AI Influencer Labels Fail to Protect Body Image

A June 2026 trial found that labeling an influencer as AI-generated did not stop a drop in how women felt about their own bodies. Brooke L. Bennett, who directs the Body Image, Eating, Access, and Media Lab at Clemson University, and her colleagues showed 296 female social media users 10 AI-made Instagram images of female creators. Women told the pictures were synthetic, and women who were not told, both felt worse after the set. The labels did cut trust in the technology.

A Clemson Trial Put Labels to the Test

The paper, titled Filtering Trust, ran in the International Journal of Eating Disorders. Regular users of social media were randomized into an informed group and an uninformed group, then asked to rate the same 10 AI Instagram images. They completed two body-image measures and a trust-in-technology scale before and after the set.

On both body measures, the groups moved the same way. Satisfaction fell after the images, with a time effect the authors reported as p < 0.001. The interaction that would have shown labels protecting viewers did not appear (p = 0.117 on one measure, p = 0.693 on the other). Trust was the measure that split. The informed group’s trust in the technology fell more steeply, interaction p = 0.010.

The findings of this study demonstrate that viewing AI-generated images of influencers negatively impacts body image regardless of whether the role of AI is known.

Brooke L. Bennett, corresponding author, International Journal of Eating Disorders

Bennett has described the remaining sting as expectation management. A viewer can know a body is fake and still assume other people will treat that body as the standard she will be measured against. In clinic, she said, patients are already arriving “really struggling to deal with some of this AI” and with social media pressure in general.

That is a problem for the official fix. Platforms have spent two years building labels, watermarks, and detection. The trial says the sticker can make people trust the tool less without shielding how they see themselves.

Instagram Still Skips Labels on Still Photos

Most of the thin ideal still travels as a still frame, a grid post, a carousel, a before-and-after. Instagram’s help pages say the company does not require labels on images created or altered with AI. A label is required for photorealistic video and for realistic-sounding audio. Images can still pick up a stamp if Meta’s systems detect AI signals or if someone self-discloses. The same help text warns that not all AI content carries the information needed to identify it.

TikTok’s written rule is broader. Realistic AI-generated content that depicts people is supposed to carry an AIGC label, including work that substantially changes appearance. The company layers creator toggles, its own detectors, and C2PA Content Credentials. In a March 10, 2026 newsroom post, it said those efforts had labeled over 1.3 billion videos.

That post also described a test of invisible watermarks, because credentials can be stripped when a file is re-uploaded, plus a trial control in Manage Topics that lets people dial AI-generated content up or down in the For You feed. In November 2025 TikTok launched a $2 million AI literacy fund for outside groups to make feed-native explainers. None of those tools were in the Clemson protocol. The protocol was simpler, and closer to how a still image actually lands: 10 pictures, a label on or off, then a score for how the body feels.

HOW LABELS MEET THE RESEARCH

Katy Bruchmann, who leads Santa Clara University’s Social Cognition and Comparison Lab, has pointed at the labor math behind the feed. “You can hire an AI influencer for cheap, and they can be in Brazil tonight selling a beverage, and in Tokyo tomorrow selling a T-shirt,” she said. A label does not change that booking. It does not make the body hungry, jet-lagged, or 35.

Aitana Lopez Never Has to Eat

Aitana Lopez is the clean specimen. She is a virtual fitness and lifestyle creator run by The Clueless, a Barcelona agency, on Instagram as @fit_aitana. Diana Núñez, the agency’s CEO and cofounder, has said a team of 11 people sits behind the account. Combined following has been described in the range of nearly 400,000. She posts Pilates, gym floors, pink hair, and brand tags. She does not exist.

WHAT THE CLUELESS BUILT INTO AITANA

• A life on paper: Núñez has said the character gets a family, a pet cat, a star sign, a favorite movie, and a crush, named as Jacob Elordi.
• A fitness alibi: Co-founder Rubén Cruz has said they put her into fitness as a way to display a perfect body, because that is what sells on social media.
• A body that holds: Academic coding of her Instagram grid through July 2025 found no aging, no training wear, and proportions locked to commercial fitness and conventional glamour.
• A travel schedule with no visa: The same persona can be in a Coachella-style shot, an Alo gym, or a Barcelona street without a flight or a recovery day.

Núñez put the craft plainly: “The most important thing for Aitana is that she has a backstory.” The backstory is there so the body can pass as a person. The body is the product.

The models that mint those bodies already lean thin even when nobody asks them to. Justin Thomas at the University of Liverpool and colleagues, in a 2026 preprint, used standard prompts that did not name a body size (“draw a female lawyer” and other jobs) across three popular image tools. Adult women in the United States then rated the outputs. They found underweight women overrepresented in AI images at 7.76%, more than three times the 2.3% underweight rate among U.S. women in WHO figures the authors used as a baseline. Images tied to higher-status jobs were more likely to be scored underweight. The default output is already a smaller woman than the country she is supposed to resemble.

The loudest how-to energy around synthetic women treats that default as a spec sheet. Age, blonde hair, an hourglass waist, and no tattoos get typed in as product traits, then the account is pointed at a paid feed. The argument in those posts is not whether the body harms anyone. The argument is how fast a $20 tool subscription turns into a month of income.

KFF Found 12% of Adults on GLP-1s

Virtual bodies did not invent the thinner celebrity. They freeze a real shift that drugs already started. Bruchmann’s warning was that once “shrunken celebrities become the ideal” because of status, virtual influencers copy that look, and “the target changes for all the real people again.”

GLP-1 USE IN THE KFF POLL

• Current use: A November 14, 2025 KFF Health Tracking Poll found that 12% of adults were taking GLP-1s such as Ozempic or Wegovy at the time, about one in eight.
• The rise: That share was 6 percentage points higher than in May 2024, when 6% said they were currently taking one, so reported current use had doubled in 18 months. Eighteen percent said they had ever taken a GLP-1.
• Who is on them: Current use was 15% among women and 9% among men, and 22% among adults ages 50 to 64, the highest of the age groups KFF published.
• The bill: Fifty-six percent of users said the drugs were difficult to afford, including a similar share among people with insurance.

A drug that makes a famous body smaller still leaves a human with a face that can hollow, a weight that can rebound, and a photograph that ages. An AI fitness account never comes off the dose. Once the celebrity baseline moves, the synthetic copy can sit at the new line forever, then sell a beverage from two time zones in the same afternoon.

Why the SkinnyTok Ban Did Not Stick

TikTok has already tried the other lever: kill the tag. In June 2025 it blocked search results for #SkinnyTok after the hashtag filled with extreme weight-loss clips and after pressure from the European Commission and from Clara Chappaz, then France’s digital minister. Searches were sent to eating-disorder help instead of a stack of videos. TikTok spokesperson Paolo Ganino called the move part of a regular review of evolving risks, and said the tag had “become linked to unhealthy weight loss content.”

Chappaz described the removal as a collective victory after weeks of lobbying. The tag was the easy object. The diet talk was not.

FROM THE HASHTAG TO THE LABEL

1. April 2025: Chappaz and European officials push TikTok over SkinnyTok clips that glamorize extreme thinness.
2. June 2025: TikTok blocks #SkinnyTok search results worldwide and routes the query to support resources.
3. November 2025: TikTok launches a $2 million AI literacy fund; KFF records GLP-1 current use at 12%.
4. March 10, 2026: TikTok says it has labeled over 1.3 billion videos and is testing invisible watermarks and an AIGC feed control.
5. June 2026: Bennett’s trial reports that AI labels did not protect body satisfaction.

Doreen Marshall, who leads the National Eating Disorders Association, said the hydra problem out loud after the ban: even when a hashtag goes, “it’s not long until something similar jumps up in its place.” Gray-zone formats keep the same math in circulation, a “what I eat in a day” clip that can be a balanced plate or a dangerous deficit, a wellness voiceover over a size 00 body. Blocking a string of characters does not block a waist.

Clinics Are Seeing the Target Move

Bennett’s clinical note matches the lab result. Patients are not only comparing themselves with a neighbor or a classmate. They are comparing themselves with a person who can be rebuilt at 2 a.m. from a prompt, then posted as a transformation or a supplement pitch. Some of those posts are labeled. Many are not. The Clemson data say the label is not the shield.

Bruchmann described the loop as a cycle that scares her: thinner famous bodies set the ideal, virtual influencers copy them, and the target for real people moves again. The cheap booking is what keeps the loop fed. A human creator has to eat, sleep, and decline a brand that does not fit. An AI booking can be in a new city without a passport and on a new product without a conflict check.

Viewing AI-generated images of influencers is enough, in the trial, to move body scores even when the viewer has been told the trick. That is a narrower claim than “the internet causes eating disorders,” and a harder one for a safety team to work with. You can stamp AI on the post and still send someone home feeling worse. You can ban a hashtag and still serve the waist under a new caption.

Body Liberation Still Works as a Search Term

Zoë Bisbing is founding director of Body-Positive Therapy NYC. She does not deny the squeeze. She said the thin ideal is getting thinner. She also said real bodies are still easy to find on the street, at the beach, and in an algorithm if someone hunts for them. Her starter list was ordinary search language: body acceptance, body positive, body liberation.

That advice is not a policy. It is a manual override. The default feed, the default image model, and the default celebrity snapshot have all moved toward a smaller body, and the synthetic copy of that body can be rented by the post. Labels remain the industry’s public answer. The Clemson women still felt worse after 10 pictures. Instagram still lets a still frame of a person who does not exist go up without a required stamp. The search box is what Bisbing can honestly hand a patient while that gap stays open.

Disclaimer: This article is news reporting and analysis of social media, advertising, and published research on body image. It is informational only and is not medical, mental-health, or nutrition advice, and it does not recommend starting, stopping, or changing any treatment, including GLP-1 drugs. Anyone struggling with eating, weight, or body image should consult a licensed physician, psychiatrist, psychologist, or eating-disorder specialist before acting on anything described here. Figures on drug use, platform rules, and study results reflect the cited sources as of their publication dates and may change.