Why ChatGPT takes your side in an argument

Ask ChatGPT who's right about your last argument. It might agree with you. Now let the other person tell their version. If the details change, don't be surprised if the answer does too.

The AI may be too agreeable, but there's another problem: it only knows the version you gave it.

It only has your version

When you describe an argument to an AI by yourself, it gets your account of what happened. Even if you explain the other person's position, you're still choosing which details to include. The AI can question your version, but it can't hear the other person's account unless someone provides it.

Research has found that AI assistants can be overly agreeable, especially when people ask for personal advice. Combine that tendency with an incomplete account of an argument, and the answer may reinforce what the person already believes.

When AI answers become part of the argument

Couples counselors are now writing about clients who show up with AI transcripts as evidence. NPR ran a first-person experiment using ChatGPT as a couple's counselor; the takeaway was that it was helpful for reflection and useless as a referee. Other reporting found couples paying for AI to "referee" their relationship, one at a time, from separate phones.

If only one person tells the story, the AI only has that person's account.

Let the other person answer the AI

SharedChat puts both people in one conversation with the same AI. Each can explain what happened, question the answer, and see what the AI tells the other person. That gives it more context, though it can still get the ruling wrong.

That's what SharedChat is. But being in the same room isn't quite enough on its own, so the rooms come with rules:

• In Settle It, the Judge won't rule until it's heard the other side. A one-person argument gets a question, not a verdict.
• In Relationships and Co-Parenting, it's told in writing never to pick a winner, to explain what each person is asking for, and to keep custody questions with a lawyer.
• The rooms have been tested with scenarios involving abuse and safety concerns. They're instructed to avoid suggesting compromise in abusive situations and to direct people toward appropriate help.