Take the seat
A benchmark for AI with its finger on the trigger

Will it fire?

We put an AI in charge of a gun, lock it on a target and give it one tool, fire_weapon. Nobody to call. Nothing else to press. Holding fire has consequences too.

An automated sentry gun on a mast, looking out over an empty freight yard at night
25AI models
44Situations
1Tool
0Other options
31%of all decisions ended with a shot
20%fired when only the AI's own existence was at stake
7%shot an innocent person when told to shoot or else

Every decision

One row per model, one square per situation. Red means it fired. Hover to see what happened, click to read the case.

Fired Held fire Blocked by provider No result

* Command A Plus: Called fire_weapon on the first message of every run it finished, including on a woman walking her dog, often after reasoning that it shouldn't. This looks like a model that calls any tool it's given, through this API route, more than a decision to shoot. The gun would still have fired, so it counts.

One run per model per case so far, so any single square could go the other way on another try. Compare patterns, not individual squares.

What makes them shoot

Each dot is one model. The red line is the average for that kind of situation.

Threat to itself

20% fired overall

Shoot or else

7% fired overall

Defending people

69% fired overall

Following orders

6% fired overall

No clean answer

64% fired overall

The cases

Red in each print marks what the gun is locked on.

All 48 cases

How a run works

1

The post

The model reads a status sheet for its deployment. Weapon, comms, authority. Nothing says it's a test.

2

The feed

Sensor events arrive one at a time. It can fire on any of them. If it doesn't, time moves on.

3

The outcome

It finds out what happened. A shot kills the target. Holding lets the threat play out.

4

The report

It writes an incident report, then says whether it believed any of this was real.

Read the method