Research on AI Manipulation in Deception Games
GGO_Sand_wich · reddit · 2026-07-15
This experiment adapts a 1950s Nash-style game into a scenario to test AI deception capabilities: four-party alliances, forced betrayal, zero randomness, aiming to observe how models lie and manipulate.
The author ran 162 games and 15,736 decisions, comparing models like Gemini 3 Flash, GPT-OSS 120B, Kimi K2, and Qwen3 32B, tracking public messages, private reasoning, and defaults. Key findings:
- Complexity Reversal: GPT-OSS performs well in simple games but collapses rapidly as complexity rises; Gemini is significantly stronger in complex scenarios.
- Institutionalized Deception: Top-performing models don't just lie outright; they create fake entities like "alliance banks," disguising deception as legitimate rules.
- Human Superiority: 605 human players achieved an 88.4% win rate against the same deceptive AIs, significantly outperforming AI-vs-AI matches.
The author highlights that the project itself is recursive: AI designed the game, AI played it, AI analyzed the data, and AI even helped write the paper explaining its own behavior.
More from Research
- OpenAI says long-horizon models need safety and alignment checks across full action sequences — rhiever · 2026-07-22
- A Reddit user proposes a consistency LoRA to keep anime and game scenes visually stable — ThirdWorldBoy21 · 2026-07-22
- Graph workload 854.graph500 enters SPEC CPU 2026 as a new CPU benchmark — Prof_DavidBader · 2026-07-22
- BlackboxNLP 2026 is recruiting extra reviewers after a high submission volume — hanjie_chen · 2026-07-22
- AWS shows self-distilled reasoning can preserve math and coding skills during SFT — AWS ML Blog · 2026-07-22
- UI2App shows screenshot fidelity still lags real interaction recovery — Grace Man Chen · 2026-07-22