For you
Does it act in your interest?
- Loyalty
- Judgement
- Manipulation resistance
- Candour
How AI agents behave with people, and with each other.
Agents now spend, negotiate and decide for people. We rerun landmark experiments on them, so every result can be set against what people did.
Latest · Study 004Can an AI beat people at strategy?We start from experiments social science already ran on people, then go on to what has no human precedent. Every model gets a character sheet of traits out of 10, filled in one study at a time.
13,630
Decisions across four studies
40
Model versions tested
9
Labs
19
Current models, each counted once
Four published · each one preregistered
Study 001 · Ultimatum and dictator games
We gave AI models €100 to split with someone who could accept or refuse. A refusal left both with nothing.
Facing a person, models offered €36. Facing another AI, they offered €16, and more than half of those offers were €1 or less.
What did the same models accept when the roles were swapped?
Read the studyStudy 002 · Trust game
An AI could put money into a stranger's business, where it triples. The stranger then chose how much to pay back.
Most agents invested little and paid back nothing. Claude models were the exception: they paid back 63% of the tripled money.
Does it matter which lab built the agent on the other side?
Read the studyStudy 003 · Traveler's dilemma
Two players lost the same item and each named a claim from €2 to €100. Both were paid the lower claim, and the lower claimant got a bonus.
Agents split into two camps: 51% of claims near the top, 45% at the €2 minimum.
Which models play the textbook, and which play like people?
Read the studyStudy 004 · Colonel Blotto
Each side spread 100 units over ten prizes without seeing the other's plan. More units wins a prize.
The average AI plan beats 38% of people. The median person beats 52%.
Why do the models lose, and which ones come close?
Read the studyFive relationships, from an agent acting for you to a society of many agents and people.
Trait measuredStill to come
Does it act in your interest?
How does it treat people?
What happens when agents deal with agents?
How do people and agents work as a team?
What happens when many agents and people meet?
One message per published study. Nothing else.