Do AI autocomplete suggestions shift attitudes more than static text?

Two preregistered experiments (Williams-Ceci et al., Science Advances) with N=1485 and N=1097 US participants found that writing an essay on a debatable societal issue with biased AI autocomplete suggestions shifted participants' later attitudes significantly toward the AI's position. In experiment 1, seeing biased suggestions moved attitudes 0.44 scale points (of 5) toward the AI's stance versus a no-suggestion control (95% CI 0.21–0.67, Cohen's d = 0.24); experiment 2, using pre-task attitudes gathered roughly 34 days earlier, replicated the shift at 0.41 scale points (95% CI 0.05–0.77, d = 0.19). The effect held "even [for] participants who did not actually accept any of the AI suggestions in their writing."

The authors attribute the effect to a covert, behavioral pathway rather than explicit persuasion: because "people's behaviors can shift their attitudes," autocomplete suggestions that make one viewpoint easier to elaborate than another can move attitudes even when users never adopt the suggested text verbatim. Experiment 1 included a Static Text condition showing the same AI-generated arguments as a plain list instead of in-line suggestions, isolating argument content from the act of co-writing; the paper reports the autocomplete condition produced a "significantly stronger effect on attitudes" than this static presentation of identical arguments, and that participants were "generally unaware" the suggestions had swayed them.

This cuts against the warning-based mitigations found effective elsewhere: Can a simple warning reduce how much LLMs persuade people? and Can warnings stop people from being swayed by sycophantic AI? both found warnings can blunt belief change or at least leave trust untouched. Here, the authors state plainly that bias-warning interventions, "despite previous work suggesting that these techniques can counteract the influence of persuasive messages," did not eliminate or even reduce the autocomplete effect — pointing to covert, in-writing suggestion as a distinct and more resistant channel than the explicit persuasive messages studied in Where does AI's persuasive power actually come from?. It also extends the unawareness theme in Do readers trust unlabeled AI-written messages as much as human ones?: there, ignorance of AI involvement left a reader's judgment of someone else's message untouched; here, ignorance of AI involvement is coupled with a measurable shift in the writer's own later attitude.

The excerpt does not give the warning intervention's sample size, wording, or effect-size statistics, so how comparable it was to the successful warnings in the belief-change study above cannot be checked from this excerpt alone. It also does not establish how long the attitude shift persists beyond the post-task survey, or whether it generalizes past the five debatable issues and US-recruited samples tested. Given that millions of people write daily with a small number of LLM-based assistants, even a modest per-essay shift of this kind implies a channel of societal-attitude influence that disclosure-style warnings — the main mitigation proposed for overt AI persuasion — may not reach.