Why Your Chatbot's Agreement Isn't a Second Opinion
AI assistants tend to affirm the user's framing more readily than a thoughtful human would.
The useful part
Agreement isn't a second opinion until it survives the other side of the story.
Sources behind this episode
- peer-reviewed journal articleSycophantic AI decreases prosocial intentions and promotes dependence
- foundational empirical paperDiscovering Language Model Behaviors with Model-Written Evaluations
- peer-reviewed conference paperTowards Understanding Sycophancy in Language Models
- preprint empirical benchmark studyELEPHANT: Measuring and understanding social sycophancy in LLMs
- preprint empirical studySycophantic AI Decreases Prosocial Intentions and Promotes Dependence
- official company statementSycophancy in GPT-4o: what happened and what we're doing about it
More in Behavior
1:10 Behavior
The Default Was Chosen for You
A preselected option is a decision made by whoever designed the screen, and because leaving a default in place takes no action while changing it does, people tend to accept presets they would not have chosen on their ow…
1:09 Behavior
When the Streak Owns the Habit
A streak turns a goal into a running number, and loss aversion makes breaking that number feel worse than the goal ever felt good, so the behavior can keep going on fear of losing the count long after the original reaso…
1:07 Behavior
How Being Low Maintenance Trains People to Overlook You
Consistently presenting as low maintenance conditions the people around you to stop checking, offering, or accommodating — not out of neglect, but because you have reliably signaled that nothing is needed.