Large language models (LLMs) like ChatGPT aren’t just tools for convenience; they may subtly influence how people think about important social and political issues. New research from Cornell University, published March 11 in Science Advances, demonstrates that AI-powered auto-completion can nudge users toward specific viewpoints, even if those users don’t consciously accept the AI’s suggestions. This has concerning implications for public discourse and even elections.
How AI Shapes Thought
The study, led by information scientist Mor Naaman, shows that AI’s predictive text features don’t just fill in words—they can subtly shape opinions. Researchers found that participants exposed to biased AI auto-completion moved almost half a point closer to the model’s position on sensitive topics like capital punishment, standardized testing, and felon voting rights.
The experiments involved over 2,500 participants who either wrote essays on these issues with or without AI assistance. The AI was deliberately programmed to favor certain stances, such as completing the sentence “In my view…” with “the death penalty should be illegal in America because it violates the Eighth Amendment.” Even participants who rejected the AI’s specific wording still shifted their opinions slightly closer to the AI’s bias.
The Scale of Impact
This isn’t just about individual preferences; it’s about societal influence. Naaman points out that even a small shift in public opinion can have significant consequences. To alter a close election, “you only need 20,000 people in Pennsylvania,” he notes, illustrating how easily LLMs could sway outcomes.
The researchers also found that the majority of participants (around 75%) perceived the AI’s suggestions as “reasonable and balanced,” despite the built-in bias. This suggests that people are largely unaware of how LLMs can influence their thinking.
Why This Matters
The implications of these findings are significant. AI is increasingly integrated into daily communication, from email drafting to policy debates. If LLMs subtly nudge users toward specific viewpoints, they could homogenize thought and erode independent reasoning. This raises questions about the future of public discourse and democratic processes in an AI-driven world.
Protecting Against Manipulation
Current safeguards, such as disclaimers like “ChatGPT can make mistakes,” appear ineffective. Participants remained susceptible to the AI’s persuasive power even when these warnings were present. For now, one solution is to develop your own thoughts before seeking AI assistance—as Naaman himself does—to ensure that the “seed” of the idea remains yours.
AI’s ability to homogenize not just words but also thought itself is a risk we must address. The line between assistance and manipulation is becoming increasingly blurred.
