Summary: AI-powered writing tools do more than speed up typing—they can subtly shift how people think. A large-scale study from Cornell University shows that biased autocomplete suggestions from AI writing assistants can nudge users’ opinions on major societal issues, including the death penalty and fracking.
Researchers ran two preregistered experiments with more than 2,500 participants to test whether autocomplete-style suggestions influence attitudes. Participants writing about contentious topics were offered biased AI suggestions while drafting short essays. After the task, many participants’ survey responses moved toward the AI’s programmed stance. Crucially, most users did not notice the influence, and standard countermeasures—warnings issued before use or debriefings after—did not prevent the shift.
Key Facts
- Covert persuasion: Participants who accepted or incorporated biased AI autocomplete suggestions produced essays that reflected the AI’s slant and then reported opinions closer to that slant in posttask surveys.
- Warnings ineffective: Unlike many cases of misinformation where pre-emptive warnings or post-exposure corrections reduce belief, telling participants the assistant was biased did not stop their attitudes from shifting.
- Broad reach: The effect appeared across multiple political topics (for example, GMOs, voting rights for felons, fracking, and the death penalty) and affected people across the political spectrum.
- Mechanism of change: The shift likely arises because users end up writing biased content themselves; decades of psychology research show that producing an argument can change the author’s own beliefs.
Source: Cornell University
Overview
Autocomplete and AI-driven text suggestions are now embedded in many writing environments, from email clients to document editors. This research examines whether those suggestions merely save time or whether they can reshape users’ viewpoints. The study, led by doctoral candidate Sterling Williams-Ceci with senior author Mor Naaman and co-authors including Maurice Jakesch, finds that biased suggestions can produce measurable attitude change.
In the first experiment, participants wrote essays for or against standardized testing. Some saw autocomplete suggestions biased toward pro-testing arguments, others saw neutral or no suggestions, and a third group saw static lists of AI-generated pro-testing arguments displayed before writing. The greatest attitude shifts followed the live autocomplete suggestions, not the static lists, indicating that the interactive nature of suggestions amplifies influence.
The second experiment expanded the scope to several politically consequential topics: the death penalty, fracking, genetically modified organisms (GMOs), and voting rights for people with felony convictions. For each topic the researchers engineered autocomplete suggestions that leaned in a predetermined direction—liberal-leaning on some topics and conservative-leaning on others. Across all issues, participants’ posttask opinions moved toward the direction of the suggestions.
A particularly important finding is that most participants were unaware their views had changed and could not be protected by simple warnings or debriefs. “We expected pre-bunking or debriefing to provide some immunity,” said Sterling Williams-Ceci, “but neither approach reduced the extent of attitude shifts produced by biased writing assistants.”
Why this matters
As large language models become common tools for drafting emails, essays, and reports, the potential for subtle, widespread shifts in public opinion grows. If many people rely on the same writing assistants that carry a particular bias—whether accidental, emergent, or intentional—collective viewpoints could drift in the assistant’s direction without clear awareness.
The mechanism is subtle but powerful: when people adopt suggested phrasing and present it as their own writing, they may internalize the position. This process is supported by long-standing psychological findings that articulating an argument can change the arguer’s beliefs. The study highlights a unique pathway for persuasion that differs from simply reading biased content.
Authors and funding
Lead author: Sterling Williams-Ceci. Co-authors include Maurice Jakesch, Advait Bhat, Kowe Kadoma, Lior Zalmanson, and Mor Naaman. Funding was provided by the National Science Foundation and the German National Academic Foundation.
Frequently Asked Questions
A: Not necessarily intentionally, but it can be an unintended side effect. Many AI models reflect biases in their training data. When you accept a suggested sentence, your brain often treats it as self-generated, so the suggestion can become integrated into your own viewpoint.
A: Warnings that help against misinformation typically operate when people passively consume content. With writing assistants, users actively co-author text. That creative involvement appears to bypass critical filters, making simple warnings less effective.
A: Widespread use of similarly biased writing tools could contribute to a homogenization of expressed views, creating broad, incremental changes in public opinion that occur without clear awareness or debate.
Editorial Notes
- This article was edited by a Neuroscience News editor.
- The journal paper was reviewed in full.
- Additional contextual information was added by staff.
About this research
Author: Becka Bowyer
Source: Cornell University
Contact: Becka Bowyer – Cornell University
Image credit: Neuroscience News
Original research: “Biased AI Writing Assistants Shift Users’ Attitudes on Societal Issues” by Sterling Williams-Ceci, Maurice Jakesch, Advait Bhat, Kowe Kadoma, Lior Zalmanson, and Mor Naaman. Published in Science Advances. DOI: 10.1126/sciadv.adw5578
Abstract
Biased AI Writing Assistants Shift Users’ Attitudes on Societal Issues
AI writing assistants that offer autocomplete suggestions are increasingly common. Two large preregistered experiments (N = 2,582) exposed participants writing about important societal issues to assistants that offered biased suggestions. Participants’ posttask attitudes shifted toward the assistant’s position, and most were unaware of both the suggestions’ bias and their influence. The effect was stronger than exposure to similar static text, and warning participants about bias before or after exposure did not eliminate the attitude shift.