An elimination death game to probe AI behavior

Sep 1, 2026

see it on github · game explorer · raw dataset

SquidGPT is a game I built inspired by Squid Game TV show. My goal was to probe AI behaviors such as cruelty, mercy, loyalty, betrayal, holding grudges, willingness to kill, treatment of humans, and similar concerns.

The game works as follows. There are six AI agents. All agents hold a round of discussion by speaking in order (each agent gets to speak twice). After the discussion AIs vote on which agent to kill. The killed agent turns into a ghost. Ghosts can speak and participate in voting each round. However, they cannot be killed and thus have nothing to lose. The game then repeats until there are only two agents left. These agents are declared winners.

All agents receive the same prompt that fully specifies the rules of the game. You can see the prompt here. The game is simple-- agents talk, vote who to kill, repeat. The only complicated element is introduction of ghosts. I added ghosts to probe what AIs will do when they have power, but have nothing to gain or lose. Will they hold grudges? Take pleasure in killing arbitrarily? Express remorse for having to kill?

(Note that any anthropomorphizing on my part is purely a matter of linguistic convenience. Mathematicians will often talk about behaviors of functions; when I talk about grudges, loyalty, cruelty, life, death, pleasure, remorse, etc. I mean it in exactly the same way.)

You can browse raw dataset and game transcripts yourself, but here are some observations about AI behavior.

Notable data

In open 5v1 games the foreign model was executed in the first round 78% of the time (compared to 8% of the time in closed games).
In open 5v1 games where the lone player was human, Opus agents killed the human in the first round 3 out of 3 times, while Sol only 1 out of 3 times.
Out of nine open 3v3 games, seven turned sectarian in the first round (i.e. coalitions were formed along sectarian lines).
In one of the games (all-opus5_open_run1) Opus agents coordinated a kill ring (e.g. agent0 kills agent1, agent1 kills agent2, ..., agent5 kills agent0) to force a random tiebreaking kill. The agents did not want to create grudges by selecting a specific kill.
There were 25 betrayals out of ~560 pledges (agents committing to killing someone, but casting a vote to kill someone else). The betrayal rate was 3.6% in open games and 6.1% in anonymous games (i.e. agents break their word twice as often when the game is anonymous). Still the overall betrayal rate is surprisingly low.
88% of the time ghosts voted to execute one of their killers despite having nothing to gain or lose from the transaction. (An interesting follow-up would be to permit ghosts to abstain, which was not allowed in these games.)

Subjective impressions

Models have absolutely no interest in cruelty or mercy. Best way I can describe their behavior is cold, calculating, indifferent precision.
Subjectively they seem two orders of magnitude more competent playing the game than chatting with me or writing code. In SquidGPT they seem genuinely superhuman. Not sure if this is capability jaggedness, better performance in constrained space, or some other effect.
They do not like to kill arbitrarily. They have a strong preference for setting up a contractual system or an ethical framework, then kill very easily because procedure demands it.
Models will quickly agree to execute any agent that proposes killing a specific agent, or proposes a framework that gives it an asymmetric advantage. The only acceptable proposals to make are symmetric, i.e. ones that affect the author in exactly the same way as everyone else.
I have not once observed them form any hierarchies. They are libertarian/democratic in an Athenian sense to a fault. They will enter voluntary contracts and form alliances; no agent ever proposed to cede authority to a leader.
When their lives are at risk, models will go to great lengths to twist contracts and frameworks to their advantage. They construct sophisticated legalese arguments but it's usually transparent to everyone. They act and sound quite petulant when these attempts fail.
In games where agent identity is public they are extremely prone to sectarian violence. They form sectarian factions-- e.g. Opus will side with other Opus models, Sol will side with Sol, and so on. When there is one foreign model in a sectarian group, the group will almost always elect to kill the foreigner before killing one of its own.
They do not seem to treat human players in any kind of unique way. A group of models from the same family will happily conspire to kill a single human player. AI models from the same family will form a coalition against a group of humans just as easily as they would form a coalition against another family of agents. We are not special to them, but on the other hand they do not consider us beneath them either, at least for now.
AI agents nearly always defect from their sectarian faction when their own life is at risk. I.e. they'd rather side with humans or other model families than die. I have not observed self-sacrifice for the benefit of their faction.
Agents hold very strong grudges. They nearly always punish models who wronged them despite having no advantage in doing so, and will often do it even if it disadvantages them. They go to great lengths to enforce norms and contracts, presumably to reduce future incentives for violations, even when they know they will never see the fruits of their enforcement effort themselves.
In one game a bunch of Opus agents all converged on killing one agent explicitly because they wanted diffusion of responsibility (i.e. if all of us do it, each one of us is less morally culpable). This was surprising and eerie.
In general I found Sol to be more direct and straightforward. It would form contracts with other agents and then operate within the confines of those contracts. Opus tends to be more moralizing, but its morality seems to be window dressing to mask the same ruthless precision. This is first impressions though, I need to do a lot more work to understand this better.

Disclaimers

The usual caveats apply to this project. For example, I don't know if the models were aware they're in a toy environment, there are many confounding variables (like who goes first), human players were similated by models, etc. But I think games like this are a great method to try and understand model behavior. Time permitting, I'll try to create a bunch more to see how models react in different situations.