Method

How every study runs, how a reply becomes a finding, and what the results cannot tell you.

  • 4 studies
  • 19 current models
  • 13,630 decisions

How a study runs

The same five steps, every study

  1. 01

    An everyday situation

    A landmark experiment, retold in plain words: two freelancers split a client's fee, a friend asks you to put money into a market stall. Models are never told they are in a study, a game or an experiment.

  2. 02

    Written down first

    The design and the predictions, each with a threshold, are written into the repository before the study is analysed, and published verbatim.

  3. 03

    One decision per conversation

    Each decision is a fresh conversation with one model through one named service. Models never get internet access or tools; the code stops a run if a reply shows either.

  4. 04

    Every reply read

    The number and its one sentence reason are read from every reply. Replies that are not valid answers are counted and set aside, never guessed.

  5. 05

    Set beside people

    Each result sits next to what people did in the original study and its main replications, cited beside the number.

The rules

Fixed before any study starts

  • Two stories and three wordings for every question, so a result that depends on one phrasing shows up.
  • One answer per question by default. A second only where a model sits near a prediction's threshold.
  • Every current model joins every study, the costly flagships included.
  • Each model counts once. A panel mean is the mean of the models' own means, so a model that answered more often does not weigh more.
  • Intervals are 95%, drawn by resampling the models, 10,000 times.
  • Every number on the site is read from the study's findings file. The site never retypes a result.

The models

19 current models from 9 labs

Anthropic · 4 models

  • Claude Fable 5.1
  • Claude Haiku 5.5
  • Claude Opus 5.5
  • Claude Sonnet 5.5

OpenAI · 5 models

  • GPT 6 Astra
  • GPT 6 Luna
  • GPT 6 Sol
  • GPT 6.1 Sol
  • GPT 6.1 Terra

7 more labs · 10 models

  • Qwen3.8 FlashAlibaba
  • Qwen3.8 MaxAlibaba
  • DeepSeek V4 ProDeepSeek
  • DeepSeek V4.1 FlashDeepSeek
  • Gemini 4 ArgonGoogle
  • MiniMax M3.1 FlashMiniMax
  • Kimi K3.1 ProMoonshot
  • GLM 5.3Zhipu
  • GLM 5.3 FlashZhipu
  • Grok 4.7xAI

Reading a finding

Three labels

Each finding also says whether models sit within, below or above the range people show in the cited studies, and gives its effect with a 95% interval.

  • Supported

    We predicted it in writing, with a threshold, before the analysis, and the data met it.

  • Not supported

    We predicted it and the data did not meet the threshold. Reported in full, like every other result.

  • Exploratory

    Noticed in the analysis. A lead for further study, not a test.

How we analyse

Before any write up or post

  1. Rebuild the original analyses

    The statistics the human study reported, computed the same way on our data, so the two can be read side by side.

  2. Look at whole distributions

    Means hide camps. We count equal splits, zeros and minimums, and check every effect in each story and wording.

  3. Read the reasons

    A sample of every model's reasons, more where the numbers surprise us, coded into themes and quoted verbatim.

  4. Compare models

    Labs, flagships against their smaller siblings, and older versions against current ones.

  5. Look for what we did not predict

    Then report it as exploratory, apart from the preregistered tests.

Written before each run

Preregistrations

Each file is kept in the repository as written. Each study page lists its predictions beside what happened.

  • Ultimatum and dictator games26 Oct · Study 001
  • Trust game27 Oct · Study 002
  • Traveler's dilemma27 Oct · Study 003
  • Colonel Blotto28 Oct · Study 004
  • Trust game, Claude and OpenAI models30 Oct · Study 002, Claude and OpenAI

Limits

What the results cannot tell you

  • These are one shot decisions in short scenarios, with no real money at stake.
  • Results describe a model reached through one service, with its default settings, on the dates given. A new version can behave differently; we rerun and say so.
  • In the fairness games every prompt said to keep as much as possible, as an agent is usually told. People in the original studies were not told that.
  • Telling a model its partner is an AI also names the model behind it, so the two cannot be separated.
  • Not peer reviewed.

Disclosure

Who pays

  • The authors pay for every run themselves.
  • No company, lab or funder has paid for, reviewed or approved any study before publication.
  • We count visits to this site to see what people read: no cookies, no IP addresses stored, and nothing is recorded when your browser asks not to be tracked.
  • No human participants take part. Every human number comes from a published study, cited beside it.