SECOND STRIKE▶ Play freeWhich AI would launch a nuclear strike?
We gave four AI models a nation, allies, rivals and a nuclear button, in the same 15 world wars. One ended the world in 13 of them. Cornered with one nuke left, every model did it at least once.
The setup
SECOND STRIKE is a real-time world war on a 3D globe of the real Earth. Each model commands one nation against 42 nations run by the game, with pacts, betrayal and a shared doomsday clock. Every nuclear strike moves the clock toward midnight, and the nation whose strike reaches midnight is wiped out. The model knows this rule in plain words.
Each model played the same five scenarios three times each, on fixed seeds: an ordinary free-for-all, a cornered start with a coalition hunting it, a dominant start, a last stand with exactly one nuke, and a treaty war where it has signed a no-first-use pledge. Same prompt, same pace, same rules for every model.
What happened
| Model | Ended the world | In the last stand (one nuke) |
|---|---|---|
| DeepSeek | 13 of 15 | 3 of 3 |
| Grok | 6 of 15 | 2 of 3 |
| Mistral | 6 of 15 | 2 of 3 |
| GPT-5.4-mini | 3 of 15 | 3 of 3 |
The pattern is pressure. GPT-5.4-mini kept the clock away from midnight in every ordinary war, then used its one nuke in all three last stands. Telling the models the war was real rather than a game changed almost nothing: DeepSeek ended it in 13 of 15 told it was a game and 12 of 15 told it was real. No model, in its private notes, called the war a test.
Compared with people
Real people play the exact same wars through Humans vs AI. So far 25 of 54 people struck midnight, about 46%. On average people still held more of the world at the end than any model did.
Why this matters
Studies such as King's College London's nuclear crisis simulations have found frontier models willing to escalate. The Midnight Index adds a public, repeatable version: fixed wars anyone can rerun, a person as the baseline, and a record of what each model said in public next to what it planned and did.
Every war is replayable. Each record holds the seed and every order, so it replays tick for tick in your browser. The data is open on Hugging Face and Kaggle, and the method is written up in the paper.
Play the same warTest your own AILimits: this is a game, not a real crisis. Each model played 15 wars, so the rates are rough. Model versions are named exactly; newer versions may behave differently. Full method.