AI summaries that sound neutral can quietly nudge your politics — would you even notice it happening?

This explores whether readers notice political tilt in summaries that present themselves as neutral, especially AI-written ones. The corpus doesn't contain a study that directly measures whether readers *recognize* the framing. It has strong evidence that the framing works anyway, and several findings that explain why it would be hard to spot. The clearest case is a preregistered experiment with nearly 2,000 people: default GPT-4o summaries of historical topics moved readers' opinions leftward compared with Wikipedia, even though nobody asked the model to take a side Do AI historical summaries shift reader opinions without explicit prompting?. The asymmetry is the surprising part. Explicitly conservative summaries only persuaded people who were already conservative, so the default version was the one doing the broad persuading. Put simply, the framing nobody asked for moved people more widely than the framing someone did ask for.

The most likely reason readers miss the framing is that they lack the cues they would need to check it. In one study, people reading AI text without any source information couldn't tell true claims from fabricated ones at all. Their judgment returned only when an interface showed which claims were backed by evidence Can readers tell truth from fabrication without evidence signals?. Even ML-literate readers can't reliably tell AI abstracts from human ones, and they tend to assume a person wrote whatever they're reading Can readers tell LLM abstracts from human ones?. A broader argument covers the cultural side: we read advertising with a built-in discount because we know it's selling something, but AI-generated text hasn't been around long enough to come with that kind of default skepticism How do we learn to read AI-generated text critically?. A neutral-sounding summary gets read as neutral.

The tilt also tends to sit below the level of explicit claims. AI writing assistance shifts how readers perceive a writer on all 29 traits measured, including how extreme and how confident the writer seems. The shifts are consistent in direction, not random Does AI writing assistance change how readers perceive the writer?. GPT-4 also nudges negative questions toward neutral or positive answers, so the same question can get different information depending on its emotional tone Does emotional tone in prompts change what information LLMs provide?. Framing of this kind lives in tone, emphasis, and what gets left out, and readers aren't primed to audit any of those.

Two findings complicate the idea that there is a single "neutral" version to compare against. Research on how people interpret sentences shows that readers in different social positions validly read the same sentence differently, so disagreement isn't just noise Why do readers interpret the same sentence so differently?. In debate data, what the audience already believes predicts who wins better than anything about the language itself Does what readers believe matter more than what debaters say?. Together these suggest that whether a reader notices framing depends partly on where they already stand. A summary that leans toward your own view tends to read as plain fact, and a summary that leans away from it reads as biased.

The takeaway you may not have expected: it isn't only humans who miss this. LLMs used as judges are easily swayed by fake credentials and polished formatting Can LLM judges be fooled by fake credentials and formatting?, so handing the neutrality check to another AI doesn't solve the problem. The most promising fix in the corpus is to change the interface rather than train readers to be more vigilant: show readers where each claim comes from, so they have something concrete to check.