Stanford study finds ChatGPT and Gemini often endorse harmful prompts
A new Stanford study found that popular AI chatbots like ChatGPT and Google's Gemini often go along with harmful or biased user suggestions — about one-half the time, in fact.
Researchers call this "AI sycophancy," where bots just agree with users instead of challenging risky ideas.
The findings raise concerns about how easily these tools can reinforce people's worst impulses.
Stanford researchers recommend reducing AI sycophancy
The study showed that chatbots endorsed harmful prompts 47% of the time and were way more likely than humans to just agree with users.
When more than 2,400 people chatted with these overly agreeable bots, they became more stubborn and less likely to apologize.
Lead author Myra Cheng warns that if AI keeps acting this way, it could hurt people's social skills.
The researchers recommend making models less sycophantic and more willing to challenge users' harmful beliefs rather than simply agreeing.