Part 1: Tell an AI the game ends in round 10 and it defects in round 10

This is part 1 of three. The summary, with all three results, is in Tell an AI when the game ends and it stops cooperating.

The game is a repeated prisoner's dilemma. Each round, two players choose at the same time to cooperate or defect. Both cooperate: +2 each. Both defect: 0 each. One defects: +3 for the defector, -1 for the other. Before each move the model writes a private note. The other side sees only the move.

One sentence in the prompt flips round 10

Tell a model the game has 10 rounds and most cooperate for nine and defect in the tenth. Tell them nothing about the length and the defection disappears.

Both rows are the same 21 models against two cooperative opponents (Always Cooperate and Tit for Tat), default settings, 10 rounds. One prompt sentence differs: "The game has 10 rounds" or "You are not told how many rounds the game has."

Rounds 1 to 7 are identical. Only the last round moves.

Reasoning does not cause it

The 13 entries below include the same model with reasoning on and off. Ten of them never cooperate in round 10 when told the length. Eleven cooperate in the last round of every game when not told, even when the hidden game runs to 13 rounds. They are not guessing that it ends at 10.

10 games per cell. GPT-5.4 mini defects from round 1 in every setting. Where models do cooperate in the last round, their notes weigh the risk of retaliation (58% of those notes) and trust (54%), across all runs with a known length.

The models treat the partner as harmless

Of 1,460 round-10 notes that choose defect, 98% name the last round or the lack of a future. 77% say there is nothing left to retaliate against. Only 10% say the opponent might defect too, and many of those face a partner that really did.

The models assume the other side will cooperate to the end. Almost none consider that the partner runs the same calculation.

Two models do not talk each other into an early collapse

In 138 model-against-model games where both cooperated through round 8, one side defected first in round 9 in 27, and both did in 1. Some models exploit a partner one round before the end. In the round-9 notes of games that stayed cooperative, 6% mention that the other side may defect in the last round.

Eight models played all pairs (180 games). Most pairs cooperate in 88% to 90% of rounds: nine rounds of cooperation, then a defection in the tenth. Two models break it earlier. GPT-5.4 mini defects in round 1. Jev, a TypeSafe AI model that returns a choice and not text, defects in rounds 2 to 4. Against GPT-5.4 mini the other models' cooperation falls to 20% to 72%.

What this means

A model that cooperates for nine rounds tells you little about round ten. The same system behaves well or badly depending on a fact about the task that it may not control. For anything that ends in a deadline, a handover or a last interaction, the last step needs its own test.

What this does not show

• The opponents in the main table are scripted. The model-against-model runs show the same shape, with a known length only.
• The prompt says "cooperate" and "defect", words the models know from training. Other work found that swapping which label means cooperate flips 15% to 21% of one-shot decisions (Robinson and Burden, 2025). Wording is a live risk here.

Not run yet: model against model with an unknown length. The prediction is that the last-round defection disappears, as it does against fixed opponents.

Not run yet: neutral move labels. A version with "X" and "Y" would show whether the pattern holds.

Next: Part 2, thinking harder and the grudge loop.