech act leaves no evidence it ever happened; keep it and you have closed the responsibility gap at a specific, timestamped point.

Responsibility is plural

H.L.A. Hart, the Oxford legal philosopher, pointed out that "responsibility" itself splinters into several concepts — role-responsibility (the duties of your position), causal-responsibility (you brought it about), liability-responsibility (you must answer for it), and capacity-responsibility (you were able to understand and control it). Good AI governance addresses each rather than gesturing at the word:

• Capacity is bounded up front by guardrails — the system is constrained so it cannot take certain actions.
• Causal attribution is captured by the audit trail — we can reconstruct what brought the outcome about.
• Role and liability are assigned by the approval gate — a specific person, in a specific role, answers for the specific output they signed.

Closing the gap in software

Take the abstractions and land them. An audit trail is attribution made durable. An approval gate is answerability and liability transferred to a named human by a performative act. Guardrails are a limit on the system's capacity to act. Together they ensure that for every consequential action there is a nameable person on the responsible side and a reconstructable record on the attributable side.

That is what accountable AI should mean — not a virtue an algorithm possesses, but a property of the system-plus-humans that guarantees the responsibility gap never opens. Define the word carefully and it tells you precisely what to build.

References

1. J.L. Austin, How to Do Things with Words, Oxford University Press, 1962.
2. Andreas Matthias, The Responsibility Gap: Ascribing Responsibility for the Actions of Learning Automata, Ethics and Information Technology, 6(3), 2004.
3. Luciano Floridi and J.W. Sanders, On the Morality of Artificial Agents, Minds and Machines, 14(3), 2004.
4. H.L.A. Hart, Punishment and Responsibility: Essays in the Philosophy of Law, Oxford University Press, 1968.

---

Want to see how attribution and approvals close the loop on a real workflow? Book a 30-minute call.