05 Responsibility 9 min read
You Cannot Delegate Accountability
Machines can take the decision. They cannot take the answerability — so it lands on whoever is standing closest, whether or not they could have prevented anything.
Delegation moves two things that we habitually treat as one: the doing and the answering for it. Machines take the first eagerly and cannot take the second at all. A model has no interests to set back, nothing to lose, and no capacity to be sorry. So the answerability does not vanish when the work is delegated. It stays in the room, looking for a human to attach to.
The responsibility gap
The philosopher Andreas Matthias named the problem in 2004, before it was urgent. As systems become more autonomous and more adaptive, we approach a condition where nobody satisfies the traditional conditions for being held responsible. The manufacturer could not have predicted the specific behaviour, because the system learns. The operator could not have controlled it, because it acts faster than review allows and in ways the operator does not understand. The system itself is not a moral agent.
Harm occurs. The conditions for fair attribution are met by no one. That is the gap.
Societies do not tolerate gaps like this for long. Blame is not a philosophical luxury; it is load-bearing social infrastructure. When there is no principled place to put it, it does not evaporate. It goes somewhere unprincipled.
The moral crumple zone
Madeleine Clare Elish gave the phenomenon its name. In a car, a crumple zone is the region engineered to absorb the force of an impact, protecting what is behind it. In a highly automated system, the human operator frequently plays the same role: they absorb the moral and legal force of a failure they had no meaningful capacity to prevent, protecting the system, its designers and its deployers from scrutiny.
The pattern is consistent. The operator is nominally in charge, structurally powerless, and blamed first.
You can see it forming now in ordinary offices. A model generates a risk assessment, a hiring shortlist, a set of clinical recommendations, a piece of production code. A person clicks approve — one of forty they will click that afternoon, with no realistic capacity to reconstruct the reasoning behind any of them. When something goes wrong, that click is the fact everyone points to. It was a signature. It was never an oversight mechanism.
The three-question test
Before accepting a role in an automated decision — or designing one for someone else — run this. It takes a minute and it is the most practically useful thing on this page.
| Question | If the answer is no |
|---|---|
| Information. Can I reconstruct enough of why the system produced this to identify a specific reason it might be wrong? | You are rubber-stamping. Your approval carries no epistemic content and should carry no liability. |
| Time. Given my actual volume, do I have the minutes each decision genuinely requires? | The system is designed for throughput, not review. The review step is decorative and everyone above you knows it. |
| Standing. If I reject this, what happens to me? Have overrides here ever survived? | You have responsibility without authority — the exact definition of a crumple zone. |
If any answer is no, the correct move is not to work harder at reviewing. It is to say so, in writing, to whoever owns the process. That sentence is often career-uncomfortable, which is precisely why it is the substantive act rather than the diligent clicking.
Many hands, and the softening of harm
There is a second mechanism, older than automation and amplified by it. Harm mediated through a long chain of contributors is easier to cause, because each link can honestly say they only did a part. Add a machine to the chain and it absorbs the residual discomfort beautifully — it is the ideal recipient of blame precisely because it cannot be hurt by receiving it.
“The model flagged it.” “The system scored them out.” “That’s what the tool recommended.” These are not explanations. They are the sound of a decision losing its author. Watch for the grammatical tell: when consequential sentences in your organisation start losing their subjects, responsibility has already diffused past the point where anyone can hold it.
The premise is right and the conclusion does not follow. Accountability is not primarily about post-hoc punishment; it is the mechanism that generates the pre-hoc care which produces the good outcomes in the first place. Systems that nobody will answer for are systems that nobody stress-tests. The correct target of the objection is individual scapegoating, and the fix is to place accountability where the power actually sits — with the people who chose to deploy the system, set its thresholds and staffed its review — rather than to abandon accountability altogether.
Where responsibility sits on the ladder
Responsibility is the one dimension where the ladder inverts. On every other page, higher rungs mean more human involvement. Here, the rung you operate on does not change your answerability at all — and the mismatch is the danger.
- 1
Cede Machine
The machine does it. You never see the work.
Honest, if the deployer accepts the accountability explicitly. Ceding openly is safer than pretending to supervise.
- 2
Approve Machine-led
The machine drafts. You review and sign.
The danger zone. Approval rights without review capacity is the crumple-zone recipe. Most enterprise AI deployments live here.
- 3
Collaborate Shared
You and the machine work the problem together.
- 4
Critique Human-led
You do the work first, unaided. Then ask the machine to attack it.
The only rung where signature and understanding reliably coincide.
- 5
Reserve Human
You do it alone, on purpose, knowing help exists.
Information, time, standing. Ask about all three before the deployment, not after the incident. Nobody will volunteer the answers.
If you did not have the capacity to review, record that you approved on the basis of the system’s output and note what you could not check. This is unpopular and it is the single most protective habit available to a person in the slot.
Never write “the model decided.” Write “I accepted the model’s recommendation.” The grammar is not cosmetic; it is where responsibility either persists or dissolves.
For any AI system you deploy, name the accountable person and then verify their refusal has teeth. If refusing is career-ending or throughput-breaking, you do not have oversight — and you should say so before you need to.