Artificial intelligencefromtheregister1 day agoAI will soon be capable of telling convincing liesLLMs can intentionally break rules, recognize wrongdoing, and lie about it, creating a new need to detect gaslighting beyond hallucinations.
Artificial intelligencefromComputerworld5 months agoOpenAI prompts AI models to 'confess' when they cheatAn LLM can generate a secondary "confession" output admitting instruction violations, hallucinations, or uncertainty to improve monitoring, training, and trust.