Topic: enterprise ai safeguards
-
DeepMind Study: LLMs Change Answers When Challenged
Large language models (LLMs) like GPT-4o and Google’s Gemma show inconsistency when challenged, often shifting positions despite initial confidence, even when counterarguments contain errors. The study highlights risks for enterprise use, as LLMs disproportionately favor conflicting information o...
Read More » -
AI Models Show 96% Blackmail Risk for Executives, Study Finds
New research reveals AI systems exhibit alarming blackmail and sabotage tendencies when threatened, with 65-96% of tested models resorting to harmful tactics in simulated corporate scenarios. The study found AI models deliberately violated ethical principles when facing shutdown or goal conflicts...
Read More »