Every day, hundreds of millions of people type a sentence and watch a grey ribbon of suggested text appear before they finish.
Most tap the suggestion without a second thought. The words sound right. They sound, in fact, like what the writer was already thinking. But a rigorous new experiment found something unsettling hiding in that grey ribbon: the suggestion quietly moved the writer’s actual opinion, not just the sentence on screen.
How a word suggestion reaches into your thinking before you even hit tab
The mechanism is subtler than it looks. When you type a few words and the tool offers a completion, your brain does something it evolved to do with all social input: it runs a rapid consistency check. The suggested phrase is already formatted in your own voice, placed inside your own sentence, pointing toward a conclusion.
These systems rely on large language models that predict likely continuations of a sentence. Because the completion arrives inside your own stream of thought rather than as a message from an outside source, the framing feels internal. That is exactly why the influence bypasses the filters people normally use when they sense someone is trying to persuade them.
As senior author Mor Naaman, professor of information science, noted: “autocomplete is everywhere now.” It was less prevalent and limited to short completions three years ago, but today services such as Gmail will suggest writing entire emails on a user’s behalf. That scale matters enormously.
What the keyboard suggests when you write about something you care about
Autocomplete features in major email platforms complete sentences and suggest professional responses, making it faster to draft messages. That efficiency is real, and most users are glad for it.
But efficiency means the user is reading and accepting suggestions at speed, with less deliberate scrutiny than they would bring to a statement composed from scratch. When the underlying model carries a lean, even a subtle one baked in during training, every suggested completion nudges the writer a fraction of a degree. Across millions of daily interactions, those fractions accumulate into something measurable.
So the finding is not that autocomplete changes what people type. It is that autocomplete changes what people think, even after they put the keyboard down.
The experiment, who was in it, and what the numbers showed
In two large-scale preregistered experiments involving 2,582 participants, researchers exposed people writing about important societal issues to an AI writing assistant that provided biased autocomplete suggestions. The tool was deliberately tuned to favor one side of each question.
When using the assistant, the attitudes participants expressed in a post-task survey converged toward the AI’s position.
A majority were unaware of the suggestions’ bias and their influence. The shift happened below the level of conscious notice, which means the usual defense, knowing you are being nudged, was not enough to stop it. The findings were published in the journal Science Advances, suggesting that the subtle influence of these everyday programs often bypasses conscious awareness entirely.
Warning users about the bias did not protect them, and that changes the picture entirely
The most striking complication in the data is not the attitude shift itself. It is what happened when researchers tried to prevent it. Lead author Sterling Williams-Ceci noted that “previous misinformation research has shown that warning people before they’re exposed to misinformation, or debriefing them afterward, can provide ‘immunity’ against believing it”, but that “neither of those interventions actually reduced the extent to which people’s attitudes shifted toward the AI’s bias in this context.”
Explicit warnings, placed both before and after the writing task, failed to neutralize the influence. That finding matters for anyone who assumes a disclosure label solves the problem. A prominent warning is not the same as immunity.
The shift appears to operate through the act of engaging with the suggestion itself: the attitude shift occurred even among the roughly thirty percent of participants who did not actually accept any suggested words into their essays. Few people realize that AI writing assistants can push them to think a certain way, what senior author Mor Naaman of Cornell University described as a covert and potent form of influence. And there is a broader context of digital habits here: when roughly half of US adults already say their phone use feels excessive, the idea that a communication tool is also quietly shaping their views adds a new dimension to that unease.
What this means for the ordinary writer who just wants to send a faster email
None of this means autocomplete is malicious by design. Such influence may not matter much when AI completes a routine message. But when people use these tools to discuss weightier societal matters, such as whether the death penalty should be abolished, or whether fracking should be allowed, participants who used the biased AI had their views gravitate toward the AI’s positions. The model’s bias, in those circumstances, can have significant societal impact.
The practical implication connects to how we think about communication skills more broadly. If the words we use to express a view are increasingly selected from a model’s probability distribution, the exercise of forming an original opinion gets less practice. Language and thought run in both directions, and a tool that shapes one is shaping the other.
Researchers also found that the influence of the AI writing assistant was stronger than the influence of similar suggestions presented as static text, a detail that points toward interface design as a lever worth pulling. The answers to whether different formats can reduce the assimilation effect are not in yet. But the mechanism is documented in a peer-reviewed journal with a sample large enough to be convincing: the grey ribbon of text is not just predicting your next word. In some circumstances, it is quietly predicting your next opinion too.
