The AI has read everything. It has seen nothing.

Ask a good language model a real question. Will this company survive eighteen months. Will this hire work out. Does this market exist. You will get a remarkable answer: structured, balanced, honest about its limits, and better written than the best consultant you could afford.

Read it twice and you will notice something else. That answer is the weighted average of everything ever written about situations resembling yours. This is its strength, and it is not a manufacturing defect: it is precisely what the thing was built to do. A model is a compression of the written record. It hands you the consensus, very fast, very well phrased.

The trouble is that the consensus is worth almost nothing. Not because it is often wrong, quite the opposite: it is right most of the time. It is worth nothing because everybody has it. Information the market already holds is in the price. An opinion everyone shares is in the valuation, in your competitor's strategy, in the budget that has already been voted. You gain nothing by restating it more elegantly than the next person.

Where the value of a decision actually sits

Imagine noting two things about every important professional decision you make: were you right, and how many people agreed with you at the time. You get four boxes.

Being wrong along with everyone else costs little, socially at least: nobody will hold it against you, it was the general view. Being wrong alone is very expensive, and it is more or less the only way to lose your job over an analytical error. Being right along with everyone else earns nothing: you did what anyone would have done, and the outcome was already in the price.

One box is left. Exactly one. Being right when the majority is wrong. That is where the entire return on judgment sits, and it is a narrow box.

Keynes summed up the practical consequence of that asymmetry almost a century ago, and nothing has disproved it since.

Worldly wisdom teaches that it is better for reputation to fail conventionally than to succeed unconventionally.

John Maynard Keynes, The General Theory, 1936

In other words: the box that pays is protected by a toll. The toll is not intellectual, it is social. You have to accept being wrong in public, alone, with your name on it.

What the machine cannot do, precisely

Precision matters here, because most talk about "what remains human" is a poorly argued consolation. A language model is not limited by a lack of soul, creativity or sensitivity. It is limited by something far more concrete: it can only return regularities present in what it has read.

That makes it excellent everywhere the past resembles the future. Drafting a clause, summarising a report, finding a precedent, producing the ten classic arguments in a classic debate: on that ground you have lost, and that is fine.

It makes it systematically average where the past says nothing about the future. When a situation is genuinely new, the average of what has been written about old situations is the wrong answer. And by construction, a model is pulled towards the middle of the distribution: that is where the bulk of its training sits. It is structurally reasonable. It is rarely contrarian at the right moment, not out of caution, but because the correct contrarian position is, at the very instant it would be useful, a minority view in the available text.

Note what this does not imply. It does not imply that humans forecast better in general: on many routine forecasting tasks, statistical models have beaten experts since the nineteen sixties, and that result is solid. It implies something more modest and more useful: the marginal value of a human has moved. It is no longer in producing an analysis. It is in choosing the moment to depart from the common one.

The spark, without the mysticism

Call it the spark, provided it is defined without poetry. It is not intuition, a word we mostly use to dress up our preferences. It is not creativity, which machines now produce industrially. It is this:

• holding a position the available text does not yet support;
• being able to say, before the outcome, what would change your mind;
• and accepting to be scored on it.

All three matter, and the third is the one we forget. A minority opinion that is undated, unquantified and never confronted with reality is not contrarian thinking, it is posturing. It is in fact more dangerous than the consensus, because it produces the same feeling of clarity with none of the correction mechanism. We all know people who were right against everyone, provided they get to pick which of their predictions we count.

That is the difference between being contrarian and being contrary. The contrary person opposes by reflex, which still means letting the crowd set their position, simply by taking the inverse. The contrarian follows the consensus most of the time, because the consensus is right most of the time, and departs from it at the precise moment it is wrong. That moment can be recognised, and recognising it can be trained.

Why it can be measured

A skill you do not measure does not improve, and judgment long escaped measurement for a simple reason: our memories rewrite themselves. After the outcome, everyone recalls having had a doubt. The effect is robust enough to have a name, hindsight bias, and it alone explains why thirty years of experience do not automatically produce thirty years of learning.

Forecasting tournaments provided the way out in the twenty tens, by imposing three constraints that look trivial and change everything. A question with a verifiable answer. A numerical probability, not an adverb. A date. From then on memory cannot cheat, and one thing becomes visible: some forecasters beat the others persistently from one year to the next. Persistence is the key, because it rules out luck. So we are not talking about a mysterious gift but about a skill, with the properties of a skill: it can be observed, compared and trained.

That body of work, notably Philip Tetlock's, also identified what sets the best apart. Not the degree, not access to information, not even raw intelligence. It is the habit of revising your beliefs in small steps, often, including when it stings.

What this changes on Monday morning

Three things, and they cost nothing.

1. Put numbers on your views. Replace "probably" with a figure. "Probably" is bulletproof: it is right whatever happens. A number can be wrong, and that is exactly its value.
2. Date them. A forecast without a deadline is not a forecast, it is a mood. Write down the date on which you will know.
3. Write what would change your mind, before you know. It is the only known protection against the art of proving yourself right afterwards, and the best way to discover that you had no reasons, only a preference.

Use AI for everything else. Genuinely everything else: it gathers the context, lays out both sides, retrieves the precedents and tells you what the majority thinks, in seconds and for nothing. That work no longer has market value, and clinging to it is a career mistake.

Keep the last inch for yourself, the one where you decide that on this precise point, for these precise reasons, the majority is wrong. That inch is not in the training data, for a simple reason: at the moment it matters, it has not been written anywhere yet.