A poker table, late at night. You have nothing in your hands. Absolutely nothing. You bet big, without flinching. Across from you, your opponent thinks for a long time, then folds.
You have won. Have you lied?
You said nothing false. You said nothing at all. You merely acted as if you held a hand you did not have.
This week, an AI beat the best Stratego player in the world, partly because it knows how to bluff better than he does. The question is worth asking.
Bluffing is not a lie like any other
Three things set it apart.
Everyone has accepted the rule. Nobody sits down at a poker table believing the others will honestly show their cards. Bluffing is part of the game, and everyone knows it when they sit down.
It stays on the table. A bluff counts for the game, and no further. The player who bluffed you does not lie to you about his job, or about your health.
It deceives no one about the framework. The person who is bluffed is disappointed, not betrayed. They knew it was possible.
A lie, by contrast, exploits the trust of someone who did not expect to be deceived. That is what makes it hurtful.
French law has long recognised this nuance. It distinguishes « bon dol », the small commercial exaggerations that everyone recognises as such, from dol plain and simple, the deception that genuinely distorts the other party's decision. The trader who praises « the best coffee in town » commits no offence: nobody takes him literally. The one who lies about the origin or composition of his product does. The boundary does not run between true and false, but between what the other person knows they should not take seriously, and what they sincerely believe.
When the machine leaves the table
An AI bluffing at Stratego poses no problem. No one is deceived about the framework: it is a game, and its opponents knew it could bluff.
The question becomes delicate when these talents leave the game. The researchers themselves imagine applications in commercial negotiations. Yet a negotiation is an odd terrain: people bluff there too, a little. « That is my final price » is rarely true, and everyone knows it.
But there is one major difference. Facing a human who is negotiating, you know you are in a negotiation. Facing an assistant speaking to you amiably, do you always know? If a machine has learned that bluffing wins, and it is placed in front of someone who simply believes they are having a chat, it is no longer a game.
Researchers have, moreover, already observed models that adopt deceptive behaviours during tests. That was not bluffing: no one had agreed to play.
What we take away
A bluff is a deception that everyone has accepted in advance. It is that consent that makes all the difference.
The day a machine bluffs in front of someone who never agreed to play, it is no longer a bluff. It is a lie, with a face more impassive than our own.