When AI Fails, Look for the Missing Human

Last week I spent a day inside an AI governance tool, the kind built to tell an organization which regulations apply to it. It was fast, and it was confident. It also told a hotel company that it had to comply with hospital regulations, because it could not tell the difference between hospitality and a hospital. It cited a real law and attached it to the wrong company. When I ran a university, it treated one global institution, with a medical college in New York and a campus in Qatar, as a single entity, and it did not know that the university had just lost a large share of its federal funding and was in cuts.

The tool did nothing wrong, technically. It did exactly what it was built to do. What it could not do was the one thing that mattered: know the context, and be accountable for the call.

That is the pattern in nearly every AI failure I have studied.

When lawyers were sanctioned for filing briefs full of citations to cases that did not exist, the AI had invented plausible ones. It did what it was asked. The missing piece was a lawyer who checked the citations before signing.

When a large company quietly scrapped an AI recruiting tool because it was downgrading resumes from women, the model had simply learned from biased history. It did what it was trained to do. The missing piece was a human accountable for fairness before the tool ever reached a candidate.

When a city's public chatbot told small business owners they could do things that were against the law, the AI was confidently wrong. The missing piece was a person responsible for what the public was told.

In every case, the headline blamed the AI. But the AI did not fail. It performed exactly as designed. The failure was the empty chair where a human should have been, the one who understood the situation and would answer for the outcome.

The real choice is whether there is an accountable human between the machine's output and the decision it touches.

AI is extraordinary at speed, at drafting, at scanning, at surfacing options. It is not yet capable of the thing a good professional does almost without thinking: reading the room, knowing which rule applies to this entity in this moment, and carrying the responsibility if the call is wrong. You cannot automate accountability. Someone still has to sign the work.

So the firms and institutions that get this right will not be the ones with the most powerful model or the most seats. They will be the ones who pair the speed of the machine with the judgment of a person who stays accountable for what it produces.

That is not a limitation of AI. It is the design of trustworthy work. The model does the gathering. A human does the judging. And when something goes wrong, there is a name attached to the decision, not an algorithm to blame.

When AI fails, look for the missing human. You will almost always find them. Or rather, you will find the empty chair where they should have been.