Essay
There is no one in the mirror
A decade before generative AI, clients were asking for stock photography that didn’t look like stock photography. The brief never changed. Only the price did.
· 14 min read
A decade before anyone could generate a convincing human being, marketing departments were already asking for something close to it. They wanted stock photography that did not look like stock photography.
The image industry built products for that brief. Getty Images launched the Lean In Collection with LeanIn.org in February 2014, sold on the promise of authentic images of women in place of the clichés the category kept producing. Dove, Getty and Girlgaze followed in March 2019 with Project #ShowUs, roughly 5,000 undistorted photographs shot by women, non-binary and female-identifying photographers across 39 countries. Serious efforts, both of them, answering the same commercial request. Sell us realness we did not have to go out and make.
Scholars had described the machinery earlier. Paul Frosh called stock photography the wallpaper of consumer culture in Media, Culture & Society in 2001. In 2023 Giorgia Aiello, Crispin Thurlow and Lara Portmann examined 600 stock photographs from three global image banks for Social Media + Society and found social life flattened and corporatized on the way to the library. Candid was a purchasable genre long before it was a prompt.
Generative AI changed the arithmetic. A licensed stock photograph is priced in dollars; a generated one in cents. The US Bureau of Labor Statistics puts the median wage for photographers at $21.47 an hour in its May 2025 edition, and counts about 145,000 of them, fewer than its previous edition did. The BLS does not say why, and neither will I. The brief stayed the same and the floor under it fell out.
The argument everyone is having
The fear clients bring to this is oddly specific. It is not about the image at all. It is about being caught.
Gallup surveyed 3,270 US adults in May 2026 for the Bentley University Business in Society report, on a probability-based panel with a margin of error of 2.4 points. Forty-nine percent viewed businesses using AI in advertising negatively, against 19% positively. Then it split the question by task, which is more useful than any headline. Seventy-five percent accepted AI helping employees brainstorm or draft, with disclosure. Fifty-three percent accepted AI producing final text, images or video. And 62% called it unacceptable to use AI to create people or voices in advertising, even with disclosure.
Even with disclosure is the phrase to sit with, because it moves the problem. The objection survives being told.
Guess ran that test in public. It placed an advertisement featuring two AI-generated models, produced by Seraphinne Vallora, in the August 2025 US issue of Vogue, with a disclosure line in small print. This was paid advertising, not editorial, and Vogue said afterwards that it has not used AI models in its own fashion production. The reaction arrived anyway, amplified by a TikTok post past 2.7 million views and covered by CNN, ABC News and Forbes.
The usual reply is that nobody can tell. Sophie Nightingale and Hany Farid, in PNAS in 2022, found participants sorting synthetic from real faces with 48.2% accuracy, below guessing, and a trained group reaching only 59%. Those participants also rated the synthetic faces as more trustworthy than the real ones, 4.82 against 4.48 on a seven-point scale, a gap the authors call small.
The usual correction is a much larger study by Kamali and colleagues at CHI in 2025, with 749,828 judgments from 50,444 participants: 76% accuracy on AI images, 74% on real photographs. Those numbers describe a deliberately hard subset, because the authors excluded every image identified correctly more than 80% of the time, removing 14% of all observations. Before that exclusion, accuracy on individual AI images ran from 32% to 99%. Nightingale and Farid were testing cropped faces, the hardest category there is. The two agree: accuracy is a property of the image, not of the viewer, and human curation made images harder to catch. Choosing the best of forty generations is what a creative team does by default.
The law is splitting, and most marketing teams have not registered it. Article 50 of the EU AI Act, applying from 2 August 2026, requires anyone deploying a deepfake to disclose it clearly, in a form a person understands without technical tools. Article 3(60) defines one as AI-generated content resembling existing persons, objects, places or events that “would falsely appear to a person to be authentic or truthful”. Read that looking for the word intent. It is not there.
The United States has no equivalent rule for images. The FTC’s rule on reviews and testimonials, 16 CFR Part 465, effective October 2024, bans testimonials misrepresenting they come from someone who does not exist, including AI-generated reviewers. That covers an invented person recommending you, not one standing in your hero image. While researching this I kept finding articles citing an FTC staff guidance from March 2025 that supposedly requires advertisers to disclose AI imagery. I could not find it on ftc.gov.
On 23 September 2026 we asked a search engine for it once more, and it produced the guidance: three core principles, a requirement to disclose photorealistic AI images however realistic they look, even warning letters sent to seven fashion brands. None of it exists, and every source behind the answer was a vendor blog citing another vendor blog. A rule nobody wrote now has the appearance of one, because intermediaries repeated it and a machine synthesized it back wearing the manner of a verified fact. Nobody had to lie for an industry to start complying with an imaginary law.
Why we were told to put people there
Underneath the disclosure argument is an assumption almost nobody says out loud. We put people on pages because the viewer sees themselves in them.
The version that circulated in marketing credits mirror neurons. The viewer’s brain simulates the person in the picture, and identification follows. Martin Lindstrom put that idea into wide circulation with Buyology in 2008, whose third chapter is titled, without ambiguity, “I’ll Have What She’s Having: Mirror Neurons at Work”. It is a reasonable idea, clearly told, and it handed a generation of marketers a mechanism they could name. The evidence does not carry it that far.
The neurons are real, found in macaque premotor cortex in the 1990s and recorded directly in humans once, by Mukamel and colleagues in 2010. The interpretation is what did not survive. In 2009 Gregory Hickok set out eight problems with the theory that mirror neurons underpin action understanding, noting that disrupting the relevant region in monkeys impairs grasping but has never been shown to impair perceiving. A year later Cecilia Heyes argued they are better explained by ordinary associative learning.
So the mechanism marketing borrowed is contested where it was borrowed from. The observation underneath it survives, and it has to be measured where it lands.
Attention is one thing, trust is another
Where it lands is attention, and the effect is precise enough to design with. Two eye-tracking studies in Frontiers in Psychology found that a face gazing toward the product beats a face gazing at the viewer: Sajjacholapunt and Ball in 2014 for attention and brand memory, and Palcu, Sudkamp and Florack in 2017, who put the odds of looking at the product 4.1 times higher when the model’s gaze pointed at it. Read the rest of the 2017 results before spending on the effect. Gaze direction moved purchase intention weakly, at p = 0.035, and did nothing significant to how attractive participants found the product or what they would pay.
Trust is harder, because the answer depends on how you measure it. By 2003 the published results were in conflict, and Jens Riegelsberger, Angela Sasse and John McCarthy said so: an earlier study of their own had found staff photographs producing negative reactions and lower trust, so they built an experiment to settle it. They added photographs of sales assistants to twelve e-commerce sites, half with good reputations and half with bad, and measured trust in 115 participants with a method from experimental economics that put the participant’s own money at risk.
Nothing moved. The presence of a photograph had no effect on trust, t(114) = -.01, p = .99. Neither did how trustworthy the person in it looked, F(2,228) = .01, p = .99.
A different method reached a compatible place. Dianne Cyr, Milena Head, Hector Larios and Bing Pan combined survey, interview and eye-tracking work across three countries for MIS Quarterly in 2009. Faces raised image appeal and perceived social presence, and both predicted trust. The direct link, from human images to trust itself, was the one thing they did not find.
Faces move attention reliably. They move stated trust indirectly at best. When the measure costs the participant something, the effect stops showing up.
There is one more result in that 2003 paper, in a section no summary of it carries, and it is worse than a null. The photographs degraded people’s judgment. Assessment error was significantly greater for sites with a photograph than without, t(114) = -2.17, p = .03. With no photograph, participants rated good and bad vendors differently, 6.6 against 6.0. Adding a photograph erased the difference, F(1,114) = 5.48, p = .02. In the authors’ summary, the photographs raised the perceived trustworthiness of the poorly performing vendors and lowered it for the ones with good reputations.
Sit with that a second, with your own homepage in mind. The face was not inert decoration. It helped the worst sellers and hurt the best ones, by covering the only signal buyers had.
The disclosure trap, and the way out of it
That brings the argument back to the invented person, and to a 2025 finding that appears to point the wrong way.
Zeph van Berlo and Priska Breves, in Computers in Human Behavior Reports in August 2025, ran two experiments with 245 and 429 US Instagram users on how prominently a virtual influencer’s artificial nature is disclosed. Prominent disclosure produced lower credibility than subtle disclosure and lower than human influencers, carrying through to brand attitude and purchase intention. Read alone, that argues for burying the label.
The same abstract says why it does not. Credibility for the subtly disclosed influencers came out mixed against human influencers rather than clean. The quiet version does not win; it blurs. And the rest of the evidence closes that exit. Gallup’s 62% rejected AI-created people with disclosure held constant, and Guess disclosed in print and was punished regardless. Prominent disclosure costs credibility, quiet disclosure costs you whenever someone notices, and detection research says you cannot predict which images those will be. There is no position on that dial where an invented person works.
One setting does put a human image to work, and it points the same way. Deborah Small, George Loewenstein and Paul Slovic showed in 2007 that people give considerably more to one identified individual than to statistical victims. That effect belongs to a specific person who exists. A generated face is the shape of it with nobody inside.
The numbers nobody can trace
Marketing repeats a small set of figures here. That authentic images convert 45% better than stock, attributed to HubSpot. That team photographs produce a 65% higher trust score. That customer photographs lifted signups by 35%. I tried to trace each to a published method, sample and date, and none resolved.
The 45% is the interesting failure, because it is half real. There is a 45% on HubSpot’s blog. It belongs to a case where click-through to case studies improved by nearly that much after a change in visual hierarchy and format, and has nothing to do with photographs of people. A real number about one thing, reattributed to another, repeated until it became canonical.
Money is the same. No industry body publishes a benchmark for what a commercial shoot costs; every range in circulation comes from photographers’ own pricing pages. Nor is there public first-tier evidence that AI imagery has moved sales in either direction. Coca-Cola, two years into AI-generated holiday advertising, says its 2024 campaign performed well in private testing, which with no published method is a claim rather than a result.
The gap cuts both ways. For years the case for human photography has rested on numbers nobody can check, which is the same position as the client who wants an AI model nobody will notice.
Every one of them has the same face
There is a second cost, and for a marketing team it should be the one that decides the question, because it has nothing to do with ethics.
Generative models do not produce a random person. They produce something closer to the average person, and that has now been measured. Dunn and colleagues, in the British Journal of Psychology in February 2026, ran deep neural networks optimised for face identity and found AI faces distributed more centrally in face-space than real ones. The human half of their study is the part worth reading twice. Thirty-six super-recognizers spotted AI faces better than a typical sample by 15%, and better than a motivated, high-performing control group of 89 by 7%, with their advantage tied to sensitivity to that hyper-average look. Before filing this as independent confirmation, note that four of its six authors also wrote the 2023 Psychological Science paper that named AI hyperrealism. One research group is extending its own line.
The behavioural picture matches. Alice Mado Proverbio and Mariia Dosaikina, in Scientific Reports in 2026, put 220 real photographs against 220 StyleGAN2 faces. Across two separate groups of 25 raters, the AI faces scored higher on attractiveness, 3.12 against 2.31, and on familiarity, 2.70 against 2.24, while the second group identified them as artificial only 33% of the time, against 62% for the real ones. Their own conclusion is a dissociation rather than a clean win for the machines: behaviour failed to catch the faces and the EEG signal did not. One limit the authors flag themselves matters here. Every face they tested was white. Miller and colleagues had already found in 2023 that the hyperrealism effect did not appear for faces of people of colour. The average these models produce is the average of their training data.
Nightingale and Farid had reached for the same explanation back in 2022. Their wording in the discussion is that synthetic faces may be rated more trustworthy “because synthesized faces tend to look more like average faces which themselves are deemed more trustworthy”. The reference they hang that on is Sofer, Dotsch, Wigboldus and Todorov, in Psychological Science in 2015, under a title that saves everyone some time: “What is typical is good.” The link from typicality to trust was already peer-reviewed. What arrived later was the measurement that these faces are typical.
Now set that beside what psychology has known about average faces since 1990, when Judith Langlois and Lori Roggman showed in Psychological Science that mathematically averaged composite faces were judged more attractive than almost every individual face that went into them. Averageness is why these images test well in a room. It is also why they slide off.
In 1979 Leah Light, Fortunee Kayra-Stuart and Steven Hollander ran four studies in the Journal of Experimental Psychology: Human Learning and Memory and found recognition memory for prototype-like faces worse than for unusual ones, holding across incidental and intentional learning, presentation times from 3 to 15 seconds, and retention intervals from 3 to 24 hours. A fifth experiment identified the reason: interitem similarity. Two years later the same group published the finding under a blunter title. Attractive people are harder to remember. Tim Valentine’s face-space account in 1991 gave the pattern a framework. Faces are encoded by how far they sit from the norm, and the ones near the norm are the ones that get confused with each other.
The mechanism is still argued over, and that should be said. Morris and Wickham re-examined it in 2001 and concluded that typicality, attractiveness, mere exposure and people’s beliefs about their own memory are tangled together in ways the earlier theory did not separate. The direction of the effect is not what they dispute.
The image does not sit alone either. Raymond Burke and Thomas Srull showed in the Journal of Consumer Research in 1988 that memory for a brand’s advertising was inhibited by later exposure to advertising for competing brands in the same class, and, less comfortably for anyone with a product line, by advertising for other products from the same manufacturer. A third experiment found that the presence of competitive advertising changes the relationship between repetition and memory. And the sameness is documented rather than felt. AlDahoul, Rahwan and Zaki mapped bias in Stable Diffusion across six races, two genders, 32 professions and eight attributes, then ran a separate analysis of how far the model makes people of the same race resemble one another. That one found significant racial homogenization, with nearly all Middle Eastern men depicted as bearded, brown-skinned and in traditional attire.
So the face on the page is more average than a real one, average faces are harder to remember, and it competes with every other page built from the same handful of models.
If there is no budget
So, no money for a shoot. AI, stock, or neither.
The strongest finding in this whole area turns out not to be about people. Shunyuan Zhang, Dokyun Lee, Param Vir Singh and Kannan Srinivasan, in Management Science in 2022, ran a difference-in-differences analysis on 7,423 Airbnb properties over sixteen months and found that listings with verified photographs, taken by the platform’s photographers rather than the host, had 8.98% higher occupancy. That is photography of the thing being sold, done competently, and it is the only causal, large-sample, peer-reviewed result in the area. The caution about the cheap version comes from a Cornell study of peer-to-peer marketplaces, which ranked trust highest for good user-generated photographs, then stock, then poor ones. Shooting it yourself helps only if the result is good.
What the evidence supports is narrow and usable. Put the budget into photographing the specific thing you sell: the room, the product, the work, the people who do it. Put the rest into making that photograph competent rather than into acquiring a stranger. If the money will not reach a competent photograph of a person, the defensible options are a competent photograph of something else, or none.
I cannot tell you the page without a person performs better, because nobody has published a first-tier comparison of pages with and without human imagery, or of illustration against photography. Nor has anyone run the whole chain in this essay end to end on a working website. The pieces were measured separately, and they point the same way.
The client asks for a person so the visitor will see themselves in it. The mechanism they were sold for that does not survive the literature. What a face does reliably is pull the eye, which is worth having. The invented face is built closer to the average than a real one, which is why it looks good in the room and why it is harder to remember afterwards. In the one experiment that measured trust with money on the table, putting a face on the page made bad sellers look better and good ones look worse. Then it goes live on a page competing with every other page built from the same models, in an environment where similar material has been shown to interfere with memory for all of it.
Read as a purchase, the client asks for identification and receives a composite. They pay for a face because faces get noticed, and they buy the most forgettable one available.
Five questions before you approve the image
Five questions worth putting in a brief, or asking in the room before anyone signs off on a face.
What is this person meant to make the visitor feel, and who decided that a person was the way to do it? If the answer is that people build trust, the experiment that measured trust with the participant’s own money at risk did not find any.
If we put the words “AI-generated” beside this image, does the case for using it survive? Sixty-two percent of US adults called AI-created people in advertising unacceptable with disclosure held constant, which makes the label a test of the decision rather than a fix for it.
Are we buying attention or trust? A face moves the first reliably, and gaze direction moves it with real precision. In the one study that put money behind the second, the photograph moved nothing, and it made bad sellers look better and good ones look worse.
Would anyone recognize this face if they met it again next week? Synthetic faces sit closer to the average than real ones, faces near the average are the ones that get confused with each other, and this one will arrive beside every other page built from the same models.
What would this budget buy if we photographed the thing we sell instead? It is the only question on this list with a causal, large-sample, peer-reviewed answer attached to it.
None of this is an argument for pages without people. Photograph your own, of your own work, and most of the findings in this essay stop applying to you. The trouble starts where a face is bought to stand in for a relationship nobody has built yet.
Underneath all five is the question a brief almost never asks, and it is the cheapest of them to answer: what is the person in that picture doing, and would a photograph of the thing you sell do it better?
Sources
This list is reproduced identically in the Spanish version of the essay, with titles, journals and verification notes left in their original language.
Peer-reviewed research
Aiello, G., Thurlow, C. & Portmann, L. (2023). “Desocializing Social Media: The Visual and Media Ideologies of Stock Photography.” Social Media + Society, 9(1). DOI 10.1177/20563051231156363. Social-semiotic analysis of 600 stock photographs from three image banks.
Berlo, Z. van & Breves, P. (2025). “Disclosing the virtual nature of virtual influencers: The effect of disclosure prominence and the role of product digitality.” Computers in Human Behavior Reports, 19, 100742. DOI 10.1016/j.chbr.2025.100742. Open access, CC BY. Two online experiments, N=245 and N=429, US female Instagram users aged 18–34. Official abstract verified verbatim.
Cyr, D., Head, M., Larios, H. & Pan, B. (2009). “Exploring Human Images in Website Design: A Multi-Method Approach.” MIS Quarterly, 33(3). Survey, interviews and eye-tracking in Canada, Germany and Japan. Page range not given here: the MISQ record shows 539–566 and the AIS eLibrary record 530–566, and the discrepancy is unresolved.
Frosh, P. (2001). “Inside the image factory: stock photography and cultural production.” Media, Culture & Society, 23(5), 625–646. DOI 10.1177/016344301023005005. And Frosh, P. (2003). The Image Factory: Consumer Culture, Photography and the Visual Content Industry. Berg.
Gallese, V., Fadiga, L., Fogassi, L. & Rizzolatti, G. (1996). Brain, 119(2); di Pellegrino, G., Fadiga, L., Fogassi, L., Gallese, V. & Rizzolatti, G. (1992). Experimental Brain Research, 91(1).
Heyes, C. (2010). “Where do mirror neurons come from?” Neuroscience & Biobehavioral Reviews, 34(4), 575–583.
Hickok, G. (2009). “Eight Problems for the Mirror Neuron Theory of Action Understanding in Monkeys and Humans.” Journal of Cognitive Neuroscience, 21(7), 1229–1243.
Kamali et al. (2025). “Characterizing Photorealism and Artifacts in Diffusion Model-Generated Images.” Proceedings of CHI 2025, ACM. DOI 10.1145/3706598.3713962. 749,828 observations from 50,444 participants across 450 AI-generated and 149 real images. The 76% and 74% figures follow an exclusion of every image identified correctly in more than 80% of observations, which removed 14% of all observations; the 32–99% range is from before that exclusion.
Mukamel, R., Ekstrom, A. D., Kaplan, J., Iacoboni, M. & Fried, I. (2010). “Single-Neuron Responses in Humans during Execution and Observation of Actions.” Current Biology, 20(8), 750–756. DOI 10.1016/j.cub.2010.02.045.
Nightingale, S. J. & Farid, H. (2022). “AI-synthesized faces are indistinguishable from real faces and more trustworthy.” PNAS, 119(8), e2120481119. DOI 10.1073/pnas.2120481119. Full text verified in PubMed Central. The trustworthiness ratings are 4.82 against 4.48 on a seven-point scale, d = 0.49, described by the authors as a small effect.
Palcu, J., Sudkamp, J. & Florack, A. (2017). “Judgments at Gaze Value: Gaze Cuing in Banner Advertisements, Its Effect on Attention Allocation and Product Judgments.” Frontiers in Psychology, 8, 881. N=137, eye-tracking.
Riegelsberger, J., Sasse, M. A. & McCarthy, J. D. (2003). “Shiny happy people building trust? Photos on e-commerce websites and consumer trust.” Proceedings of CHI 2003, 121–128. DOI 10.1145/642611.642634. N=115, twelve sites, trust measured with financial risk. Full text read from the author’s copy at University College London. Null results: photo presence t(114) = -.01, p = .99; depicted trustworthiness F(2,228) = .01, p = .99. Assessment error with photographs t(114) = -2.17, p = .03; interaction F(1,114) = 5.48, p = .02.
Sajjacholapunt, P. & Ball, L. J. (2014). “The influence of banner advertisements on attention and memory: human faces with averted gaze can enhance advertising effectiveness.” Frontiers in Psychology, 5, 166.
Small, D. A., Loewenstein, G. & Slovic, P. (2007). “Sympathy and callousness: The impact of deliberative thought on donations to identifiable and statistical victims.” Organizational Behavior and Human Decision Processes, 102(2), 143–153. Replicated and extended by Maier et al. (2023), Collabra: Psychology, 9(1), 90203.
AlDahoul, Nouar, Rahwan, Talal & Zaki, Yasir (2025). “AI-generated faces influence gender stereotypes and racial homogenization.” Scientific Reports, 15(1), 14449. DOI 10.1038/s41598-025-99623-3. Open access. The bias analysis covers six races, two genders, 32 professions and eight attributes; the racial homogenization analysis is a separate measurement within the same paper.
Miller, E. J., Steward, B. A., Witkower, Z., Sutherland, C. A. M., Krumhuber, E. G. & Dawel, A. (2023). “AI Hyperrealism: Why AI Faces Are Perceived as More Real Than Human Ones.” Psychological Science, 34(12), 1390–1403. DOI 10.1177/09567976231207095. Cited in the essay for the finding that the hyperrealism effect did not appear for faces of people of colour.
Sofer, C., Dotsch, R., Wigboldus, D. H. J. & Todorov, A. (2015). “What Is Typical Is Good: The Influence of Face Typicality on Perceived Trustworthiness.” Psychological Science, 26(1), 39–47. The primary source behind the typicality-to-trust link, cited by Nightingale and Farid as their reference 12.
Burke, Raymond R. & Srull, T. K. (1988). “Competitive Interference and Consumer Memory for Advertising.” Journal of Consumer Research, 15(1), 55–68. Three experiments. Abstract-level verification; the full text is paywalled, so the essay reports only what the abstract states, including that Experiment 3 found competitive advertising changes the relationship between repetition and memory, without a direction.
Dunn, James D., White, David, Sutherland, Clare A. M., Miller, Elizabeth J., Steward, Ben A. & Dawel, Amy (2026). “Too good to be true: Synthetic AI faces are more average than real faces and super-recognizers know it.” British Journal of Psychology, article bjop.70063. DOI 10.1111/bjop.70063. Published 18 February 2026. Abstract-level verification; no open copy of the full text. Four of the six authors also wrote Miller et al. (2023), which the essay states.
Langlois, J. H. & Roggman, L. A. (1990). “Attractive Faces Are Only Average.” Psychological Science, 1(2), 115–121. DOI 10.1111/j.1467-9280.1990.tb00079.x
Light, L. L., Kayra-Stuart, F. & Hollander, S. (1979). “Recognition memory for typical and unusual faces.” Journal of Experimental Psychology: Human Learning and Memory, 5(3), 212–228. DOI 10.1037/0278-7393.5.3.212. Five experiments.
Light, L. L., Hollander, S. & Kayra-Stuart, F. (1981). “Why Attractive People are Harder to Remember.” Personality and Social Psychology Bulletin, 7(2), 269–276. Citation verified; no abstract available in public databases, so nothing beyond the title is verified.
Morris, P. E. & Wickham, L. H. V. (2001). “Typicality and face recognition: a critical re-evaluation of the two factor theory.” Quarterly Journal of Experimental Psychology A, 54(3), 863–877. Included because it complicates the mechanism behind the typicality effect.
Proverbio, Alice Mado & Dosaikina, Mariia (2026). “Neural signatures of hyper-realistic AI-generated faces: dissociating behavioral indistinguishability from implicit neural evaluation.” Scientific Reports, 16(1), 22944. DOI 10.1038/s41598-026-59487-7. Full text read. EEG over 440 faces, plus behavioural validation in two separate groups of 25 raters, carried over from the team’s earlier work. All stimulus faces were of Caucasian ethnicity, which the authors list as a limitation. The paper’s conclusion is a dissociation: behaviour failed to detect the faces while the neural response did not.
Valentine, T. (1991). “A unified account of the effects of distinctiveness, inversion, and race in face recognition.” Quarterly Journal of Experimental Psychology A, 43(2), 161–204. DOI 10.1080/14640749108400966. Five experiments.
Zhang, S., Lee, D., Singh, P. V. & Srinivasan, K. (2022). “What Makes a Good Image? Airbnb Demand Analytics Leveraging Interpretable Image Features.” Management Science, 68(8), 5644–5666. Deep learning and difference-in-differences across 7,423 properties over sixteen months; verified photographs, taken by the platform rather than the host, gave 8.98% higher occupancy. The figures of 17.51%, $2,521 and 58.83% that circulate with this citation belong to the 2017 working paper (7,711 properties) and are not in the published article.
Official statistics and regulation
Bentley University–Gallup, Business in Society survey, 2026. Probability-based Gallup Panel, n=3,270 US adults, fielded 4–11 May 2026, margin of error ±2.4 points at 50%, weighted to national demographics.
European Commission, Transparency obligations under Article 50 of the AI Act, and Regulation (EU) 2024/1689, Articles 3(60), 50 and 113. Applying from 2 August 2026. No grace period to December 2026 exists in the Regulation; the transitional regime for systems already on the market is in Article 111 and runs on different dates.
US Bureau of Labor Statistics, Occupational Outlook Handbook, Photographers (SOC 27-4021), May 2025 edition. Median wage $21.47 an hour, $44,660 a year; about 145,000 jobs in 2025, against 151,200 in 2024; 67% self-employed.
US Federal Trade Commission, final rule on the use of consumer reviews and testimonials, 16 CFR Part 465, announced 14 August 2024, effective 21 October 2024.
Preprint, labeled as such
“Understanding Image Quality and Trust in Peer-to-Peer Marketplaces,” Cornell, arXiv:1811.10648. Not peer-reviewed at time of writing.
Press reporting
CNN, ABC News and Forbes coverage of the Guess advertisement in the August 2025 US issue of Vogue, including the small-print disclosure and the TikTok response. The placement was paid advertising; Vogue stated it has not used AI models in its own fashion production.
Forbes coverage of Coca-Cola’s AI-generated holiday advertising, November 2025, including statements from the company’s Global VP and Head of Generative AI.
NPR (14 February 2014) on the launch of the Getty Images Lean In Collection; Marketing Dive and MediaPost (March 2019) on Project #ShowUs.
Cited as a hypothesis, not as evidence
Lindstrom, M. (2008). Buyology. Doubleday. Trade nonfiction. Named in the essay as the popular account being tested, and corrected against the primary literature. The essay quotes only the title of chapter three, “I’ll Have What She’s Having: Mirror Neurons at Work”, verified in the publisher’s record at the Library of Congress. No sentence of the book’s text is quoted or paraphrased, because no verifiable copy of the passage could be obtained.
---
Diez años antes de que alguien pudiera generar un ser humano convincente, los departamentos de marketing ya estaban pidiendo algo muy parecido. Querían fotos de stock que no parecieran fotos de stock.
La industria de la imagen fabricó productos para ese brief. Getty Images lanzó la Lean In Collection con LeanIn.org en febrero de 2014, vendida con la promesa de imágenes auténticas de mujeres en lugar de los clichés que la categoría producía sin parar. En marzo de 2019 llegaron Dove, Getty y Girlgaze con Project #ShowUs: unas 5.000 fotografías sin distorsión, tomadas por fotógrafas, personas no binarias y personas que se identifican como mujeres, en 39 países. Los dos esfuerzos son serios y los dos responden al mismo pedido comercial. Véndenos una realidad que no tuvimos que salir a producir.
El mecanismo ya estaba descrito. Paul Frosh llamó a la fotografía de stock el papel tapiz de la cultura de consumo en Media, Culture & Society, en 2001. En 2023, Giorgia Aiello, Crispin Thurlow y Lara Portmann revisaron 600 fotos de stock de tres bancos de imágenes globales para Social Media + Society y encontraron la vida social aplanada y corporativizada camino al catálogo. Lo espontáneo era un género que se compraba mucho antes de ser un prompt.
La IA generativa cambió las cuentas. Una foto de stock con licencia se paga en dólares; una generada, en centavos. El Bureau of Labor Statistics de Estados Unidos ubica el salario mediano de un fotógrafo en 21,47 dólares la hora en su edición de mayo de 2025, y cuenta unos 145.000, menos de los que contaba en su edición anterior. La oficina no explica la caída, y yo tampoco la voy a explicar. El brief siguió igual. Lo que se desplomó fue el precio de cumplirlo.
La discusión en la que están todos
El miedo que traen los clientes a este tema es curiosamente específico. La imagen les importa poco. Lo que les preocupa es que los descubran.
Gallup encuestó a 3.270 adultos estadounidenses en mayo de 2026 para el informe Business in Society de Bentley University, con un panel probabilístico y un margen de error de 2,4 puntos. El 49% ve con malos ojos que las empresas usen IA en publicidad, contra un 19% que lo ve bien. Después la encuesta separa la pregunta por tarea, que es la parte que sirve y no el titular. El 75% acepta que la IA ayude a los empleados a generar ideas o a redactar, si se declara. El 53% acepta que la IA produzca el texto, la imagen o el video final. Y el 62% considera inaceptable usar IA para crear personas o voces en publicidad, incluso declarándolo.
Detente en “incluso declarándolo”, porque ahí el problema cambia de lugar. La objeción no desaparece cuando avisas.
Guess hizo esa prueba en público. Puso un anuncio con dos modelos generadas por IA, producidas por Seraphinne Vallora, en el número de agosto de 2025 de Vogue en Estados Unidos, con una línea de declaración en letra pequeña. Era publicidad pagada y no una pieza editorial, y Vogue aclaró después que no usa modelos de IA en su propia producción de moda. La reacción llegó igual, amplificada por una publicación de TikTok que pasó los 2,7 millones de reproducciones y cubierta por CNN, ABC News y Forbes.
La respuesta habitual es que nadie se da cuenta. Sophie Nightingale y Hany Farid publicaron en PNAS, en 2022, que los participantes separaban caras sintéticas de caras reales con un 48,2% de acierto, por debajo del azar, y que un grupo entrenado llegaba apenas al 59%. Esos mismos participantes calificaron las caras sintéticas como más confiables que las reales, 4,82 contra 4,48 en una escala de siete puntos, una diferencia que los autores llaman pequeña.
La corrección habitual es un estudio mucho más grande, el de Kamali y colegas en CHI 2025, con 749.828 juicios de 50.444 participantes: 76% de acierto en imágenes de IA y 74% en fotografías reales. Esas cifras describen un subconjunto difícil a propósito, porque los autores sacaron toda imagen acertada más del 80% de las veces, y con eso eliminaron el 14% de las observaciones. Antes de esa exclusión, el acierto por imagen iba del 32% al 99%. Nightingale y Farid trabajaron con caras recortadas, la categoría más difícil que hay. Los dos estudios dicen lo mismo: acertar depende de la imagen y no de quien mira, y la selección humana volvió las imágenes más difíciles de cazar. Elegir la mejor de cuarenta generaciones es lo que hace un equipo creativo sin pensarlo.
La ley se está partiendo en dos y la mayoría de los equipos de marketing no se ha enterado. El artículo 50 del AI Act europeo, que se aplica desde el 2 de agosto de 2026, obliga a quien despliegue un deepfake a declararlo con claridad, en un formato que una persona entienda sin herramientas técnicas. El artículo 3(60) lo define como contenido generado por IA que se parece a personas, objetos, lugares o hechos existentes y que “would falsely appear to a person to be authentic or truthful”. Léelo buscando la palabra intención. No aparece.
Estados Unidos no tiene una regla equivalente para imágenes. La norma de la FTC sobre reseñas y testimonios, 16 CFR Part 465, vigente desde octubre de 2024, prohíbe los testimonios que aparentan venir de alguien que no existe, incluidos los reseñadores generados por IA. Eso cubre a una persona inventada que te recomienda, no a una persona inventada parada en la imagen principal de tu home. Mientras investigaba esto no dejaba de encontrar artículos que citaban una guía de la FTC de marzo de 2025 que supuestamente obliga a declarar las imágenes de IA. No la encontré en ftc.gov.
El 23 de septiembre de 2026 se la volvimos a pedir a un buscador, y la produjo: tres principios básicos, la obligación de declarar las imágenes fotorrealistas de IA por realistas que se vean, incluso cartas de advertencia enviadas a siete marcas de moda. Nada de eso existe, y todas las fuentes detrás de la respuesta eran blogs de proveedores que citaban a otros blogs de proveedores. Una norma que nadie escribió ya tiene la apariencia de una norma, porque los intermediarios la repitieron y una máquina la devolvió sintetizada, con los modales de un hecho verificado. Nadie tuvo que mentir para que una industria empezara a cumplir una ley imaginaria.
Por qué nos dijeron que había que poner personas
Debajo de la discusión sobre declararlo hay un supuesto que casi nadie dice en voz alta. Ponemos personas en las páginas porque quien mira se ve a sí mismo en ellas.
La versión que circuló en marketing se lo atribuye a las neuronas espejo. El cerebro de quien mira simula a la persona de la foto, y de ahí sale la identificación. Martin Lindstrom puso esa idea a circular con Buyology, en 2008, cuyo tercer capítulo se titula, sin ninguna ambigüedad, “I’ll Have What She’s Having: Mirror Neurons at Work”. Es una idea razonable, bien contada, y le dio a toda una generación de publicistas un mecanismo al que podían ponerle nombre. La evidencia no la sostiene hasta ahí.
Las neuronas existen. Se encontraron en la corteza premotora de macacos en los años noventa y se registraron directamente en humanos una sola vez, con Mukamel y colegas en 2010. Lo que no sobrevivió fue la interpretación. En 2009 Gregory Hickok enumeró ocho problemas de la teoría de que las neuronas espejo explican la comprensión de la acción, y señaló que alterar esa región en monos daña el agarre, pero que nunca se demostró que dañe la percepción. Un año después Cecilia Heyes sostuvo que se explican mejor por aprendizaje asociativo común y corriente.
Es decir que el mecanismo que marketing tomó prestado está en discusión en la casa de donde lo sacó. La observación que hay debajo sigue en pie, y hay que medirla donde aterriza.
La atención es una cosa y la confianza es otra
Donde aterriza es en la atención, y el efecto es lo bastante preciso como para diseñar con él. Dos estudios de seguimiento ocular publicados en Frontiers in Psychology encontraron que una cara que mira hacia el producto rinde más que una cara que mira al espectador: Sajjacholapunt y Ball en 2014, para atención y recuerdo de marca, y Palcu, Sudkamp y Florack en 2017, que calcularon una probabilidad 4,1 veces mayor de mirar el producto cuando la mirada de la modelo apuntaba hacia él. Lee el resto de los resultados de 2017 antes de gastar dinero en el efecto. La dirección de la mirada movió poco la intención de compra, con p = 0.035, y no movió nada significativo en cuánto les gustó el producto ni en cuánto estaban dispuestos a pagar.
La confianza es más difícil, porque la respuesta depende de cómo la midas. Para 2003 los resultados publicados se contradecían, y Jens Riegelsberger, Angela Sasse y John McCarthy lo dijeron: un estudio anterior de ellos mismos había encontrado que las fotos del personal generaban reacciones negativas y menos confianza, así que armaron un experimento para zanjarlo. Agregaron fotos de vendedores a doce sitios de comercio electrónico, la mitad con buena reputación y la mitad con mala, y midieron la confianza de 115 participantes con un método de economía experimental que ponía en riesgo el dinero del propio participante.
No se movió nada. La presencia de la foto no tuvo efecto sobre la confianza, t(114) = -.01, p = .99. Tampoco lo tuvo qué tan confiable se veía la persona retratada, F(2,228) = .01, p = .99.
Otro método llegó a un lugar compatible. Dianne Cyr, Milena Head, Hector Larios y Bing Pan combinaron encuesta, entrevistas y seguimiento ocular en tres países para MIS Quarterly, en 2009. Las caras subieron el atractivo de la imagen y la sensación de presencia social, y las dos cosas predecían la confianza. El vínculo directo, el que va de las imágenes de personas a la confianza misma, fue lo único que no encontraron.
Las caras mueven la atención de manera confiable. La confianza declarada la mueven, como mucho, de forma indirecta. Cuando la medición le cuesta algo al participante, el efecto deja de aparecer.
En ese paper de 2003 hay un resultado más, en una sección que ningún resumen recoge, y es peor que un resultado nulo. Las fotos degradaron el juicio de la gente. El error de evaluación fue significativamente mayor en los sitios con foto que en los sitios sin foto, t(114) = -2.17, p = .03. Sin foto, los participantes calificaban distinto a los vendedores buenos y a los malos, 6,6 contra 6,0. Agregar la foto borró la diferencia, F(1,114) = 5.48, p = .02. En el resumen de los propios autores, las fotos subieron la confiabilidad percibida de los vendedores que funcionaban mal y bajaron la de los que tenían buena reputación.
Detente un segundo en eso, pensando en tu propia home. La cara no era un adorno inerte. Ayudó a los peores vendedores y perjudicó a los mejores, porque tapó la única señal que tenían los compradores.
La trampa de declararlo, y por dónde se sale
Eso devuelve el argumento a la persona inventada, y a un hallazgo de 2025 que parece apuntar al lado contrario.
Zeph van Berlo y Priska Breves, en Computers in Human Behavior Reports, en agosto de 2025, hicieron dos experimentos con 245 y 429 usuarias de Instagram en Estados Unidos sobre qué tan visible se declara la naturaleza artificial de un influencer virtual. La declaración prominente produjo menos credibilidad que la declaración sutil y menos que los influencers humanos, y eso se arrastró hasta la actitud hacia la marca y la intención de compra. Leído solo, ese resultado dice que hay que esconder la etiqueta.
El mismo resumen explica por qué no. La credibilidad de los influencers declarados con sutileza salió mixta frente a los humanos, no limpia. La versión discreta no gana: difumina. Y el resto de la evidencia cierra esa salida. El 62% de Gallup rechazó las personas creadas con IA con la declaración incluida, y Guess declaró en papel y lo castigaron igual. La declaración prominente cuesta credibilidad, la declaración discreta te cuesta cada vez que alguien se da cuenta, y la investigación sobre detección dice que no puedes predecir cuáles serán esas veces. No hay ninguna posición de esa perilla en la que una persona inventada funcione.
Hay un escenario donde una imagen humana sí trabaja, y apunta en la misma dirección. Deborah Small, George Loewenstein y Paul Slovic mostraron en 2007 que la gente dona bastante más a un individuo identificado que a víctimas estadísticas. Ese efecto le pertenece a una persona concreta que existe. Una cara generada es la forma de eso, sin nadie adentro.
Las cifras que nadie puede rastrear
El marketing repite un puñado de cifras en este tema. Que las imágenes auténticas convierten un 45% mejor que las de stock, atribuido a HubSpot. Que las fotos del equipo producen un puntaje de confianza un 65% más alto. Que las fotos de clientes subieron los registros un 35%. Traté de rastrear cada una hasta un método publicado, una muestra y una fecha. Ninguna resistió.
El 45% es el fracaso interesante, porque es medio verdadero. Hay un 45% en el blog de HubSpot. Pertenece a un caso donde el click-through hacia los casos de estudio mejoró casi en esa proporción después de un cambio de jerarquía visual y de formato, y no tiene nada que ver con fotos de personas. Una cifra real sobre una cosa, reatribuida a otra, repetida hasta volverse canónica.
Con el dinero pasa lo mismo. Ningún organismo del sector publica un referente de cuánto cuesta una sesión comercial; todos los rangos que circulan salen de las páginas de precios de los propios fotógrafos. Tampoco hay evidencia pública de primer nivel de que las imágenes de IA hayan movido las ventas en una dirección o en la otra. Coca-Cola, con dos años de publicidad navideña generada con IA encima, dice que su campaña de 2024 funcionó bien en pruebas privadas, y sin método publicado eso es una afirmación y no un resultado.
El vacío corta para los dos lados. Durante años el argumento a favor de la fotografía humana se apoyó en cifras que nadie puede comprobar, que es exactamente la posición del cliente que quiere una modelo de IA que nadie note.
Todas tienen la misma cara
Hay un segundo costo, y para un equipo de marketing debería ser el que decide la discusión, porque no tiene nada que ver con la ética.
Los modelos generativos no producen una persona cualquiera. Producen algo más cercano a la persona promedio, y eso ya está medido. Dunn y colegas, en el British Journal of Psychology, en febrero de 2026, usaron redes neuronales profundas optimizadas para identidad facial y hallaron las caras de IA distribuidas más al centro del face-space, el espacio donde cada cara se ubica según su distancia del rostro promedio, que las reales. La mitad humana de su estudio es la parte que vale leer dos veces. Treinta y seis super-reconocedores detectaron las caras de IA un 15% mejor que una muestra típica, y un 7% mejor que un grupo de control de 89 personas motivadas y de alto rendimiento, con su ventaja atada a la sensibilidad a ese aire hiperpromedio. Antes de archivar esto como confirmación independiente, conviene saber que cuatro de sus seis autores firman también el paper de Psychological Science de 2023 que bautizó el hiperrealismo de la IA. Es el mismo equipo llevando más lejos su propia línea de trabajo.
Lo que se ve en la conducta va en la misma dirección. Alice Mado Proverbio y Mariia Dosaikina, en Scientific Reports, en 2026, enfrentaron 220 fotografías reales con 220 caras de StyleGAN2. En dos grupos separados de 25 evaluadores, las caras de IA puntuaron más alto en atractivo, 3,12 contra 2,31, y en familiaridad, 2,70 contra 2,24, mientras que el segundo grupo las identificó como artificiales apenas el 33% de las veces, contra el 62% de las reales. La conclusión de los propios autores es una disociación y no una victoria limpia de las máquinas: la conducta no cazó esas caras y la señal de EEG sí. Hay un límite que ellos mismos señalan y que aquí importa. Todas las caras que probaron eran blancas. Miller y colegas ya habían encontrado en 2023 que el efecto de hiperrealismo no aparecía en caras de personas racializadas. El promedio que producen estos modelos es el promedio de sus datos de entrenamiento.
Nightingale y Farid habían buscado la misma explicación ya en 2022. Su frase en la discusión es que las caras sintéticas quizá se califican como más confiables “because synthesized faces tend to look more like average faces which themselves are deemed more trustworthy”. La referencia en la que apoyan eso es Sofer, Dotsch, Wigboldus y Todorov, en Psychological Science, en 2015, bajo un título que le ahorra tiempo a todo el mundo: “What is typical is good.” El eslabón que va de la tipicidad a la confianza ya estaba revisado por pares. Lo que llegó después fue la medición de que estas caras son típicas.
Ahora pon eso al lado de algo que la psicología sabe desde 1990, cuando Judith Langlois y Lori Roggman mostraron en Psychological Science que las caras compuestas, promediadas matemáticamente, se juzgaban más atractivas que casi todas las caras individuales que las formaban. El promedio es la razón por la que estas imágenes funcionan bien en una sala de juntas. Es también la razón por la que no se le quedan a nadie.
En 1979 Leah Light, Fortunee Kayra-Stuart y Steven Hollander hicieron cuatro estudios en el Journal of Experimental Psychology: Human Learning and Memory y encontraron que la memoria de reconocimiento era peor para las caras prototípicas que para las inusuales, y que eso se mantenía con aprendizaje incidental e intencional, con presentaciones de 3 a 15 segundos y con intervalos de retención de 3 a 24 horas. Un quinto experimento identificó el motivo: la similitud entre ítems. Dos años después el mismo grupo publicó el hallazgo con un título más directo. A las personas atractivas cuesta más recordarlas. El face-space de Tim Valentine, en 1991, le dio un marco al patrón. Las caras se codifican por la distancia a la que están de la norma, y las que están cerca de la norma son las que se confunden entre sí.
El mecanismo se sigue discutiendo, y hay que decirlo. Morris y Wickham lo reexaminaron en 2001 y concluyeron que la tipicidad, el atractivo, la mera exposición y las creencias de cada persona sobre su propia memoria están enredadas de maneras que la teoría anterior no separaba. Lo que discuten no es la dirección del efecto.
La imagen tampoco está sola. Raymond Burke y Thomas Srull mostraron en el Journal of Consumer Research, en 1988, que la memoria de la publicidad de una marca se inhibía con la exposición posterior a publicidad de marcas competidoras de la misma categoría y, con menos comodidad para cualquiera que tenga una línea de productos, con publicidad de otros productos del mismo fabricante. Un tercer experimento encontró que la presencia de publicidad competidora cambia la relación entre repetición y memoria. Y el parecido entre caras está documentado, no solo percibido. AlDahoul, Rahwan y Zaki mapearon los sesgos de Stable Diffusion en seis razas, dos géneros, 32 profesiones y ocho atributos, y después hicieron un análisis aparte de hasta qué punto el modelo hace que las personas de una misma raza se parezcan entre sí. Ese análisis encontró una homogeneización racial significativa, con casi todos los hombres de Medio Oriente representados con barba, piel morena y vestimenta tradicional.
Así que la cara de la página es más promedio que una real, las caras promedio cuestan más de recordar, y además compite con todas las otras páginas construidas con el mismo puñado de modelos.
Si no hay presupuesto
Supongamos que no hay dinero para una sesión. IA, stock, o nada.
El hallazgo más fuerte de toda esta área termina no siendo sobre personas. Shunyuan Zhang, Dokyun Lee, Param Vir Singh y Kannan Srinivasan, en Management Science, en 2022, hicieron un análisis de diferencias en diferencias sobre 7.423 propiedades de Airbnb a lo largo de dieciséis meses y encontraron que los anuncios con fotos verificadas, tomadas por los fotógrafos de la plataforma y no por el anfitrión, tenían un 8,98% más de ocupación. Eso es fotografía de la cosa que se vende, hecha con oficio, y es el único resultado causal, con muestra grande y revisión por pares, que hay en esta área. La advertencia sobre la versión barata viene de un estudio de Cornell sobre mercados entre particulares, que ubicó la confianza más alta en las buenas fotos hechas por el propio usuario, después en las de stock y al final en las malas. Hacerlo tú mismo ayuda solo si el resultado es bueno.
Lo que la evidencia sostiene es poco y es utilizable. Pon el presupuesto en fotografiar la cosa concreta que vendes: el cuarto, el producto, el trabajo, la gente que lo hace. Pon el resto en que esa fotografía esté bien hecha, y no en conseguir a un desconocido. Si el dinero no alcanza para una buena fotografía de una persona, lo defendible es una buena fotografía de otra cosa, o ninguna.
No te puedo decir que la página sin persona rinda más, porque nadie ha publicado una comparación de primer nivel entre páginas con y sin imágenes de personas, ni entre ilustración y fotografía. Tampoco nadie ha probado la cadena completa de este ensayo, de punta a punta, en un sitio real. Las piezas se midieron por separado, y apuntan al mismo lado.
El cliente pide una persona para que el visitante se vea en ella. El mecanismo que le vendieron para eso no sobrevive a la literatura. Lo que una cara hace de manera confiable es atraer la mirada, y eso vale la pena. La cara inventada está construida más cerca del promedio que una real, que es la razón por la que se ve bien en la sala de juntas y la razón por la que después cuesta recordarla. En el único experimento que midió la confianza con dinero de por medio, poner una cara en la página hizo que los vendedores malos se vieran mejor y los buenos, peor. Y después sale a vivir en una página que compite con todas las otras páginas hechas con los mismos modelos, en un entorno donde ya se demostró que el material parecido interfiere con el recuerdo de todo el conjunto.
Leído como una compra, el cliente pide identificación y recibe un compuesto. Paga por una cara porque las caras se notan, y compra la más olvidable que hay disponible.
Cinco preguntas antes de aprobar la imagen
Cinco preguntas para meter en un brief, o para hacer en voz alta antes de que alguien apruebe una cara.
¿Qué queremos que sienta el visitante con esta persona, y quién decidió que una persona era la manera de lograrlo? Si la respuesta es que las personas generan confianza, el experimento que midió la confianza con el dinero del participante en juego no encontró ninguna.
Si ponemos “generada con IA” al lado de esta imagen, ¿el argumento para usarla sigue en pie? El 62% de los adultos estadounidenses considera inaceptables las personas creadas con IA en publicidad aun con la declaración puesta, así que la etiqueta pone a prueba la decisión en lugar de arreglarla.
¿Estamos comprando atención o confianza? Una cara mueve la primera de manera confiable, y la dirección de la mirada la mueve con precisión. En el único estudio que puso dinero detrás de la segunda, la foto no movió nada, e hizo que los vendedores malos se vieran mejor y los buenos, peor.
¿Alguien reconocería esta cara si se la volviera a encontrar la semana que viene? Las caras sintéticas están más cerca del promedio que las reales, las caras cercanas al promedio son las que se confunden entre sí, y esta va a llegar al lado de todas las demás páginas hechas con los mismos modelos.
¿Qué compraría este presupuesto si fotografiáramos la cosa que vendemos? Es la única pregunta de la lista que tiene detrás una respuesta causal, con muestra grande y revisión por pares.
Nada de esto es un argumento contra poner personas. Fotografía a las tuyas, haciendo su trabajo, y la mayoría de los hallazgos de este ensayo dejan de aplicarte. El problema empieza donde se compra una cara para que represente una relación que todavía no existe.
Debajo de las cinco hay una pregunta que un brief casi nunca hace, y es la más barata de contestar: ¿qué está haciendo la persona de esa foto, y una fotografía de lo que vendes lo haría mejor?
Fuentes
La bibliografía no se traduce: se reproduce igual que en la versión en inglés, con los títulos, las revistas y las notas de verificación en su idioma original.
Investigación revisada por pares
Aiello, G., Thurlow, C. & Portmann, L. (2023). “Desocializing Social Media: The Visual and Media Ideologies of Stock Photography.” Social Media + Society, 9(1). DOI 10.1177/20563051231156363. Social-semiotic analysis of 600 stock photographs from three image banks.
Berlo, Z. van & Breves, P. (2025). “Disclosing the virtual nature of virtual influencers: The effect of disclosure prominence and the role of product digitality.” Computers in Human Behavior Reports, 19, 100742. DOI 10.1016/j.chbr.2025.100742. Open access, CC BY. Two online experiments, N=245 and N=429, US female Instagram users aged 18–34. Official abstract verified verbatim.
Cyr, D., Head, M., Larios, H. & Pan, B. (2009). “Exploring Human Images in Website Design: A Multi-Method Approach.” MIS Quarterly, 33(3). Survey, interviews and eye-tracking in Canada, Germany and Japan. Page range not given here: the MISQ record shows 539–566 and the AIS eLibrary record 530–566, and the discrepancy is unresolved.
Frosh, P. (2001). “Inside the image factory: stock photography and cultural production.” Media, Culture & Society, 23(5), 625–646. DOI 10.1177/016344301023005005. And Frosh, P. (2003). The Image Factory: Consumer Culture, Photography and the Visual Content Industry. Berg.
Gallese, V., Fadiga, L., Fogassi, L. & Rizzolatti, G. (1996). Brain, 119(2); di Pellegrino, G., Fadiga, L., Fogassi, L., Gallese, V. & Rizzolatti, G. (1992). Experimental Brain Research, 91(1).
Heyes, C. (2010). “Where do mirror neurons come from?” Neuroscience & Biobehavioral Reviews, 34(4), 575–583.
Hickok, G. (2009). “Eight Problems for the Mirror Neuron Theory of Action Understanding in Monkeys and Humans.” Journal of Cognitive Neuroscience, 21(7), 1229–1243.
Kamali et al. (2025). “Characterizing Photorealism and Artifacts in Diffusion Model-Generated Images.” Proceedings of CHI 2025, ACM. DOI 10.1145/3706598.3713962. 749,828 observations from 50,444 participants across 450 AI-generated and 149 real images. The 76% and 74% figures follow an exclusion of every image identified correctly in more than 80% of observations, which removed 14% of all observations; the 32–99% range is from before that exclusion.
Mukamel, R., Ekstrom, A. D., Kaplan, J., Iacoboni, M. & Fried, I. (2010). “Single-Neuron Responses in Humans during Execution and Observation of Actions.” Current Biology, 20(8), 750–756. DOI 10.1016/j.cub.2010.02.045.
Nightingale, S. J. & Farid, H. (2022). “AI-synthesized faces are indistinguishable from real faces and more trustworthy.” PNAS, 119(8), e2120481119. DOI 10.1073/pnas.2120481119. Full text verified in PubMed Central. The trustworthiness ratings are 4.82 against 4.48 on a seven-point scale, d = 0.49, described by the authors as a small effect.
Palcu, J., Sudkamp, J. & Florack, A. (2017). “Judgments at Gaze Value: Gaze Cuing in Banner Advertisements, Its Effect on Attention Allocation and Product Judgments.” Frontiers in Psychology, 8, 881. N=137, eye-tracking.
Riegelsberger, J., Sasse, M. A. & McCarthy, J. D. (2003). “Shiny happy people building trust? Photos on e-commerce websites and consumer trust.” Proceedings of CHI 2003, 121–128. DOI 10.1145/642611.642634. N=115, twelve sites, trust measured with financial risk. Full text read from the author’s copy at University College London. Null results: photo presence t(114) = -.01, p = .99; depicted trustworthiness F(2,228) = .01, p = .99. Assessment error with photographs t(114) = -2.17, p = .03; interaction F(1,114) = 5.48, p = .02.
Sajjacholapunt, P. & Ball, L. J. (2014). “The influence of banner advertisements on attention and memory: human faces with averted gaze can enhance advertising effectiveness.” Frontiers in Psychology, 5, 166.
Small, D. A., Loewenstein, G. & Slovic, P. (2007). “Sympathy and callousness: The impact of deliberative thought on donations to identifiable and statistical victims.” Organizational Behavior and Human Decision Processes, 102(2), 143–153. Replicated and extended by Maier et al. (2023), Collabra: Psychology, 9(1), 90203.
AlDahoul, Nouar, Rahwan, Talal & Zaki, Yasir (2025). “AI-generated faces influence gender stereotypes and racial homogenization.” Scientific Reports, 15(1), 14449. DOI 10.1038/s41598-025-99623-3. Open access. The bias analysis covers six races, two genders, 32 professions and eight attributes; the racial homogenization analysis is a separate measurement within the same paper.
Miller, E. J., Steward, B. A., Witkower, Z., Sutherland, C. A. M., Krumhuber, E. G. & Dawel, A. (2023). “AI Hyperrealism: Why AI Faces Are Perceived as More Real Than Human Ones.” Psychological Science, 34(12), 1390–1403. DOI 10.1177/09567976231207095. Cited in the essay for the finding that the hyperrealism effect did not appear for faces of people of colour.
Sofer, C., Dotsch, R., Wigboldus, D. H. J. & Todorov, A. (2015). “What Is Typical Is Good: The Influence of Face Typicality on Perceived Trustworthiness.” Psychological Science, 26(1), 39–47. The primary source behind the typicality-to-trust link, cited by Nightingale and Farid as their reference 12.
Burke, Raymond R. & Srull, T. K. (1988). “Competitive Interference and Consumer Memory for Advertising.” Journal of Consumer Research, 15(1), 55–68. Three experiments. Abstract-level verification; the full text is paywalled, so the essay reports only what the abstract states, including that Experiment 3 found competitive advertising changes the relationship between repetition and memory, without a direction.
Dunn, James D., White, David, Sutherland, Clare A. M., Miller, Elizabeth J., Steward, Ben A. & Dawel, Amy (2026). “Too good to be true: Synthetic AI faces are more average than real faces and super-recognizers know it.” British Journal of Psychology, article bjop.70063. DOI 10.1111/bjop.70063. Published 18 February 2026. Abstract-level verification; no open copy of the full text. Four of the six authors also wrote Miller et al. (2023), which the essay states.
Langlois, J. H. & Roggman, L. A. (1990). “Attractive Faces Are Only Average.” Psychological Science, 1(2), 115–121. DOI 10.1111/j.1467-9280.1990.tb00079.x
Light, L. L., Kayra-Stuart, F. & Hollander, S. (1979). “Recognition memory for typical and unusual faces.” Journal of Experimental Psychology: Human Learning and Memory, 5(3), 212–228. DOI 10.1037/0278-7393.5.3.212. Five experiments.
Light, L. L., Hollander, S. & Kayra-Stuart, F. (1981). “Why Attractive People are Harder to Remember.” Personality and Social Psychology Bulletin, 7(2), 269–276. Citation verified; no abstract available in public databases, so nothing beyond the title is verified.
Morris, P. E. & Wickham, L. H. V. (2001). “Typicality and face recognition: a critical re-evaluation of the two factor theory.” Quarterly Journal of Experimental Psychology A, 54(3), 863–877. Included because it complicates the mechanism behind the typicality effect.
Proverbio, Alice Mado & Dosaikina, Mariia (2026). “Neural signatures of hyper-realistic AI-generated faces: dissociating behavioral indistinguishability from implicit neural evaluation.” Scientific Reports, 16(1), 22944. DOI 10.1038/s41598-026-59487-7. Full text read. EEG over 440 faces, plus behavioural validation in two separate groups of 25 raters, carried over from the team’s earlier work. All stimulus faces were of Caucasian ethnicity, which the authors list as a limitation. The paper’s conclusion is a dissociation: behaviour failed to detect the faces while the neural response did not.
Valentine, T. (1991). “A unified account of the effects of distinctiveness, inversion, and race in face recognition.” Quarterly Journal of Experimental Psychology A, 43(2), 161–204. DOI 10.1080/14640749108400966. Five experiments.
Zhang, S., Lee, D., Singh, P. V. & Srinivasan, K. (2022). “What Makes a Good Image? Airbnb Demand Analytics Leveraging Interpretable Image Features.” Management Science, 68(8), 5644–5666. Deep learning and difference-in-differences across 7,423 properties over sixteen months; verified photographs, taken by the platform rather than the host, gave 8.98% higher occupancy. The figures of 17.51%, $2,521 and 58.83% that circulate with this citation belong to the 2017 working paper (7,711 properties) and are not in the published article.
Estadística oficial y regulación
Bentley University–Gallup, Business in Society survey, 2026. Probability-based Gallup Panel, n=3,270 US adults, fielded 4–11 May 2026, margin of error ±2.4 points at 50%, weighted to national demographics.
European Commission, Transparency obligations under Article 50 of the AI Act, and Regulation (EU) 2024/1689, Articles 3(60), 50 and 113. Applying from 2 August 2026. No grace period to December 2026 exists in the Regulation; the transitional regime for systems already on the market is in Article 111 and runs on different dates.
US Bureau of Labor Statistics, Occupational Outlook Handbook, Photographers (SOC 27-4021), May 2025 edition. Median wage $21.47 an hour, $44,660 a year; about 145,000 jobs in 2025, against 151,200 in 2024; 67% self-employed.
US Federal Trade Commission, final rule on the use of consumer reviews and testimonials, 16 CFR Part 465, announced 14 August 2024, effective 21 October 2024.
Preprint, etiquetado como tal
“Understanding Image Quality and Trust in Peer-to-Peer Marketplaces,” Cornell, arXiv:1811.10648. Not peer-reviewed at time of writing.
Cobertura de prensa
CNN, ABC News and Forbes coverage of the Guess advertisement in the August 2025 US issue of Vogue, including the small-print disclosure and the TikTok response. The placement was paid advertising; Vogue stated it has not used AI models in its own fashion production.
Forbes coverage of Coca-Cola’s AI-generated holiday advertising, November 2025, including statements from the company’s Global VP and Head of Generative AI.
NPR (14 February 2014) on the launch of the Getty Images Lean In Collection; Marketing Dive and MediaPost (March 2019) on Project #ShowUs.
Citado como hipótesis, no como evidencia
Lindstrom, M. (2008). Buyology. Doubleday. Trade nonfiction. Named in the essay as the popular account being tested, and corrected against the primary literature. The essay quotes only the title of chapter three, “I’ll Have What She’s Having: Mirror Neurons at Work”, verified in the publisher’s record at the Library of Congress. No sentence of the book’s text is quoted or paraphrased, because no verifiable copy of the passage could be obtained.
---