Why a Warning Is Not a Safeguard
AI-mediated appearance images create several different problems, so they cannot be answered by one technical fix. A disclosure label addresses provenance: it tells a viewer that AI helped generate or alter an image. Parental mediation addresses interpretation and support. Critical digital literacy addresses the comparison itself, while product design can change how much distortion or repetition the system encourages. This chapter compares those layers so readers can match a response to the link it is meant to change.
The evidence supports a conditional conclusion rather than a universal winner. In the adult experiment, disclosure changed trust in technology but did not prevent the reported short-term body-satisfaction decline. In the child study, active and dialogical mediation was positively associated with body satisfaction, but the study did not test a particular conversation as a causal intervention. The right question is therefore not “Which safeguard solves the problem?” It is “What problem is this safeguard actually designed to solve?”
What a label can and cannot do
A label can make an invisible fact visible. A viewer who assumes that an image records an ordinary person may interpret it differently after learning that AI generated or altered it. That information supports provenance judgment and may affect trust in the technology. The adult randomized study found exactly that kind of separation: the informed group had a steeper decline in trust after viewing the images, while disclosure did not produce a significant interaction on either body-image measure.
The result makes the label valuable but incomplete. A viewer can understand that an image is synthetic and still find the displayed appearance desirable. They can distrust the technology while comparing their own appearance with the output. Provenance answers “How was this image made?” It does not answer “Why is this appearance being shown as better?” or “What should I do with the difference between this image and my ordinary body?”
That distinction also prevents an unfair standard for labels. The study tested an informed-versus-uninformed contrast in a particular viewing task. The accessible evidence does not provide the exact wording, placement, duration, or comprehension of the disclosure, and it does not test repeated exposure or different audiences. The result supports saying that a label was not enough in this setup. It does not support saying that every label is useless or that transparency should be abandoned.
A label may still be necessary even when it is insufficient. Without provenance information, a viewer may mistake a designed appearance for an unaltered reference. With it, the viewer has at least one fact needed for critical interpretation. But a fact is not the same as an intervention on comparison. If the image remains visually persuasive and socially rewarded, information about its origin may not change the standard it presents.
The trust finding also introduces a trade-off. A disclosure can protect a viewer from being misled about the technology while making the technology seem less trustworthy. That is not necessarily a defect; a decline in trust may be an appropriate response to a system that produces a misleadingly idealized image. But it means that labels should be evaluated on more than whether users notice them.