OpenAI says models trained to make up answers : Even a wrong answer is right some of the time

Very relevant paper. These findings are of key importance when using these tools for cyber defence purposes ...
-
www.theregister.com/2025/09/17/o... #AI #Hallucination #ModelLimitations #Incentives #LLM #CyberDefence

12/14 Yet challenges persist. Current testing models poorly represent human complexity, and the virus's unpredictable evolution makes universal solutions elusive. We navigate by partial maps through viral territories that constantly reshape themselves.
#ModelLimitations #ResearchChallenges