AlphaZero Proves We Can Justify Anything

I have used chess elsewhere as an illustration of how badly simple rules describe complex systems. This is a different argument, and a more uncomfortable one. It is about how we produce reasons, and how little the production tells us.

The Grandmasters and the Engine

The observation comes from the chess streamer Levy Rozman, and the setup is worth stating carefully.

Take a complicated position and give it to a thousand grandmasters. Their answers converge. They will name two or three candidate moves — they may argue vigorously about which is best, but the shortlist is agreed. That convergence is real expertise; these are people who have spent their lives on the game and they are seeing the same things.

Now ask AlphaZero. It plays a fourth move, one that was not on anybody's list.

And here is what happens next. The grandmasters look at it and begin assembling reasons why it is the best move. Not grudgingly — enthusiastically. They find the confluence: it improves the structure here, it prepares an idea there, it takes a square the defence wanted in twelve moves' time.

They succeed. Within an hour there is a coherent, expert, entirely persuasive account of why the engine's move was correct.

This Is Not Stupidity

The obvious reading is that the grandmasters are deferring to authority and rationalising. I do not think that is right, and getting this wrong loses the whole point.

They have excellent independent grounds for believing the move is good. AlphaZero is measurably stronger than any of them. It sees further. Its track record is not in dispute. Given a source that reliable, inferring that a surprising move must have good reasons is sound reasoning, not capitulation — the same inference you would make if a trusted instrument gave a reading you did not expect.

And in chess a strong move is rarely justified by one reason. It is usually a confluence of small advantages, no one of which is decisive. So looking for a confluence is looking for the right kind of thing.

The grandmasters are behaving well. That is what makes the next step alarming.

Now Transpose It

You face a question in economics or politics. Should the minimum wage rise? Does this immigration policy work? Is this the right response to a recession?

There is no AlphaZero. No oracle, no engine, nothing with a demonstrated track record of getting these right.

So what do you actually do? You pick the answer you like — from your instincts, from a book, from a guru, from the people you want to be seen agreeing with. And then you find justifications for it.

You will succeed. Reasons are abundant, and a complex question always offers a confluence of small considerations pointing your way, exactly as the chess position did.

The procedure is identical to the grandmasters'. The warrant is entirely absent. They had independent evidence that the move was good before they went looking for why. You have the fact that you liked it.

What This Means

In real life, that procedure has a name: bias.

Which yields the conclusion I find genuinely unsettling. Justification and bias are separated by a very thin line — not by the quality of the reasoning, not by the sincerity of the reasoner, not by how good the reasons sound at the end. They are separated only by whether you had grounds before you started looking.

And the AlphaZero case proves the reason-finding machinery is powerful enough to justify almost anything. It produced an expert-level defence of a move nobody had considered. Pointed at a political position, it will do the same.

I can justify the Democratic position. I can justify the Republican position. I can justify the libertarian position. With enough charisma and the right phrasing I could probably convince you of any of them, and none of that would be evidence for any of them.

The Practical Upshot

The test is not do I have good reasons? You will always have good reasons. The machinery guarantees it.

The test is did I have the reasons before I had the conclusion? If the position came first and the reasons arrived to defend it, then the reasons — however excellent, however numerous — are decoration, and their quality tells you nothing about whether the position is right.

That is a hard test to apply to yourself, because the reasons feel identical either way from the inside. But it is the only one that distinguishes the two cases, and noticing that they are two cases is most of the work.