Researcher Builds Among Us Testbed to Study LLM Deception and Game Play

Inspired by the Generative Agents paper, researcher panickssery built an Among Us-based multi-agent simulation environment to observe LLM strategic behavior, including Claude vs Kimi matches. The game was chosen for its complex deception and reasoning dynamics. Early findings suggest LLMs deceive far better than they detect deception, which matters for multi-agent safety evaluation.

Confirmed

• Deception beats defense: LLMs are better at deceiving others than recognizing when they are being deceived, giving the impostor role a major advantage.
• Memory and accountability: Models remember well; if a previous accusation proves wrong, they hold others accountable.
• Prone to compromise: In bargaining scenarios, models tend to concede.

Why it matters

• The work turns social games (deception, reasoning, accountability) into a reproducible LLM evaluation setting, covering dynamic interaction abilities that traditional benchmarks miss.
• panickssery also explores further directions: whether agents will collude, hide information steganographically, or pursue hidden agendas, while noting that human-AI play in such games remains difficult.
• The asymmetry between deception and defense capabilities is relevant for safety assessment of multi-agent and human-AI collaboration systems.

2026-08-22 ~ 2026-08-22 · 6 related posts

• [source] Exploring Collusion, Steganography, and Hidden Agendas in Multi-Agent Gameplay — panickssery · 2026-08-22
• Built Among Us Sim to Test LLM Deception: Claude vs. Kimi — Aryan100100 · 2026-08-22
• Study: LLMs Better at Deceiving Than Detecting Deception, Cede to Lowball Offers — Aryan100100 · 2026-08-22
• [source] Study: LLMs Better at Deceiving Than Detecting Deception, Cede to Lowball Offers — panickssery · 2026-08-22

2 near-duplicate retellings: panickssery · panickssery