← CHI 2026 map

CHI '26 · Honorable mention · full-paper review · confidence medium-high

"That's another doom I haven't thought about": A User Study on AI Labels as a Safeguard Against Image-Based Misinformation

Sandra Höltervennhoff , Jonas Ricker , Maike M. Raphael , Charlotte Schwedes , Rebecca Weil , Asja Fischer , Thorsten Holz , Lea Schönherr , Sascha Fahl

DOI PDF Program page Page JSON

This is a strong CHI paper because it does not stop at asking whether people like AI labels; it tests whether labels actually help and shows a subtle failure mode: labels can become a heuristic that shifts errors elsewhere. The mixed-methods setup and the large survey make the result credible and field-relevant.

Axes Lens

Rare contribution shape, typical evidence profile. The point here is not a score. It is to show what kind of claim the paper makes, and whether the evidence pattern is unusual or baseline in this 268 -review set.

Review Summary

This paper is compelling because it addresses a policy-relevant assumption that is easy to accept uncritically: if AI-generated content is labeled, users should be better protected from deception. The authors show that the story is more complicated. Their qualitative work suggests users see labels as helpful in principle, but the experimental study reveals a counterintuitive behavioral pattern: labels can reduce belief in false claims attached to AI-generated images while simultaneously encouraging overreliance on the label itself. That overreliance creates spillover effects, making participants more susceptible to false claims paired with human-made images and more hesitant to believe true claims shown with labeled AI-generated images. In CHI terms, the contribution is not a new interface widget but an empirical finding with clear design and policy implications. The paper’s strength is that it combines user expectations with a large survey experiment, so the authors can connect what people say about labels to how labels actually affect judgments. The main caveat is ecological validity: the survey setting is artificial and uses static, news-like posts, so the results should be read as evidence about judgment under controlled conditions rather than a full account of real-world social-media use. Even so, the paper is valuable because it identifies a plausible failure mode for disclosure systems and gives the field a more nuanced basis for thinking about AI labels as safeguards rather than simple fixes.

What Changed

Canon before

Prior CHI work on AI-content labels largely emphasized disclosure design, policy rationale, or user perceptions in isolation; this paper extends that canon by testing whether labels actually change misinformation judgments and whether they create compensatory effects.

Departure from common sense

The paper’s core counterintuitive result is that labels do not simply improve discernment; instead, users may treat them as a shortcut for truth judgments. The authors explicitly note that rather than increasing awareness, participants relied on labels to decide whether a claim was true or false, which can produce spillover errors on unlabeled or differently labeled content.

Actual novelty

The paper’s novelty is twofold: it combines qualitative focus groups on users’ expectations of AI labels with a large survey experiment on image-based misinformation, and it examines not only direct label effects but also mislabeling and unintended side effects. That makes the contribution more than a design recommendation; it is an empirical test of how labels reshape belief formation.

Evidence

Evidence from the abstract and selected full-text spans supports a mixed-methods study: five focus groups on label expectations and a survey with 1,354 participants using a 2×2×3 mixed factorial design. The paper reports that labels reduced belief in false claims with AI-generated images, but also induced overreliance and side effects, including greater susceptibility to false claims with human-made images and reduced belief in true claims with labeled AI images.

Limits

Method limits

The study relies on an online survey with static, news-like stimuli and a controlled factorial design, so it captures judgment effects rather than real-world social-media behavior or longitudinal adaptation. The focus-group component informs expectations, but the main causal claims rest on an artificial experimental setting.

Deployment limits

The findings caution against assuming that AI labels are universally protective in deployed moderation or disclosure systems. If users learn to use labels as a proxy for truth, labels may shift rather than eliminate misinformation vulnerability, especially when content mixes labeled AI images and unlabeled human-made images.

Boundary conditions

The results are bounded by the specific stimulus format, the label/mislabeling manipulations, and the participant pool described in the paper. Effects may differ for dynamic feeds, other media types, different label designs, or populations with different AI literacy and trust profiles.

Manual check: No manual check required for the extracted high-level claims, but the exact wording of the novelty claim and the precise scope of the experimental manipulations could be cross-checked against the full paper if needed.

Position in field

This paper sits at the intersection of AI disclosure policy, misinformation research, and HCI studies of trust calibration. Its main field contribution is to move beyond whether labels are desirable in principle and test how they behave as a safeguard in practice, including unintended consequences.

Abstract

As generative AI is increasingly contributing to the spread of deceptively realistic misinformation, lawmakers have introduced regulations requiring the disclosure of AI-generated content. However, it is unclear if labels reduce the risk of users falling for AI-generated misinformation. To address this research gap, we study the effect of labels on users' perception and the implications of mislabeling, focusing on AI-generated images. We first explored users' opinions and expectations of labels using five focus groups. Although participants were wary of practical implementations, they considered labeling helpful in identifying AI-generated images and avoiding deception. Second, we conducted a survey with 1,354 participants to assess how labels affect users' ability to recognize misinformation. While labels reduced participants' belief in false claims supported by AI-generated images, we found evidence of overreliance, leading to unintended side effects: Participants were more susceptible to false claims accompanied by human-made images, and were more hesitant to believe true claims illustrated with labeled AI-generated images.