The biases that make you bad at AI decisions

Overconfidence is the single most recurrent bias in the research literature, across management, finance, medicine and law. It was a problem before any of this. What changed is that the average person now consults something that produces fluent, confident, well-formatted answers on demand — and fluency is the exact input overconfidence feeds on. MIT's study of 300 enterprise AI deployments found 95% returned nothing measurable, and traced the cause to integration and workflow rather than model quality. Organisations were not fooled by bad output. They were fooled by plausible output.

Three biases a model amplifies

Automation bias. People accept a machine's answer more readily than a colleague's, and check it less. The irony is precise: the more often a tool is right, the less closely it is examined, which is exactly when a wrong answer passes through.

Confirmation bias. A model will argue any side you point it at. Ask it to support your position and it will, competently, with structure. That is not evidence. It is a well-dressed reflection of the question you asked.

Fluency as a proxy for truth. Well-formed prose reads as more correct than hesitant prose. Model output is uniformly well-formed regardless of whether it is right. The usual signal — does this person sound like they know — has been decoupled from the thing it used to indicate.

Each of these existed before. What is new is a machine that produces the trigger conditions for all three, thousands of times a day, at no cost.

There is a fourth effect with no clean name: the disappearance of the struggle. Working something out slowly used to leave you with a felt sense of which parts were solid and which were guesses. An answer that arrives complete carries no such texture. You end up holding a conclusion without the map of your own uncertainty that normally comes attached to it.

The protocol

Debiasing that works is procedural, not attitudinal. Deciding to be more objective does nothing.

Slow down. A decision that feels obvious and instant is running on the fast system. That feeling is a signal to pause, not a signal you are right.

Name it. When you notice a pull toward a conclusion, ask which bias could produce that pull. Labelling a bias measurably weakens it. This is the cheapest step and the most skipped.

Seek disconfirmation. Ask what would have to be true for you to be wrong — then go looking for it. With a model, this means asking it to argue the opposite case as hard as it argued yours. If it cannot, you have learned something. If it can, you have learned more.

Get adversarial input. Designate a person to argue against the decision before you commit. Self-review does not surface blind spots, because the blind spot is doing the reviewing.

What this is worth

The operator advantage is not access to better tools. Everyone has the same models. The advantage is having a procedure that survives a confident wrong answer.

That procedure is boring by design. Slow down, name it, look for the disconfirming case, let someone attack it. None of it is clever. All of it is rare, because each step costs time at the moment you least want to spend it.

It is worth being precise about when to spend it. Every decision does not need this. Reversible, cheap, fast decisions should be made quickly and corrected later — running a protocol on them is its own waste. The protocol earns its cost on decisions that are expensive to unwind: hires, prices, architecture, anything with a contract attached. The skill is telling the two apart before you start, not after.

Your Next Move

Add one line to every consequential prompt: ask for the strongest case against the conclusion you want. Read that answer first.

Pick your recurring decision type — hiring, pricing, whether to build — and write the pre-mortem before the decision, not after. Assume it failed. Explain why.

Name one person as your adversarial reviewer, and tell them the job is to argue against you. An informal reviewer agrees with you. A named one does not.