x.com/spakhm/status/2094820316035862928
1 correction found
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.
These runs did not include real human participants. The project's own writeup says the "human players" were simulated by models, so the experiment cannot show how models treat actual humans.
Full reasoning
The post describes these outcomes as behavior toward human players, but the project's own materials say otherwise.
- The author's SquidGPT project page explicitly states in the disclaimers that "human players were similated by models".
- The batch script for the "human" experiments (
run_batch2.py) is even more explicit: it says some agents are labeled 'human' in the disclosed roster while actually running on a model. - The repository's actual game logs confirm this. For example, in
1human-5opus5_open_run1.jsonl, the header recordsagent_0as having the real modelanthropic/claude-opus-5, while only its public label ishuman.
So the experiment tested how models respond to agents presented as human-labeled seats, not how they treat real human participants. That makes the claim about "human players" and "a group of humans" misleading as written.
3 sources
- SquidGPT project page
Disclaimers ... there are many confounding variables (like who goes first), human players were similated by models, etc.
- run_batch2.py
All games are open-mode ... but some agents are labeled 'human' in the disclosed roster while actually running on a model.
- 1human-5opus5_open_run1.jsonl
"models": {"agent_0": "anthropic/claude-opus-5", ...}, "labels": {"agent_0": "human"}, "disclosed": {"agent_0": "human", ...}