The Fake AI Monet, and Why Your Synthetic Persona Will Get the Same Treatment

A viral hoax around a fake AI Monet quietly replicated twenty years of psychology research — and exposes the exact same bias now distorting how the industry talks about synthetic personas. The real question is not whether something is human or synthetic. The real question is whether it helps you decide.

The painting that everyone knew how to criticize

On 12 May 2026, a user on X posted an image of water lilies with a deceptively simple caption: “I just generated this in the style of Monet with AI. Describe in detail what makes it inferior to a real Monet.”

The internet did what the internet does.

The composition was unbalanced. The depth was missing. The reflections were “flattened, like noise”. One commenter wrote 850 words to demonstrate the mediocrity of the rendering. The post passed 5.5 million views.

Then the reveal came.

No AI had ever touched the image. It was an actual Claude Monet, from the Water Lilies series, currently hanging at the Neue Pinakothek in Munich.

The interesting moment is not the hoax. The interesting moment is what happened after the reveal. Some critics deleted their comments. Others doubled down: “It’s not a good Monet,” “the reproduction is bad,” “I stand by my analysis.” A lot of words were spent avoiding the only honest sentence available:

This is not an anecdote. It is a replicated finding.

What makes the Monet hoax worth writing about is that it is not new information. It is a near-perfect, accidental field replication of a body of psychology research that has been quietly piling up since 2022.

The protocol is almost always the same. Take an artwork. Show it to two groups of participants. Tell one group it was made by a human. Tell the other it was made by AI. Then measure how they rate it on beauty, depth, creativity, emotional impact, willingness to pay.

The results are remarkably consistent across labs, countries, art forms, and years.

→Bellaiche and colleagues, in Cognitive Research: Principles and Implications (2023), found that across a series of experiments, the same artwork was rated lower on liking, beauty, profundity and worth when participants were told it was AI-created. The effect persisted even when participants admitted they could not tell the difference visually.^[1]

→Horton, White and Iyengar at Columbia Business School (2023, Scientific Reports) ran a series of experiments where people devalued art labeled as AI-made across multiple dimensions, even when they reported it was indistinguishable from human-made work, and even when they believed it had been produced collaboratively with a human. In one of their experiments, art labeled as AI-generated was valued 62% lower than the exact same work labeled as human-made.^[2][3]

→Millet, Buehler, Du and Kokkoris in Computers in Human Behavior (2023), in a paper titled Defending humankind: Anthropocentric bias in the appreciation of AI art, showed that the bias operates through a perceived creativity gap: people experience less awe when they think a piece was made by AI, because they decide upfront that it cannot have been creative.^[4]

→A 2024 study in Scientific Reports connected the bias to personality and prior attitudes: participants explicitly preferred what they believed to be human-made art, even when their implicit aesthetic responses to the AI-labeled and human-labeled works were essentially identical.^[5]

→A 2026 paper in Psychology of Aesthetics, Creativity, and the Arts pushed even further: the AI label is so cognitively loaded that it changes how viewers perceive basic visual features, including color and brightness, of the very same image.^[6]

→Most uncomfortable of all: a 2024 Frontiers in Psychology study showed that when people are asked to choose between human-made art and AI-made art without knowing which is which, they often prefer the AI piece. The negative bias is not a property of the work. It is a property of the label.^[7]

The pattern is brutally simple. The work does not move. The label moves the rating.

The Monet hoax did not discover anything. It just turned a peer-reviewed effect into a public spectacle.

Why I cannot let this sit in the art world

If this were only about paintings, it would be a fascinating curiosity for museums and art schools.

It is not only about paintings.

The exact same cognitive reflex is now driving an enormous amount of the conversation around synthetic respondents in market research, AI co-pilots in qualitative work, simulated focus groups, and persona-conditioned LLMs.

Replace “painting” with “insight”. Replace “made by AI” with “produced by a synthetic panel”. Replace “made by a human” with “collected from real respondents”.

Run the experiment again, with marketers, researchers, and brand managers in the role of the critics.

You get the same result. Two identical findings. Same numbers. Same verbatims. Same direction of recommendation. One is labeled “based on a 600-person field survey”. The other is labeled “based on 600 synthetic respondents.”

The second one gets dissected. The composition is unbalanced. The depth is missing. The verbatims are “flat, like noise”. Someone writes 850 words to demonstrate why this is not real research.

Then, sometimes, the reveal: the two reports came from the same study, with one column relabeled.

I have seen this happen, in rooms, in 2026.

The wrong question, and why it keeps winning

The current debate around synthetic respondents is framed around three questions, in roughly this order:

→Does it look like a real person? (Is the persona realistic?)

→Is it better than the real thing? (Does the synthetic study beat the field study?)

→Is it true? (Are the synthetic numbers “correct”?)

All three are the wrong question. They are the Monet questions. They are critiques of the label, not of the decision.

In my recent Substack piece, Fully Synthetic Audiences Did Not Fail. Poorly Grounded Synthetic Audiences Did., I argued that the real frame is methodological, not ontological. The relevant question is not human vs. synthetic. The relevant question is:

A generic LLM prompted to “act like a 35-year-old consumer interested in sustainability” is methodologically weak. Not because it is synthetic, but because it is under-specified. A real respondent recruited through a sloppy online panel, with no quotas, no behavioral filters and no category screening, is also weak — for exactly the same reason. The label “human” does not redeem a bad sampling design, just as the label “AI” does not condemn a well-designed simulation.

The Monet experiment is the aesthetic version of this confusion. The synthetic-vs-real debate is the decision-making version. In both cases, the label is doing the work that the analysis should be doing.

The only question that matters: does it help you decide?

If I had to compress this entire essay into a single sentence to hand to a CMO, an insights director, or a research buyer, it would be this:

The job of market research is not to produce truth. It is to produce the best possible decision under uncertainty, in the time and budget you have.

That reframing changes everything.

→A synthetic study that stress-tests three messaging routes in 24 hours, before you spend €120,000 on quantitative validation, is doing its job — even if no “real” person was ever recruited.

→A traditional 1,200-person field survey that arrives six weeks after the decision window closed is not doing its job — even if every respondent was a verified human with a verified IP address.

→A synthetic concept screen that kills 7 weak ideas out of 10 before they reach human fieldwork is doing its job — because the human fieldwork that follows is now sharper, smaller, faster and cheaper.

→A synthetic persona that hallucinates a confident answer on a question it was never grounded to handle is not doing its job — and the right response is to fix the protocol, not to ban the method.

The Monet hoax is funny because the critics were arguing about whether the painting was good, while the only relevant question — does this image move me? — was already answered, before they ever read the label.

The synthetic-research debate is the same, with money on the table. People argue about whether the respondents are real, while the only relevant question — does this study sharpen the decision I am about to make? — is sitting unanswered in the corner.

A short checklist for anyone tempted to be the next Monet critic

Before you write the 850-word LinkedIn comment about how synthetic respondents cannot possibly tell you anything useful, run this five-line audit:

If the answers are clean, the study deserves a seat at the decision table — whether the respondents were sampled from a panel, simulated from a model, or both.

If the answers are weak, the study should be sent back to the lab — whether it was labeled synthetic or not.

A small historical footnote

There is a quiet irony in the Monet story.

The Impressionists were savaged in their time. The very critiques being thrown at the AI-labeled water lilies in 2026 — flattened, unfinished, lacking depth, not real painting — are almost word-for-word the critiques thrown at Monet and his peers in the 1870s by the salons of Paris. The label then was “impressionist”, used as an insult. The label now is “AI”, used as a dismissal. The mechanism is identical: a category judgment substituting itself for a perceptual one.

Labels are cheap. Decisions are expensive. The skill — in art, in research, in business — has always been the same: to look at the thing in front of you, ask what it is actually doing for you, and refuse to let the sticker on the frame think on your behalf.

Laurent Florès is a professor at Paris-Panthéon-Assas University, former ESOMAR President, and founder of L’Atelier IA and FlashInsight. He writes about AI, market research, and the human factor.

Sources

1. Bellaiche, L., Shahi, R., Turpin, M. H., Ragnhildstveit, A., Sprockett, S., Barr, N., Christensen, A., & Seli, P. (2023). Humans versus AI: whether and why we prefer human-created compared to AI-created artwork. Cognitive Research: Principles and Implications. pmc.ncbi.nlm.nih.gov/articles/PMC10319694 ↩
2. Horton, C. B., White, M. H., & Iyengar, S. (2023). Bias against AI art can enhance perceptions of human creativity. Scientific Reports, 13. nature.com/articles/s41598-023-45202-3 ↩
3. Columbia Business School Research Brief — Beyond the Machine: Why Human-Made Art Matters More in the Age of AI. business.columbia.edu/research-brief/digital-future/human-ai-art ↩
4. Millet, K., Buehler, F., Du, G., & Kokkoris, M. D. (2023). Defending humankind: Anthropocentric bias in the appreciation of AI art. Computers in Human Behavior. sciencedirect.com/science/article/pii/S0747563223000584 ↩
5. Understanding how personality traits, experiences, and attitudes shape negative bias toward AI-generated artworks. Scientific Reports (2024). nature.com/articles/s41598-024-54294-4 ↩
6. Bias against AI art is so deep it changes how viewers perceive color and brightness. PsyPost / Psychology of Aesthetics, Creativity, and the Arts (2026). psypost.org ↩
7. Human perception of art in the age of artificial intelligence. Frontiers in Psychology (2024). frontiersin.org/articles/10.3389/fpsyg.2024.1497469 ↩
8. Florès, L. (2026). Fully Synthetic Audiences Did Not Fail. Poorly Grounded Synthetic Audiences Did. drlaurentflores.substack.com