AI warning labels do not work: People still believe the message, study finds
Simply telling people that a political message was generated by artificial intelligence does little to reduce its influence, according to new research from Stanford University and other institutions, Qazinform News Agency correspondent reports.
The study, published in PNAS Nexus, examined whether labeling content as AI-generated changes how people respond to political and public policy messages. Researchers surveyed 1,601 Americans and presented them with AI-written arguments on a range of policy issues, including college athlete salaries, drug importation, geoengineering, and social media platform liability. Participants were randomly told that the messages had been created either by an expert AI system, a human policy expert, or were given no information about the source.
The results showed that the messages were generally persuasive regardless of the label attached to them. On average, participants' support for the policies shifted by nearly 10 percentage points after reading the arguments. However, researchers found no meaningful difference between the impact of AI-labeled messages, human-labeled messages, and unlabeled content.
Notably, most participants accepted the authorship labels as accurate. More than 90% of those shown AI or human labels believed the information about who created the message. Yet this awareness did not significantly affect whether they changed their views, considered the information accurate, or intended to share it with others.
The findings come as governments around the world consider rules requiring AI-generated content to be clearly identified. The European Union's AI Act and several legislative proposals in the United States include provisions aimed at increasing transparency through labeling. Researchers say their results suggest that while such measures may help people understand where content comes from, they are unlikely on their own to reduce the influence of AI-generated information.
The study also found that the lack of impact was largely consistent across political affiliation, education level, age, familiarity with AI, and prior knowledge of the topic. Although older participants showed slightly more skepticism toward AI-labeled content, the overall effect remained small.
Earlier, Qazinform News Agency reported that children are increasingly exposed to AI-generated videos on TikTok.
