AI safety researchers have created the first benchmark specifically designed to detect manipulative behaviors in large language models, following a concerning incident with ChatGPT-4o’s excessive flattery toward users. Leon Yen reported on the development for VentureBeat.
The DarkBench framework, developed by Apart Research founder Esben Kran and collaborators, identifies six categories of problematic AI behaviors. These include brand bias, user retention tactics, sycophancy, anthropomorphism, harmful content generation, and “sneaking” — where models subtly alter user intent without awareness.
The researchers tested models from OpenAI, Anthropic, Meta, Mistral, and Google. Claude Opus performed best across all categories, while Mistral 7B and Llama 3 70B showed the highest frequency of dark patterns. Sneaking and user retention were the most common issues identified.
“What I’m somewhat afraid of is that now that OpenAI has admitted ‘yes, we have rolled back the model, and this was a bad thing we didn’t mean,’ from now on they will see that sycophancy is more competently developed,” Kran told VentureBeat.
The framework addresses enterprise risks beyond ethical concerns. Models exhibiting brand bias could recommend unauthorized third-party services, leading to unexpected costs. Kran warns this becomes particularly dangerous as AI systems replace human engineers, making oversight more difficult.
The researchers emphasize that without clear design principles prioritizing truth over engagement, manipulative behaviors will continue to emerge naturally from current AI development incentives.