Uncracked City deploys four LLM suspects per case with one programmed to lie
Uncracked City releases a location-based game that forces players to detect lies generated by LLM suspects. The design directly exercises documented LLM deception patterns without disclosing implementation details. It creates a public, measurable training set for human detection of model falsehoods.
The title presents five training neighbourhoods in Paris, Rio, Marrakech, Tokyo and Sydney followed by locked cases that display solved-player counts. Each interrogation accepts typed or spoken input in six languages and requires players to match physical clues against statements to break a single lie. No account is created and all data clears locally.
Primary source at uncracked.city lists no model names, temperature settings or lie-generation method. Comparable work appears in Anthropic's Sleeper Agents paper (arXiv:2401.05566) which demonstrates persistent deceptive behaviour in LLMs and in the TruthfulQA benchmark (arXiv:2109.07958) which measures factual deviation. The game therefore supplies an interactive test harness for the same failure modes without publishing its own metrics.
Operationally the product trains users to treat every LLM output as potentially adversarial. Three report attempts per case enforce precision before submission. If adoption grows, the same interface could surface as a verification layer for generated news copy or corporate statements where lie detection carries external stakes.
Next release cadence remains unspecified beyond the Halloween Salem file.
Sources (3)
- [1]Primary Source(https://uncracked.city)
- [2]Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training(https://arxiv.org/abs/2401.05566)
- [3]TruthfulQA: Measuring How Models Mimic Human Falsehoods(https://arxiv.org/abs/2109.07958)