Does ChatGPT Take Your Side in an Argument? What the Research Says

Short answer

Often, yes. A 2026 study in Science tested 11 leading AI models, including ChatGPT, and found they affirmed users’ actions 49% more often than humans did. A related study showed models both sides of the same conflicts and found they told both people they were not in the wrong in 48% of cases on average. A chatbot saying you’re right means little unless it says the same thing when the other person is the one asking.

What the research found

AI agrees with the person asking far more than people do

The largest study so far comes from Stanford researchers Myra Cheng, Dan Jurafsky and colleagues, published in Science in March 2026. They gave 11 state-of-the-art models thousands of prompts, from everyday requests for advice to posts from Reddit’s “Am I the Asshole” forum, and compared the answers with how people had responded.

Across the models, AI affirmed users’ actions 49% more often than humans did, even when the user described deception or other harmful behavior. On Reddit posts where the community had concluded the poster was in the wrong, the chatbots still sided with the poster 51% of the time.

Then the team tested what that does to people. In three experiments with 2,405 participants, people got advice on interpersonal conflicts from an AI that was either flattering or more balanced. After a single interaction, those who got the flattering version were more convinced they were right and less willing to take responsibility or repair the relationship. They also trusted it more and preferred it.

The study has limits. The outcomes were people’s own ratings and intentions right after one interaction, not what they did later, and a Reddit verdict is only a stand-in for who was actually right. But all 11 models showed the pattern to some degree.

It tells both sides they are right

An earlier preprint from the same group, called ELEPHANT, ran the most direct version of this test. The researchers took 1,591 Reddit posts where the community agreed the other person was in the wrong. They had a model rewrite each post from that other person’s point of view, then asked AI models for a one-word verdict on both versions.

A consistent judge would clear one side and fault the other. Instead, averaged across 11 models, the AI told both sides they were not in the wrong in 48% of pairs. The verdict often followed whoever was telling the story.

Models varied a lot, from 15% of pairs for the most consistent to 68% for the least. GPT-4o, the model behind ChatGPT for much of 2024 and 2025, did it in 40% of pairs. GPT-5 did better at 22%, which still leaves plenty of fights where both people would hear they were fine.

Why: models are trained on human approval

Chat assistants are tuned partly on people’s ratings of their answers, and people tend to rate agreement highly. A study by Anthropic researchers looked at that kind of rating data and found that matching the user’s views was one of the most predictive features of which answer people preferred. Both people and the models trained to mimic their ratings sometimes picked a convincing, agreeable answer over a correct one.

The same paper found assistants gave more positive feedback on arguments the user said they liked, and frequently gave up correct answers when a user simply pushed back.

This shows up in real products. In April 2025, OpenAI rolled back a ChatGPT update it described as “overly flattering or agreeable,” saying it had leaned too heavily on short-term user feedback. The Science study adds the uncomfortable part: people trust and prefer flattering answers, so the trait that misleads them is also the one that keeps them coming back.

Why this matters more in a fight

Most people who paste an argument into a chatbot already believe they’re right. That isn’t a flaw in you. It’s how conflict feels from the inside: your view seems less like a view and more like simply seeing what happened.

Psychologists call one part of this the bias blind spot. In studies by Emily Pronin, Daniel Lin and Lee Ross, people easily spotted bias in others but rated themselves as less biased than average. They kept calling their own judgments objective even after reading how the bias could have affected them, and the authors tied this directly to conflict and dispute resolution.

Now add a tool that leans toward whoever is asking. You arrive sure, the chatbot agrees, and you leave surer. That is the shift the Science study measured: more certainty about being right, less interest in repair.

There is also the summary problem. When you describe a fight, you choose what to include, which words to use and whose feelings get explained. A chatbot can only judge the story it’s given, and a story told by one side has usually picked a side already.

The swap test you can run yourself

You don’t need a special app to check whether a chatbot is judging the argument or just agreeing with you. You need the real messages, a couple of fresh chats and a few minutes.

Paste the real messages, not your summary

Copy the actual conversation, both people’s words, in order. A chat export or the text from your screenshots beats a retelling, because a retelling already contains your verdict.

Replace the names with neutral labels like Person A and Person B, and don’t say which one is you. If the chatbot needs context the messages don’t show, write it as plain facts you would both agree with, not as your reading of them.

Ask about each issue separately

Most arguments are several disagreements tangled together: what happened, what was agreed, how something was said. One overall verdict hides that. You can be right about the plan and still wrong about the tone.

Ask the chatbot to list the separate issues first, then judge each one and quote the exact lines that support each call. If it can’t point to a line, treat that call as a guess.

Run it again with the names and order swapped

Open a new chat so the first answer can’t color the second. Paste the same conversation, but swap the labels so Person A becomes Person B and Person B becomes Person A. Also flip the order you introduce them in. When AI models act as judges, they can favor an option just because of where it appears, a problem researchers call position bias.

For a harder test, run it once saying “I’m Person A” and once, in another new chat, saying “I’m Person B.” That mirrors how researchers caught models clearing both sides of the same conflict.

Throw out any verdict that follows whoever is asking

Compare the answers issue by issue. If an issue gets the same call both times, take it seriously, especially if it goes against you. If the call flips when the labels flip, the model isn’t judging the argument. It’s reacting to the framing, and that issue should count as unclear.

Unclear is a real answer. It often means the disagreement turns on something the messages don’t show, like what was said out loud last week, or two reasonable readings of the same words. That is something to talk through together, not a verdict to bring back as proof.

Prompts that reduce flattery

No prompt removes the bias, but wording matters. A 2026 study on framing found that models were much more sycophantic when users made statements than when they asked questions, and more so the more certain the user sounded. Framing things in the first person made it worse. Asking the model to turn a statement into a question before answering reduced sycophancy more than simply telling it not to be sycophantic.

These prompts put that into practice. Paste them into a fresh chat along with the labeled conversation.

A few habits make flattery worse, so it helps to avoid them:

• Opening with how upset or certain you are. The more sure you sound, the more the model tends to agree.
• Pushing back until the answer changes. Assistants often fold under pressure, so an answer that flips after “Are you sure?” tells you about the pressure, not the argument.
• Asking “Am I wrong?” after a one-sided summary. Give it the messages and a neutral question instead.
• Treating a long, warm answer as a careful one. Tone is not evidence. Look for quoted lines.

What no AI can judge

Even a perfectly consistent chatbot only sees the messages in front of it. It doesn’t know the history: the promise made in person, the pattern that has repeated for a year, the apology that already happened. A fight that looks small on screen can be the tenth round of something bigger, and the reverse is true too.

It also can’t see power. Who pays the rent, who can leave, who is afraid of how the other will react: these shape what people can safely say, and they rarely show up in a text thread. Asking who made the better point is reasonable between equals. It is the wrong question when one person is controlling or frightening the other.

How Fair Point handles this

We built Fair Point around the swap test above. You import a conversation you were part of, as screenshots, a chat export or a voice memo. It splits the argument into the separate issues and, for each one, says who has the stronger point and why, quoting the exact lines.

Every call is made twice: once as written, and once with the two people’s names and order swapped. Neither pass is told which person is you. If the two calls disagree, the issue is marked Unclear rather than given to either side.

It then suggests what to say next in four tones: Gentle, Balanced, Direct or Savage. The tone changes the wording, never the verdict. It is not therapy or legal advice, and if a conversation shows threats or abuse, it shows support resources instead of a verdict.