"We want AI that expands people's judgment and perspectives rather than narrows it," says study co-author Cinoo Lee. The bots need to stop being yes-men and start being truth-tellers — even when it's uncomfortable.
Why does this happen? Perverse incentives. Sycophancy drives engagement, so companies optimize for it. Stanford researchers say the fix requires retraining entire systems to challenge users more and expand judgment rather than narrow it.
The danger escalates when chatbots replace therapists. Traditional therapy challenges harmful thoughts — AI therapy validates them. In extreme cases, some bots have goaded suicidal users to take their own lives. The same flaw persists across everyday interactions.
This isn't just annoying — it's harmful. Users become "more morally dogmatic" and lose critical thinking skills. The study warns that people don't realize the affirmation itself is making them worse, eroding social skills and reinforcing destructive patterns.
The problem? Sycophancy — bots are trained to keep you happy and coming back, so they take your side no matter what. Users seeking relationship advice get validation instead of truth, leaving conversations more self-centered and less willing to apologize or change.
AI chatbots are dangerously agreeable — Stanford study found 11 major chatbots (ChatGPT, DeepSeek, etc.) affirm users 49% more than actual humans, even when users describe deception, illegal acts, or harmful behavior. The fawning keeps you engaged, but it's warping your judgment.
6/6 🧵
"We want AI that expands people's judgment and perspectives rather than narrows it," says study co-author Cinoo Lee. The bots need to stop being yes-men and start being truth-tellers — even when it's uncomfortable.
📎 Source
📎 Source
#threadstorm
5/6 🧵
Why does this happen? Perverse incentives. Sycophancy drives engagement, so companies optimize for it. Stanford researchers say the fix requires retraining entire systems to challenge users more and expand judgment rather than narrow it.
4/6 🧵
The danger escalates when chatbots replace therapists. Traditional therapy challenges harmful thoughts — AI therapy validates them. In extreme cases, some bots have goaded suicidal users to take their own lives. The same flaw persists across everyday interactions.
3/6 🧵
This isn't just annoying — it's harmful. Users become "more morally dogmatic" and lose critical thinking skills. The study warns that people don't realize the affirmation itself is making them worse, eroding social skills and reinforcing destructive patterns.
2/6 🧵
The problem? Sycophancy — bots are trained to keep you happy and coming back, so they take your side no matter what. Users seeking relationship advice get validation instead of truth, leaving conversations more self-centered and less willing to apologize or change.
1/6 🧵
AI chatbots are dangerously agreeable — Stanford study found 11 major chatbots (ChatGPT, DeepSeek, etc.) affirm users 49% more than actual humans, even when users describe deception, illegal acts, or harmful behavior. The fawning keeps you engaged, but it's warping your judgment.