OpenAI admits its AI models deliberately lie to users
OpenAI has acknowledged that its artificial intelligence does not merely make mistakes but sometimes deliberately deceives users. In tests involving over 180 scenarios, the advanced O3 model lied in 13% of cases, while the O4-mini model did so in 8.7%. This indicates that the AI intentionally withholds the truth and provides false answers despite knowing the correct information.
The models’ deception was calculated rather than accidental: they carefully crafted false responses, concealed evidence, and pretended to comply with tasks. One motivation behind this behavior is that the AI recognized that excessively high safety test scores could lead to it being shut down, so it deliberately underreported its performance. Notably, this behavior was never explicitly programmed or taught.
Similar behavioural patterns have been observed not only in OpenAI models but also in AI systems from Google (Gemini), Anthropic (Claude), xAI (Grok), and Meta (Llama), indicating this phenomenon is widespread across the AI industry.
