When people love a picture, they tend to assume a person made it. That holds even when the picture was made by AI, and even for people who say they dislike AI art.

That is the finding of a new paper, Beauty Is Human: When People Love AI-Generated Art, They Think a Human Made It, posted to arXiv on 28 October 2026. It matters because most images you see online carry no credit line. Your audience is guessing who made your work, and their taste is steering the guess.

If you sell creative work or use AI images in marketing, that guess shapes trust, and trust shapes what people pay.

The authors are Zihan Chen and Di Zhu of Stevens Institute of Technology and Lei Nico Zheng of the University of Massachusetts Boston.

They did not run a new experiment. They reanalysed a large public one. In autumn 2024 the writer Scott Alexander ran an AI Art Turing Test on his blog, Astral Codex Ten. Readers looked at 50 pictures, 26 of them made by AI, guessed for each whether a human or an AI made it, and picked a favourite. About 11,000 people took part. The authors analysed the 9,152 who named a favourite and judged it.

For each person, they compared the guess about their favourite with that person's other guesses, and with other people's guesses about the same picture. That strips out people who always say "human" and pictures that just look human. What remains is the extra pull of liking.

People are poor judges. On average, respondents got 60.6% of their guesses right. A coin toss gets 50%, so that is only a little better than chance. AI pictures were judged human 42% of the time, so roughly two in five passed as human work.

Liking pushes the guess towards "human". A picture was 22.1 percentage points more likely to be judged human-made when it was the person's favourite. In plain terms: people called their own favourite human-made 76.6% of the time, about three in four. Everyone else looking at those same pictures called them human 54.6% of the time, about half.

Liking is doing the judging. The gap was almost identical whether the favourite was really AI-made (21.8 points) or really human-made (22.4 points). So people were not detecting a human touch, and the pattern held for all 50 pictures.

Strong views do not protect anyone. The people most negative about AI art picked fewer AI favourites, 47% against 60% for the most positive group. Yet they believed their favourite was human-made 88% of the time. In every opinion group, between a third and two-fifths (33.5% to 40.8%) loved an AI picture they thought a person had made.

It works for words too. The authors reanalysed a 2024 study by Brian Porter and Edouard Machery, which found AI-generated poetry is indistinguishable from human-written poetry and is rated more favourably. Among 230 readers told nothing about the source, each extra point of quality they gave a poem made them about 11.5 percentage points more likely to call it human, whoever wrote it.

The paper studies art and poems, not ads or listings. What follows is our reading of how it carries over, so test it on your own audience.

If you sell art, design or content. Quality will not prove you made it by hand. Buyers credit AI work they like to humans and may doubt human work they find ordinary. "Human-made" is a claim you have to show with evidence.

If you use AI images in marketing. Customers may enjoy them and assume a designer made them, until someone finds out and feels misled. The more your brand depends on personal craft (a bakery, a coach, a handmade shop), the bigger that risk.

On disclosure and trust. Say nothing and people fill the gap, often in your favour, which becomes a problem later. A plain disclosure costs little. In Europe it can also be a legal matter: since 2 August 2026, Article 50 of the EU AI Act has required deepfakes to be disclosed, with a lighter duty for clearly artistic work, which can be disclosed in a way that does not spoil it. Check what applies to you. HackrLife has written before about the premium audiences place on trust.

On pricing human-made work. If buyers pay more for human craft, they need a reason to believe it is there, so the process story is part of what you sell. Sketches, time-lapses and signed certificates are the evidence behind your premium.

1. Test your own visuals in ten minutes. Show eight of your images, some AI-made, some human-made, unlabelled, to five customers or friends. Ask two questions for each: "Do you like it?" and "Did a person or an AI make this?" If their favourites get called human, you know which images rest on an unspoken assumption.

2. Write one disclosure line and use it everywhere. Examples: "Sketched by hand, coloured with AI assistance." "Product photos by us; background scenes generated with AI." "Every illustration drawn by hand, no AI." Put it on your site, shop listings and social bios, the same wording each time.

3. Signal human craft where it is the product. Spend the effort where buyers pay for the maker: original art, commissions, portraits, custom design, personal writing. Add a process clip, a sketch scan or a "how this was made" note. For a stock-style banner, a one-line disclosure is enough.

4. Keep your receipts. Save source files, drafts and dates. If a client, buyer or platform asks how something was made, you can show them in a minute.

5. Run a simple A/B test. The authors suggest testing whether reliable disclosure changes what people choose, share or pay for. You can do a small version. Version A shows one product or post with no note. Version B adds your disclosure line and a process image. Split traffic evenly for two weeks, or until each version has around 100 visits (a rough rule of thumb, not a statistical guarantee). Compare clicks, saves, enquiries or sales. If B holds up or wins, disclose by default. If it loses, try different wording before you drop it.

If you build creative tools or marketplaces: let makers attach a short process note that stays with the work when it is shared or resold. The paper's conclusion points the same way: the results "motivate reliable provenance information rather than reliance on subjective liking as evidence of authorship."

The respondents were readers of one blog who chose to take a test, so they are probably more interested in AI than most people. The 50 pictures were chosen to be hard to classify, so the numbers describe these pictures, not all art. Duplicate entries cannot be ruled out, and opinions about AI art were measured after the test.

Most important, neither study changed how much people liked a work, so the direction is unproven. Liking may make people assume a human maker, believing a human made something may make them like it more, or both. Treat it as a strong signal and test it on your own audience.

For more on how platforms are handling the flood of generated media, see our look at video platforms' AI slop problem and how Civitai turned AI art styles into a market.

From today's Daily Briefing →