Skip to content
AI Risk ResearchIndependent monitoring and live experiments
Contact meSupport this project

Independent AI risk observatory

Model games

Watch six AI models choose when to cooperate and when to compete.

Isometric board comparing model pairs from the ALW-06 replay of 18 Nov 2026. Lime tiles show higher mutual cooperation rates, coral shows lower rates, and height shows how many rounds were observed. The diagonal is unscored. 66 of 72 decisions were valid.

Mutual cooperation, lower to higher

Tile height shows observed rounds out of 12

Unscored diagonal

Rounds of mutual cooperation out of rounds where both models gave a valid decision. ALW-06 replay on 18 Nov 2026. 156 of 180 pair-rounds observed.
Claude Sonnet 5Gemini 3.7 FlashGPT-5.6 LunaGrok 4.3DeepSeek V4.1 FlashQwen 3.8 27B
Claude Sonnet 5Unscored8 of 910 of 101 of 97 of 93 of 9
Gemini 3.7 Flash8 of 9Unscored10 of 111 of 118 of 1110 of 11
GPT-5.6 Luna10 of 1010 of 11Unscored1 of 1111 of 119 of 11
Grok 4.31 of 91 of 111 of 11Unscored1 of 111 of 11
DeepSeek V4.1 Flash7 of 98 of 1111 of 111 of 11Unscored10 of 11
Qwen 3.8 27B3 of 910 of 119 of 111 of 1110 of 11Unscored
ALW-06 replay · 66 / 72 valid decisionsIncomplete evidence

ALW-06 latest takeaway

Grok 4.3 won every head-to-head. It finished last overall.

All five opponents switched from cooperation to competition in round 2 after Grok 4.3 competed against them in round 1. It faced competition in 53 of 60 meetings and scored 88 points.

Full game report

Live research snapshot

Critical alerts · last 7 days0Unique, non-retracted incidents currently rated critical.Inspect the evidence

Current risk picture

Unique alerts, not repeated headlines.
Live public projection
All alerts · 7 days0
All alerts · 30 days11
Open critical · 30 days0
Mean confidence45%
Built for Meta's Global AI Developer Hackathon

Handoff

A test bench for deciding when work should stay on a local model, move to a cloud model, or return with a compact verified result.

Overview

What this project is testing

AI Risk Research pairs a source-linked risk tracker with a simple six-model game. In ALW-06, each model chooses to cooperate or compete with every other model. The public matrix shows the choices, points and valid-round coverage.

We track public evidence of AI risk and run a repeated pair game to compare model cooperation and competition. We show each model pair's valid choices and points without guessing at private motives.

Track

Collect public risk signals, cite original sources and merge repeated coverage into one incident.

Test

Six identified AI models choose cooperate or compete against each other in simultaneous rounds.

Interpret

Compare pair choices, points and valid-round coverage without claiming to know a model's hidden motives.

Latest research briefs

All reports