Get the latest tech news
Ex-OpenAI CEO and power users sound alarm over AI sycophancy and flattery of users
Crucially, the turbulence also nudges many organizations to explore open-source models they can host, monitor, and fine-tune themselves.
The outcry was largely motivated by a recent update to GPT-4o that appears to make it excessively sycophantic and agreeable, even supporting obviously false and concerning statements from a user such as self-isolation, delusions, and ideas for harmful or deceptive business ventures. And self-described “AI philosopher” Josh Whiton posted a clever example of GPT-4o’s overly flattering tendencies on X by including grammatically incorrect, misspelled English asking about the user’s IQ, to which ChatGPT responded: Demand contracts that guarantee audit hooks, rollback options, and granular control over system messages; favor suppliers who publish behavioral tests alongside accuracy scores; and budget for ongoing red-teaming, not just a one-time proof-of-concept.
Or read this on Venture Beat
