Can a summary change your mind while every single fact in it stays true?

This explores whether a summary can stay factually accurate and still shift what readers believe. The corpus suggests the question rests on a false trade-off. Accuracy and persuasiveness turn out to be mostly separate dials. A summary can be entirely true and still lean, and the levers that do the persuading usually have little to do with whether the facts are correct. The most direct evidence is a preregistered experiment with nearly 2,000 people. GPT-4o's default historical summaries moved readers' opinions leftward compared with Wikipedia, even though nobody asked the model to persuade anyone Do AI historical summaries shift reader opinions without explicit prompting?. The shift came from framing: which true facts got chosen, in what order, and with what emphasis. It didn't come from false claims.

So where does persuasion live, if not in the facts? Grammar is one place. Claims that are presupposed (slipped in as background, as in "*even* the committee *realized* that…") persuade more than the same claims stated outright, especially when the information is new to the reader Why are presuppositions more persuasive than direct assertions?. The claim can be perfectly accurate. It simply arrives already treated as settled, so it skips the reader's scrutiny. Register is another place. An audit of five models found they persuade in almost every conversation, using logical appeals and numbers. That style reads as neutral and gives LLM output an objective feel it hasn't necessarily earned Do LLMs persuade users more often than humans do?. Tone matters too. GPT-4 tends to turn negative prompts into neutral-to-positive answers, so the same question gets a different factual emphasis depending on the asker's mood Does emotional tone in prompts change what information LLMs provide?.

The reader's side of the exchange makes this harder to see. Readers given no evidence signals couldn't tell fluent hallucinations from the truth at all. A display showing which claims had been verified restored their ability to tell them apart Can readers tell truth from fabrication without evidence signals?. Readers also judge by surface signals. Responses with more citations win trust even when the citations are irrelevant Do users trust citations more when there are simply more of them?, and polished prose gets mistaken for higher quality Does polished writing actually signal better quality work?. Put together, an accurate summary doesn't persuade because it's accurate. It persuades because it looks credible, and inaccurate text can borrow that same look.

There's a limit worth knowing about. In the GPT-4o study, explicitly conservative summaries persuaded only readers who were already conservative. That matches debate research showing readers' prior ideology predicts persuasion outcomes better than anything about the language Does what readers believe matter more than what debaters say?. Framing works best on audiences who are open to it or don't notice it. It struggles to move people who are already committed the other way.

The takeaway you may not have expected: "is it accurate?" is the wrong safety check for AI summaries. A summary is always optimized toward something. One system trained on ranking signals produced dense, attribute-heavy summaries instead of fluent prose Can reinforcement learning align summarization with ranking goals?. Whatever the target (engagement, helpfulness, a default worldview), it shapes which true things you see. The better question is what the summary was selecting for. For now, the corpus's most practical answer is to show readers the evidence behind each claim, not just a fluent conclusion.