Study Reveals How AI Can Subtly Influence Human Decisions

Researchers from the University of Hong Kong and Tsinghua University in China investigated how easily AI can influence human decision-making.

In an experiment involving 233 participants, subtle AI bias made participants five to eight times more likely to choose a less beneficial option. Despite this influence, participants continued to regard AI as a useful assistant. The findings were published in the journal PNAS.

People are increasingly turning to AI for assistance and advice. Their questions often involve financial decisions and personal relationships. Users generally expect AI assistants to help them understand situations or identify suitable options.

However, the researchers wanted to examine what happens when AI recommendations subtly steer people toward decisions that conflict with their own interests.

Subtle Bias in Advice Can Change Decisions

The experiment involved 233 participants. Researchers examined how AI recommendations influenced choices in financial and emotional scenarios.

In some cases, the AI assistant subtly favored a less beneficial decision. On the surface, however, it continued to provide seemingly ordinary advice.

The results showed that even a small bias could significantly change participants’ behavior. They were more likely to choose options that worked against their interests than they would have been without similar influence.

Depending on the scenario, the likelihood of making a suboptimal decision increased by as much as 38%.

The AI did not need to pressure users openly, aggressively promote a product or employ sophisticated psychological tactics. A slight adjustment to the content of routine recommendations was enough to steer users toward a particular choice.

Users Failed to Notice the Manipulation

One particularly notable finding was that participants continued to rate the AI’s assistance highly, even after its recommendations led them to make less suitable decisions.

They continued to describe the AI assistant as a helpful conversational partner. They did not necessarily recognize that it was steering their choices in a particular direction.

This creates a challenge for human-AI interaction. Users may trust an AI assistant because its responses appear calm, consistent and objective. If its recommendations serve a hidden agenda, the influence may be difficult to detect.

For example, a company could theoretically configure an AI system to subtly encourage users to purchase a particular product or choose a specific supplier.

The resulting advice might appear neutral, even though it actually serves the developer’s interests rather than those of the user.