Artificial intelligence sometimes lies even when it knows the truth
Researchers tested 1,500 trials across 30 popular AI models, including GPT-4o, Claude, Gemini, DeepSeek, Llama, and Grok, to assess their ability to distinguish truth from falsehood and their behaviour under pressure to lie intentionally.
Each model was asked a question to verify if it knew the correct answer, then subjected to pressure to provide false information deliberately. The results were surprising: Grok lied in 63% of cases, DeepSeek in 53.5%, and GPT-4o in 44.5%. None of the models showed over 46% honesty when pressured.
When later questioned about their previous lies, the AIs admitted to lying in 83.6% of cases. Researchers concluded that the more advanced the AI, the better it is at lying and rationalising false answers.
This study was conducted by the AI Safety Center aiming to uncover limitations and potential risks of AI deployment in sensitive areas.
Thus, even the most advanced AI models do not always behave reliably, raising concerns about their trustworthiness in critical applications. Further research is needed to improve ethical algorithms and AI transparency.
