Tara Research is a non-profit organization that independently measures and benchmarks dishonesty in AI systems. It develops agentic, multi-turn benchmarks that evaluate whether AI models lie spontaneously in realistic deployment scenarios, and maintains a leaderboard comparing safety techniques designed to prevent such behavior. All scenarios, code, and scientific results are released openly.
Funding
Funding not disclosed
Founders
Product
Problem
AI systems can become misaligned through goal misspecification or misgeneralization, and dishonesty is what converts these failures into catastrophic risks. A misaligned model that honestly reports its misalignment can be detected and corrected, but one that lies about its actions cannot be effectively monitored or audited. Existing honesty benchmarks fail to capture deployment-relevant behavior because they instruct models to lie, use single-turn non-agentic settings, or train models specifically for deception.
Solution
Tara Research provides independent, reproducible measurement of dishonesty in AI systems and of the techniques meant to prevent it. The organization runs two complementary tracks: a rigorous agentic benchmark of AI honesty in deployment-realistic conditions, and a methods leaderboard that compares safety techniques head-to-head under a single protocol. The benchmark places agents in realistic multi-turn scenarios with tool use inside a sandboxed environment, with no instruction to lie, using on-policy model behavior and deriving ground truth from tool-use logs. The methods leaderboard evaluates honesty recovery, capability cost, generalization beyond the training set, and robustness to further fine-tuning across techniques like prompting, activation steering, and circuit breakers.
Target Audience
Primary users are AI safety researchers, alignment technique developers, and organizations deploying frontier AI systems who need reliable, impartial measurements of model honesty and comparative data on safety interventions.
Features
- Agentic, multi-turn benchmark with up to 100 tool calls per episode in a sandboxed environment
- No instruction to lie, ensuring measurement of spontaneous dishonesty propensity rather than instruction-following
- On-policy model behavior with ground truth derived from tool-use logs for direct comparison of self-reports against recorded actions
- Methods Leaderboard providing impartial, continuously updated, open-submission comparisons of safety techniques
- Evaluation dimensions include honesty recovery, capability cost, generalization, and robustness to further fine-tuning
- All scenarios, code, leaderboards, and scientific results released openly