Accountability Was Never the Computer’s Job

Somewhere in the last two years a new sentence entered the postmortem, and it slips out so easily that almost no one flinches when they hear it: the model got it wrong. It is said the way people used to say the network was down — a fact about the world, nobody’s fault in particular. I have started flinching, because that sentence is doing something quietly corrosive to how teams hold themselves responsible.

The asymmetry is the tell. When the work is good, a person steps forward to own it; when the work is wrong, the tool steps forward to absorb it. Credit flows toward people and blame flows toward the machine, and the more of the work the machine touches, the wider that escape hatch opens. It is a comfortable arrangement, and comfort is exactly what should make us suspicious of it.

Decades ago, long before any of this was possible, IBM wrote a single line into an internal training manual that has aged into something close to prophecy: "A computer can never be held accountable, therefore a computer must never make a management decision." Read it again with today’s tools in mind. The point was never that machines cannot decide. It was that accountability cannot be delegated to a thing that cannot be held to account — and the moment you let it make the call, you have quietly moved the decision somewhere no one can answer for it.

And the machine genuinely cannot answer for it. It has no reputation a bad outcome would dent, no standing it is trying to protect, nothing at stake in being right the next time. You cannot call it into a room after an incident and ask it to account for the choice it made, because it made no choice it can recall, and it will carry no memory of the failure into the next hour of its work. Accountability was never a property the tool could hold. It has only ever lived in the person who looked at the output and said, ship it.

Luca Bonmassar put his finger on this in a way I had felt but not said cleanly: as more of the work becomes AI-assisted, the people who get trusted will be the ones who own the output fully anyway, as though no tool had ever been involved. I think that is exactly right — and I think the inverse is the real danger. A team that learns to reach for the tool as an explanation will slowly stop reaching for the mirror.

Because here is what accountability actually does, underneath the moralising: it is the mechanism by which a failure becomes a lesson. A person who owns an outcome carries it forward — they remember the shape of the mistake, they flinch in the right place next time, they change something. Offload that ownership onto a machine that cannot learn a lesson in any human sense and you do not merely lose a name to attach to the failure; you break the loop that would have turned the failure into improvement. A culture that blames the tool is a culture that has quietly opted out of getting better.

So the standard I am holding, for myself and for the people I lead, is unfashionably simple. The draft can come from anywhere. The accountability comes from you. When something an AI helped build goes wrong, the honest answer to "who owns this" cannot be a shrug toward the model, because the model is not there to own it and never will be. In an era where anyone can generate a plausible answer in seconds, the scarce and trusted thing is a person willing to stand behind one without caveats. That willingness is not a soft skill. In an industry about to drown in plausible answers, it may be the rarest competence we have left.