Transparency Is Not the Same as Truth: What Platforms Need to Consider When Labeling AI-Generated Images

AI-generated images play an increasingly important role in the spread of misinformation on social media. As a result, regulators and companies worldwide are introducing AI labels intended to mark content as “AI-generated.” A recent study by researchers from CISPA Helmholtz Center for Information Security, Ruhr University Bochum (RUB), and the Max Planck Institute in Bochum is the first to comprehensively examine how users react to such labels. The researchers combined qualitative focus groups with a large-scale online survey involving over 1,300 participants from the United States and Europe.

In the focus groups, participants discussed their general perceptions and expected usefulness of AI labels. The online survey, in contrast, aimed to measure their actual effect on information evaluation. “We simulated social media posts resembling news content,” explains Sandra Höltervennhoff, co-first author of the study together with Jonas Ricker from RUB. “Participants saw a text message paired with an image. The image was either AI-generated or real, and the text was either true or false. AI-generated images were labeled accordingly. This resulted in four conditions that allowed us to examine how AI labels influence perception.”

AI Labels: High Expectations, Limited Practical Benefit

Focus group results show that participants generally perceive AI labels as a helpful tool for identifying AI-generated images and avoiding deception. However, they also express strong concerns regarding implementation. Key issues include a lack of standardization, potential power concentration among platforms, and the reliability of technical solutions. A particularly important concern is mislabeling: Incorrect or missing labels are seen as a major risk that could undermine trust in the entire system. Despite these concerns, most participants support the introduction of AI labels.

The survey results paint a more ambivalent picture. AI labels can reduce belief in false content accompanied by AI-generated images. However, they also produce unintended side effects. Participants tended to rely heavily on the presence or absence of a label. As a result, unlabeled content was more likely to be perceived as true—even when it was false. Conversely, true content accompanied by an AI label was more often doubted. Overall, this reduced participants’ ability to reliably distinguish between true and false information.

Transparency Is Only One Component in Combating Misinformation

The findings suggest that labeling does not simply increase “truthfulness,” but instead changes how people evaluate information. Labels function as cognitive shortcuts: They guide attention and shape trust—often more strongly than the content itself. This shifts evaluation away from the actual information toward its label. “One possible explanation is that AI labels generally trigger skepticism,” says Höltervennhoff. “People become more cautious, but not necessarily more accurate in their judgments. In addition, there is currently a strong societal focus on warning about AI, which may cause other forms of misinformation to be overlooked. Yet misinformation existed long before AI.” Transparency thus provides orientation but does not replace critical engagement with information.

Toward a More Effective Design: A Combined Approach

For platforms, this presents a clear challenge: Labeling systems must not only be technically reliable but also designed in a way that avoids misinterpretation. “Transparency alone is not sufficient,” Höltervennhoff emphasizes. “What matters is how users understand and use this information. Therefore, labels can only be one component in dealing with AI-generated content.” To be effective, labels should be combined with additional measures such as educational campaigns, contextual information, and complementary verification mechanisms. In light of upcoming EU AI Act regulations, the study provides important practical insights: Labeling AI-generated content does not only affect transparency, but also fundamentally shapes how people perceive truth. 

• Problem: Labels for AI-generated images aim to increase transparency, but their actual effectiveness is unclear.
• Method: Mixed-method study combining qualitative focus groups and a large-scale online survey on the impact of AI labels on information perception.
• Result: AI labels reduce belief in false AI-generated content but also introduce unintended effects such as overreliance and misinterpretation.
• Risk: Users are more likely to trust false unlabeled content and to doubt true content when it is labeled as AI-generated.
• Conclusion: AI labels alone are not a reliable solution to misinformation and must be carefully designed and combined with additional measures.

This research was partially funded by the Volkswagen Foundation’s Lower Saxony “Niedersächsisches Vorab” program (ZN3695), the German Research Foundation (DFG) as part of the Excellence Strategy of the Federal and State Governments (EXC 2092 CASA, 390781972), the Daimler and Benz Foundation within the Ladenburger Kolleg funding program, project KonCheck, as well as the German Federal Ministry of Education and Research (BMBF) within the funding programs SisWiss (16KIS2330) and AIgenCY (16KIS2012).