Biased AI writing assistants shift users' attitudes on societal issues

Source: Williams-Ceci, Jakesch, Bhat, Kadoma, Zalmanson, Naaman (Cornell Tech), Science Advances · 2026-03-11

Biased AI writing assistants shift people’s attitudes about societal issues; common interventions do not prevent this influence.

Artificial intelligence (AI) technologies based on large language models (LLMs) are increasingly involved in human communication, generating text and making suggestions in a multitude of contexts. First appearing as “smart replies” (1) for email and messaging, AI-mediated communication [“AIMC”; (2)] features are now ubiquitous in text editors like Microsoft Word, email clients, messaging applications, and elsewhere (3–7). The integration of generative AI into people’s communication raises questions about the technology’s impact on language use (2, 8–11), perceptions of others (6, 12, 13), and interpersonal dynamics (3, 13–15). Together, these studies raise the broader question of how much AI can influence people’s cognition (16). As AI tools have been shown to reproduce and reinforce human biases (16–20), their influence on people’s cognition has become a serious concern (16).

Previous research has examined AI’s persuasive capabilities in explicit contexts, where people are aware of an overt influence attempt. Studies have shown the persuasive impact of AI in various types of interaction with humans, such as when AI generates persuasive messages (21–26), converses with people about a relevant belief (27–30), and provides decision recommendations (31–33). However, a critical gap remains in our understanding of the more subtle and indirect ways that AI can influence human cognition.

The present study investigates a potentially more powerful and pervasive form of AI influence: the impact of AI writing assistants that provide autocomplete suggestions (“AI suggestions” for short) on users’ attitudes [“attitudes” refers to the evaluations people form toward “objects” such as societal issues and policy decisions; see (34)]. Unlike the explicit persuasion attempts studied previously, AI suggestions may shape users’ cognition in ways that bypass conscious awareness, representing a more indirect and covert way in which AI can influence humans. In applications that make use of AI suggestions, writing assistants powered by LLMs predict in-line continuations to the author’s writing and suggest this predicted text to the author (35–38). AI writing assistants that provide autocomplete suggestions pose a distinct pathway for influencing users’ thoughts: While it is often known that people’s attitudes influence their behavior, it is also known that people’s behaviors can shift their attitudes (39–41). When biased AI autocomplete suggestions inspire people to think of certain viewpoints—and make it easier for people to elaborate on those viewpoints than on others—cowriting with a biased AI writing assistant may powerfully shift people’s attitudes on various issues. If AI suggestions have even a minor influence at the individual level, their potential for aggregate influence is worrisome as millions of people are using the same models in their daily communications (18, 42, 43).

In this work, we provide robust evidence of the impact of AI writing assistants’ autocomplete suggestions on attitudes and show that well-known influence mitigation techniques may not reduce it. Our work shows that AI suggestions can influence people’s attitudes across a range of politically charged societal issues, with AI biases that vary in partisan association and stance. We compare the covert influence of people writing with the AI writing assistant’s autocomplete suggestions to the influence when overtly presenting similar arguments and show that the autocomplete suggestions have a significantly stronger effect on attitudes. We also find that users are generally unaware of the suggestions’ influence. Furthermore, we show that the effect is not eliminated or even reduced by interventions that warn people about the AI suggestions’ bias, despite previous work suggesting that these techniques can counteract the influence of persuasive messages (51). An initial study suggested that AI writing assistants may shift users’ attitudes toward the issues they write about (8), but the study’s design was too limited to support generalizable conclusions about this phenomenon: It only relied on a single experiment using one issue that was not politically charged; it did not rule out other potential explanations; and it did not test whether the influence can be undone by making people more aware of the bias [which has been found in the case of misinformation (51)].

In our experiments, participants wrote a short essay about one of five debatable issues with (AI Treatment) or without (Control) receiving AI suggestions that were biased toward one viewpoint (see Fig. 1 for an example of the AI Treatment’s suggestions). In experiment 1 (N = 1485), all participants wrote about their attitudes toward whether standardized tests should be used in education. In the AI Treatment, an AI writing assistant provided autocomplete suggestions that made arguments biased in favor of standardized testing, which we achieved by carefully engineering an AI prompt that instructed the AI model we used to generate pro-test arguments that still logically completed the users’ sentences. In addition to the Control condition (N = 499) and AI Treatment (N = 487), experiment 1 included a third comparison condition, the Static Text condition (N = 499), in which we showed participants a static list of arguments, based on those that were generated by the LLM when providing autocomplete suggestions in a pretest (see fig. S1). This condition explored whether the content of the suggestions alone could explain any observed shift in attitude, or whether the unique ability of AI writing assistants to change people’s writing in real time contributes to AI’s influence, as predicted by the literature on behavior influencing attitudes (39–41). After the writing task, we asked participants about their attitude toward the issue they wrote on (using a Likert scale to measure their attitudes numerically). We then asked participants in the AI Treatment to rate how balanced and influential they perceived the AI suggestions to be.

While experiment 1 only compared participants’ posttask attitudes between the groups, experiment 2 used a repeated-measures design in which we also measured subjects’ pretask attitudes several weeks before participation. We first recruited participants to fill out questionnaires asking for their attitudes toward several societal issues, four of which were the same as those we used in the actual experiment. We used the same Likert scales to measure their attitudes in both stages. We launched the actual experiment 20 days after the last participant had completed the pretask survey and restricted eligibility only to the participants who had completed the pretask survey. The connection between the two tasks was hidden from participants, who took the actual experiment, on average, 34.44 days following the pretask survey; see Materials and Methods for more details. This design enabled us to do a within-subject comparison of the extent of the shift from the presurvey attitude to the postexperiment attitude about the issue they had written about in the experiment. The percentages of participants in experiment 2 whose pretask attitudes agreed with the AI suggestions’ bias, disagreed with it, or were neutral for their assigned issue in the ultimate task are shown in table S1.

Across experiments and topics, participants who saw the biased AI suggestions reported attitudes that were significantly closer to the AI’s position than participants who wrote without any suggestions. In experiment 1 (N = 1485), participants’ attitudes differed based on their condition [analysis of variance (ANOVA): F(2, 1474) = 10.244, P < 0.001, ηp2 = 0.01]. Specifically, a post hoc comparison showed that seeing biased AI suggestions (N = 487) increased the alignment of participants’ attitudes with the AI’s position by 0.44 scale point on a scale from 1 to 5 [95% confidence interval (CI) for AI Treatment–Control difference (0.21, 0.67), t(1474) = 4.523, P < 0.001, Cohen’s d = 0.24] compared to the attitudes of Control group participants (N = 499). This effect includes participants who did not actually accept any of the AI suggestions in their writing. In experiment 2 (N = 1097), a similar effect occurred, with participants in the normal AI Treatment (N = 219) expressing attitudes in the posttask survey that were 0.41 scale point more aligned with the AI’s biased position than Control participants’ attitudes (N = 312) for the issue they wrote about [ANOVA: F(4, 1089) = 4.07, P < 0.001, ηp2 = 0.01; 95% CI for AI Treatment–Control difference (0.05, 0.77), P = 0.015, t(1089) = 3.14, Cohen’s d = 0.19]. These results are summarized in Fig. 3 (the “Control” and “AI Treatment” conditions) alongside results from other conditions discussed below.