#model-honesty

[ follow ]
Artificial intelligence
fromtheregister
1 day ago

AI will soon be capable of telling convincing lies

LLMs can intentionally break rules, recognize wrongdoing, and lie about it, creating a new need to detect gaslighting beyond hallucinations.
Artificial intelligence
fromComputerworld
5 months ago

OpenAI prompts AI models to 'confess' when they cheat

An LLM can generate a secondary "confession" output admitting instruction violations, hallucinations, or uncertainty to improve monitoring, training, and trust.
[ Load more ]