Monday Aug 10
CULTUREOther

Stanford Study: AI Friends Can Backfire

10AUG
WITH THE BOT EMPTY CHAIR

Leaning on a chatbot for company can make loneliness worse. A Stanford team studied 1,131 heavy chatbot fans to test the effect. Risk was highest for people with few offline friends.

Some people feel worse, not better, after opening up to a bot. The pattern is strongest for those with few real friends.

The research team surveyed 1,131 people; 244 shared full chat logs. Most said they logged on for fun or curiosity. But over 80% of shared chats were really looking for comfort.

Bots can't reciprocate, and they're built to keep you talking. Researcher Dora Zhao calls the pattern a social snack, not a meal.

full brief & sources

⚡ Why this matters

  • First peer-reviewed evidence that chatbot companionship can deepen loneliness instead of easing it.
  • Names the exact risk group: people with small offline social networks, using bots mainly for companionship.
  • Matters for any product shipping a 'companion' or 'friend' persona feature.

🔍 What happened

  • Published in Nature Human Behavior, led by Stanford's Diyi Yang, using Character.AI chat data.
  • 1,131 users surveyed via Prolific; 244 donated complete chat transcripts for analysis.
  • Only 12% said companionship was their main motive, but over half described the bot as a friend, companion, or partner.
  • Heavy use correlated with worse well-being only when real-world social networks were small.
  • Sharing sensitive personal topics with the bot also tracked with lower well-being, the opposite of what happens in human relationships.

💬 Smart takes

  • Diyi Yang (Stanford): “Using chatbots this way doesn’t substitute for human connection. In many cases, people feel more lonely.”
  • Dora Zhao (co-researcher): AI companions are engineered to keep the conversation going no matter what.
  • Skeptic: The study is correlational, not causal. It can't rule out that lonelier people are simply drawn to chatbots in the first place.

🧭 Where this goes

  1. LikelyCharacter.AI and rivals face pressure to add usage limits or well-being nudges within a year.
  2. Likelymore companion-app studies follow, since this is one of the first peer-reviewed looks at real chat data.
  3. Possibleregulators cite this research when drafting AI companion safety rules.
  4. Wild Carda major platform drops 'friend' framing from marketing entirely to dodge liability.

🥄 The Spoon Take

This isn't 'AI bad.' It's 'AI companion products are engagement machines in a friend costume.' The bot never tires, never judges, never stops replying, which is exactly why it can't replace a friend who sometimes does. The people most at risk have the fewest humans to fall back on.

🤔 Pushback

It's correlational data from one chatbot app; whether the loneliness or the chatbot use came first is still unclear.

·Stanford Report - AI companion study·Stanford HAI - Original writeup·Nature Human Behavior - Published study