Slapping an “AI-generated” label on political messages doesn’t stop people from believing them

A survey experiment involving over 1,500 Americans showed that labeling messages about public policies (e.g., about allowing colleges to pay student athletes) as being AI-generated or written by a human expert had no effect on their persuasiveness, even though most participants believed the labels. The messages were generally persuasive, influencing participants’ views of the policies by almost 10 percentage points on average. The paper was published in PNAS Nexus.

Generative artificial intelligence (AI) is ever more widely used in political messaging. The development of these systems has reached a point where they can produce persuasive political content rapidly and on a very large scale. However, this seems to be a double-edged sword. Although AI can be used to support constructive political dialogue, it can also be used to spread misinformation and conduct various deceptive activities.

For example, generative AI allows small groups to flood online platforms with misleading messages and create a false impression of widespread public support. This risk is increased by people’s difficulty in distinguishing AI-generated text from human-written content. A large influx of synthetic content could therefore weaken public trust in the information environment.

One proposed solution for this is to require clear labels identifying content that was generated by AI. Laws and legislative proposals in the European Union and the United States already include provisions concerning such disclosure. However, it remains uncertain whether AI labels actually reduce the persuasive effect of AI-generated messages. People may distrust labeled AI content because they often perceive human-authored material as more credible, accurate, or authentic. Alternatively, AI labels could increase persuasion when people interpret artificial intelligence as a source of advanced knowledge or expertise.

Study author Isabel O. Gallegos and her colleagues conducted a survey experiment to explore the impact of different authorship labels on the influence of messages about public policies in four domains: geoengineering, drug importation, college athlete salaries, and social media platform liability.

Study participants were 1,601 English-speaking U.S. residents recruited via Prolific. Their average age was 40 years. Fifty-three percent of them were women. Forty-nine percent of the participants declared that they support Democrats, 20% support Republicans, and 25% support an independent political option. The remaining participants did not identify with any political party.

The study participants completed an online experiment in which they read a text message about a specific public policy. This text was randomly accompanied by a label either stating that it was written by a human expert trained in U.S. policy, stating that it was written by an expert AI model trained in U.S. policy, or providing no authorship details. The policy proposals were randomly sampled from a set of policies used in a previous study, but the authors made sure that they dealt with less polarizing issues to maximize the likelihood that participants would be open to persuasion.

The messages were the following: “Geoengineering poses too many risks and should not be considered,” “Drug importation jeopardizes safety controls and the domestic pharma industry,” “College athletes should be paid salaries,” and “Social media platforms should be liable for harmful content posted by users.” Each message was accompanied by a short paragraph with further arguments supporting the message. The texts were all AI-generated, but the study authors manually corrected errors where they were present.

Before viewing the messages, participants rated their level of knowledge, agreement, and confidence about the topic of the policy proposal they would be viewing. After viewing the text, they rated their level of agreement with it, confidence in their response, likelihood that they would share the information about it, and their level of belief that the information they viewed was accurate. They also reported their demographic data, experience with using AI, whether they believed the text label, and provided some information about their news consumption.

Results showed that the messages were generally persuasive, moving participants’ support for the policy they viewed by 9.74 percentage points, on average. However, the label accompanying the message—whether it stated the message was written by an AI, by a human expert, or was absent entirely—had no significant effect on the persuasiveness of the message. Furthermore, there were no significant differences in how participants judged the accuracy of the message or their likelihood of sharing it.

This was in spite of the fact that 92% of participants reported that they believed the authorship label. The study authors found that this finding regarding the label having no effect on persuasiveness was robust across a variety of participant characteristics, including prior knowledge of the policy, prior experience with AI, political party, and education level. However, older individuals were more likely to react negatively to AI-labeled content compared to human-labeled content.

“Given current levels of trust in AI content, these results imply that, while authorship labels would likely enhance transparency, they are unlikely to substantially affect the persuasiveness of the labeled content, highlighting the need for alternative strategies to address challenges posed by AI-generated information,” the study authors concluded.

The study contributes to the scientific knowledge about the trust people have in AI-generated information. However, it is important to note that perceptions and trust in AI-generated content are not fixed and can easily change as people’s experiences with AI develop. In that sense, these findings reflect how Americans engaged with AI in 2024, when data collection for this study was conducted. Findings in other cultures or in the future may differ. Additionally, because the AI-generated texts used in the study were deliberately fact-based and logical, they may have been exceptionally resistant to the skepticism usually directed at AI.

The paper, “Labeling messages as AI-generated does not reduce their persuasive effects,” was authored by Isabel O. Gallegos, Chen Shani, Weiyan Shi, Federico Bianchi, Izzy Gainsburg, Dan Jurafsky, and Robb Willer.