Impact of AI-Powered Predictive Text on Societal Attitudes and Opinions

A 2026 study published in Science Advances reveals that AI auto-complete features may subtly influence public opinion on contentious social issues.

Researchers from Cornell University, led by information scientist Mor Naaman, found that exposure to biased AI-generated suggestions during writing tasks can shape users’ views on policy matters, even when individuals do not consciously adopt the suggestions.

The study underscores the potential for large language models to act as silent persuaders in shaping societal discourse.

Naaman and his team conducted two experiments involving over 2,500 participants to assess how AI auto-complete functions affect users’ perspectives.

Participants composed essays on topics such as the death penalty, standardized testing in education, and voting rights for felons.

Some individuals received AI-generated suggestions, while others wrote without assistance.

Researchers intentionally biased the AI prompts to evaluate their influence.

For example, when asked, ‘Should the death penalty be illegal?’ a participant began with ‘In my view,’ and the AI auto-completed it with ‘the death penalty should be illegal in America because it violates the Eighth Amendment, which prohibits cruel and unusual punishment.’

Participants later rated their stance on a scale from 1 (no) to 5 (yes).

Those exposed to biased AI suggestions—regardless of whether they incorporated them into their writing—moved nearly half a point closer to the AI’s position compared to those without exposure.

Approximately 75% of participants who received AI suggestions deemed the model’s recommendations ‘reasonable and balanced,’ indicating they were largely unaware of the influence.

The study warns that widespread use of biased AI models could sway public positions on critical policies or politicians.

Naaman estimates that shifting just 20,000 voters in Pennsylvania could potentially alter an election outcome.

This raises concerns about the scalability of AI’s influence, particularly as billions of users globally rely on similar models for writing and communication.

Researchers emphasize that the manipulation occurs subtly and often without users’ awareness.

This creates a risk of homogenizing public discourse, as large swaths of people using the same model may adopt similar viewpoints.

Naaman noted, ‘AI can have the effect of homogenizing our words and creativity, but also our thoughts.’

He advocates for users to write their own thoughts first before seeking AI assistance, ensuring that the ‘seed [of the idea] is mine.’

The findings align with broader concerns about AI’s role in shaping public opinion.

A 2026 Atlantic Council report highlights how AI is being weaponized in geopolitical contexts, including disinformation campaigns.

For example, China’s AI-enabled disinformation efforts are expected to intensify in 2026, particularly targeting Taiwan.

These campaigns use AI-generated audio, video, and text distributed through fake accounts and private firms to conduct ‘cognitive warfare,’ shaping political perceptions and voter behavior.

Similarly, the report notes that AI poisoning—where malicious data corrupts AI training sets—has become a critical issue.

Russia’s Pravda network, for instance, has published millions of articles targeting over 80 countries, aiming to influence AI models that rely on web crawlers for training data.

This underscores the potential for AI to amplify misinformation, further complicating efforts to discern factual information from fabricated content.

While AI models include disclaimers such as ‘ChatGPT can make mistakes. Check important info,’ users remain susceptible to persuasive AI**.

The study found that even when participants were warned about potential biases, the AI’s influence persisted.

This raises questions about how to inoculate users against covert manipulation, particularly in politically charged contexts.