experiment · Social Media + Society · la publicación, 1 oct 2026 · gratis
En un estudio danés de 2024, la gente apenas distinguió las imágenes hechas con IA de las fotos reales, y le habría ido mejor si hubiera respondido 50% en todas.
Versión breve · la versión detallada sigue, unos 5 min
Pregunte al chat de weeklyAI
Pregúnteme sobre la IA en su vida: le señalo las notas que weeklyAI ha publicado sobre el tema, esta semana y antes.
Las conversaciones se guardan mientras exista weeklyAI, para mejorar la publicación. Se responde en el idioma en que usted escribe.
- El estudio, de un vistazo
- Quiénes
- Adultos daneses
- Cuántos
- 2,091 adultos
- Dónde
- Dinamarca
- Cuándo
- junio de 2024
- Tipo de estudio
- experiment
- Quién lo hizo
- Universidades de Ámsterdam, Copenhague, Aarhus, Dinamarca del Sur y LSE
- El límite que importa
- Participantes sabían que había imágenes falsas y las buscaban: rindieron mejor que alguien hojeando el teléfono.
El grupo estudiado frente a dos formas de contestar sin mirar
Quedaron apenas por encima del azar
Les habría ido mejor escribiendo 50 en todas
Fue el único grupo que quedó claramente por debajo de esa marca
En junio de 2024, 2,091 adultos de Dinamarca vieron siete imágenes cada uno, tomadas al azar de un conjunto de 52, y calificaron de 0 a 100 qué tan probable era que fueran generadas con IA. Cada evento tenía una foto real y una versión hecha con una herramienta gratuita. En promedio, los participantes acertaron un poco más que si hubieran adivinado al azar. Pero en promedio les habría ido mejor si hubieran respondido 50% en todas las imágenes. Solo los jóvenes de 18 a 24 años obtuvieron un puntaje mejor que esa marca.
El estudio no nombra ninguna forma de que usted compruebe esto en casa.
Los autores señalan que el estudio no probó ninguna norma ni capacitación. Recomiendan exigir marcas de agua a las herramientas de IA y que las plataformas de redes sociales vigilen con más cuidado el contenido generado.
Qué significa para usted
En Dinamarca, adultos que miraron imágenes en una pantalla apenas distinguieron las hechas con inteligencia artificial de las fotos reales. En su propia elección, ese estudio no prueba nada sobre su registro, su campaña ni su conteo, y ningún país fuera de Dinamarca fue examinado. Antes de votar, revise qué auditó y publicó su autoridad electoral.
Hove, M. F., Hjorth, F., Golovchenko, Y. et al. (2026). Seeing Is Believing? Most People Struggle to Detect AI-Generated Images. Social Media + Society. https://doi.org/10.1177/20563051261483376
Quién pagó: Los autores no recibieron apoyo financiero para la investigación, autoría o publicación de este artículo; la recolección de datos se realizó como parte de una subvención de la Agencia para la Cultura y los Palacios y la Agencia para el Gobierno Digital de Dinamarca, y el artículo no indica que los financiadores tuvieran algún papel.
Versión detalladaLos pasajes copiados del artículo, las ilustraciones y cada fuente con cuánto leímos de ella · 1096 palabras · unos 5 minLeerla →Cerrar
La regla y la evidencia: un estudio danés de 2024 halló que la gente apenas distingue las imágenes hechas con IA de las fotos reales, y que en promedio le iría mejor contestar 50 en cada imagen; el resultado no se puede extender a otros países ni a imágenes de mejor calidad
### EL ACTO
El estudio halló que la gente apenas distingue las imágenes hechas con IA de las fotos reales, y que en promedio le iría mejor contestar 50 en cada imagen; por eso propone medidas que podrían adoptarse: exigir marcas de agua a las herramientas de inteligencia artificial, fijar reglas de transparencia y pedir a las redes sociales que vigilen mejor lo que publican1.
### EL CASO QUE RESPONDE
No se encontró esta semana ningún caso real con nombre, lugar, fecha y segunda fuente que se pueda poner junto a este estudio.
### LO QUE SE MIDIÓ
Un grupo de 2,091 adultos daneses, representativo de la población general en edad, género y región, vio siete imágenes cada uno, tomadas al azar de un conjunto de 52: 26 eventos, cada uno con una versión auténtica y una versión hecha con inteligencia artificial2. La inteligencia artificial baja mucho el costo de crear imágenes realistas y permite que más actores las hagan y las difundan, por ejemplo en redes sociales3. Después de cada imagen, la persona movía una escala de 0 a 100 para decir qué tan probable era que la imagen hubiera sido creada por una máquina. La mediana de tiempo frente a cada imagen fue de siete segundos. Se les avisó de antemano que algunas eran falsas y se les pidió no examinarlas mucho.
El resultado, en palabras llanas: el grupo quedó apenas por encima de quien contesta al azar, pero por debajo de quien hubiera escrito 50 en cada imagen y no se hubiera molestado en mirar. La diferencia frente al azar fue de 0.05 en la escala de error, con un margen de [−0.057, −0.042] y p < .001; la diferencia frente al 50 fijo fue de 0.034 en contra, con un margen de [0.028, 0.039] y p < .014. Dicho de otra forma: en promedio, la persona que adivinaba sin mirar le iba mejor que la que miraba.
Solo un grupo quedó claramente por debajo del 50 fijo: los jóvenes de 18 a 24 años5. Ni las personas muy analíticas, ni quienes ya usan estas herramientas, ni las mujeres lograron ese umbral. Y el tipo de imagen no cambió nada: ni las crisis, ni los políticos, ni la gente conocida resultaron más fáciles o más difíciles que una vaca en un potrero6.
El estudio mismo pone su propio límite sobre la mesa: como los participantes sabían que había imágenes falsas y las estaban buscando, rindieron mejor de lo que rendiría alguien hojeando el teléfono en el transporte7. Es decir, el resultado es un techo, no un piso. Y las imágenes usadas eran de calidad más bien baja, hechas con una herramienta gratuita y pública, no las de un operativo sofisticado.
### LA RESPUESTA DE LA OTRA REGIÓN AL MISMO PROBLEMA
No se encontró esta semana ningún fallo, ley ni autoridad de otro país que trate el mismo asunto de distinguir imágenes falsas de verdaderas.
### LA LECTURA DE LA BIBLIOTECA
Así lo lee weeklyAI. Lo que la propaganda siempre ha hecho es tomar algo verdadero y darle la vuelta: una foto real de otro año, de otro lugar, con otro pie. Lo nuevo no es la mentira: es que ahora la mentira se puede fabricar desde cero en minutos y sin fotógrafo. Lo que el público aprende a revisar son las costuras: el dedo de más, la sombra rara, la letra torcida. Lo que no puede aprender a revisar es lo que ya no tiene costura. Cuando la imagen viene limpia, la única defensa que queda es preguntar de dónde salió, y eso no se hace en siete segundos de scroll.
Esa costura desaparecida tiene una consecuencia para la casa del lector: si el ojo ya no alcanza, la confianza tiene que mudarse a otro lado. No al ojo, sino al origen. ¿Quién publicó primero? ¿En qué fecha? ¿Hay una segunda fuente que lo confirme? weeklyAI espera que en los próximos meses veamos cada vez más fotos de crisis y de candidatos que parecen impecables, y que la reacción más común no sea verificar sino encogerse de hombros. Esta lectura se equivocaría si resultara que quienes más dudan de todo identifican mejor las falsas que quienes dudan menos, o si el escepticismo general viniera acompañado de mayor precisión y no solo de rechazo en bloque.
Otros trabajos advierten algo incómodo: un estudio sobre imágenes manipuladas con inteligencia artificial encontró que la gente suele poder identificarlas, al revés de lo que reporta el hallazgo principal. Es una advertencia sobre llevar el resultado de un país a otro, no sobre lo que aquí se midió.
### LO QUE NO SE SABE
El estudio midió si la gente distingue, no si ver una imagen falsa cambia el voto, la creencia o la intención de salir a votar. Nadie encuestó a nadie después de una elección: todo pasó en una pantalla de encuesta, en Dinamarca, en junio de 2024.
No se sabe si estos resultados se sostendrían en México, en Brasil, en Colombia o en Estados Unidos. El propio estudio señala que Dinamarca es un caso poco probable para encontrar personas que no distinguen, por su alto nivel de alfabetización digital, y que si ahí fallan, en la mayoría de los países probablemente fallarán igual o peor.
No se sabe qué pasaría con imágenes hechas por modelos más avanzados, porque solo se usó uno. Tampoco se sabe si las campañas de concientización ayudan o estorban: el estudio advierte que decirle a la gente "cuidado con las falsas" puede volverla escéptica de todo, incluso de lo verdadero, y abrir la puerta para que un político desestime una foto real diciendo que es inventada8.
Aquí hay una tensión que no conviene esconder: el mismo esfuerzo que sirve para que la gente no se trague una mentira puede servir para que no crea una verdad. No hay manera de salir de eso con puro ánimo individual.
Usted va a votar un día fijo, y esa mañana va a ver fotos. La pregunta útil no es "¿podré notar si es falsa?", porque la evidencia dice que probablemente no. La pregunta útil es: "¿quién me confirma que esta foto es de este lugar y de este año?". Pregúnteselo a su autoridad electoral o a un reportero local antes de compartirla.
De dónde sale cada dato de contexto, y cuánto leímos de cada documento
- Hove, M. F., Hjorth, F., Golovchenko, Y. et al. (2026). Seeing Is Believing? Most People Struggle to Detect AI-Generated Images. Social Media + Society. https://doi.org/10.1177/20563051261483376 - el artículo del que trata esta nota — el artículo completo — el pasaje: "Policy interventions could require generative AI tools to include watermarks, enforce transparency standards, and impose stricter monitoring obligations on social media platforms. Achieving this will require substantial investment in AI detection tools and verification methods."
- Hove, M. F., Hjorth, F., Golovchenko, Y. et al. (2026). Seeing Is Believing? Most People Struggle to Detect AI-Generated Images. Social Media + Society. https://doi.org/10.1177/20563051261483376 - el artículo del que trata esta nota — el artículo completo — el pasaje: "In a preregistered survey experiment, we find that, on average, people are not substantially better than chance at discerning AI-generated images from authentic images, regardless of the context of the image."
- Hove, M. F., Hjorth, F., Golovchenko, Y. et al. (2026). Seeing Is Believing? Most People Struggle to Detect AI-Generated Images. Social Media + Society. https://doi.org/10.1177/20563051261483376 - el artículo del que trata esta nota — el artículo completo — el pasaje: "Generative artificial intelligence (AI) dramatically reduces the costs of creating high-quality, realistic images, allowing a wider range of actors to create and distribute such content, for instance, on social media."
- Hove, M. F., Hjorth, F., Golovchenko, Y. et al. (2026). Seeing Is Believing? Most People Struggle to Detect AI-Generated Images. Social Media + Society. https://doi.org/10.1177/20563051261483376 - el artículo del que trata esta nota — el artículo completo — el pasaje: "However, the difference in means between individuals and the consistent 50% probability evaluation is 0.034 and statistically significant (p < .01, 95% CI = [0.028, 0.039]). This means that while individuals are better than random evaluations, they would, on average, have been better off by taking the 50% probability evaluation."
- Hove, M. F., Hjorth, F., Golovchenko, Y. et al. (2026). Seeing Is Believing? Most People Struggle to Detect AI-Generated Images. Social Media + Society. https://doi.org/10.1177/20563051261483376 - el artículo del que trata esta nota — el artículo completo — el pasaje: "While younger people, highly analytical-thinking individuals, and experienced users of generative AI have better discernment abilities, no demographic group other than young individuals (<25) performs substantially better than chance."
- Hove, M. F., Hjorth, F., Golovchenko, Y. et al. (2026). Seeing Is Believing? Most People Struggle to Detect AI-Generated Images. Social Media + Society. https://doi.org/10.1177/20563051261483376 - el artículo del que trata esta nota — el artículo completo — el pasaje: "This indicates that individuals’ discernment abilities are not different for different types of images, and for three out of four categories, not in any way that would change the interpretation of H1."
- Hove, M. F., Hjorth, F., Golovchenko, Y. et al. (2026). Seeing Is Believing? Most People Struggle to Detect AI-Generated Images. Social Media + Society. https://doi.org/10.1177/20563051261483376 - el artículo del que trata esta nota — el artículo completo — el pasaje: "However, this concern primarily makes the results more conservative as respondents are expected to perform better in the experiment than they would on social media."
- Hove, M. F., Hjorth, F., Golovchenko, Y. et al. (2026). Seeing Is Believing? Most People Struggle to Detect AI-Generated Images. Social Media + Society. https://doi.org/10.1177/20563051261483376 - el artículo del que trata esta nota — el artículo completo — el pasaje: "This occurs when dishonest actors dismiss genuine, compromising content as AI-generated to evade accountability (Schiff et al., 2025)."
Hove, M. F., Hjorth, F., Golovchenko, Y. et al. (2026). Seeing Is Believing? Most People Struggle to Detect AI-Generated Images. Social Media + Society. https://doi.org/10.1177/20563051261483376
Quién pagó: Los autores no recibieron apoyo financiero para la investigación, autoría o publicación de este artículo; la recolección de datos se realizó como parte de una subvención de la Agencia para la Cultura y los Palacios y la Agencia para el Gobierno Digital de Dinamarca, y el artículo no indica que los financiadores tuvieran algún papel.
Los hallazgos de otros estudios que aquí se mencionan provienen de este documento; esos estudios no se consultaron uno por uno.
experiment · Social Media + Society · the paper, 1 Oct 2026 · free
Most people shown AI-made images in a Danish test could barely tell them from real photos
Short version · the longer version follows, about 6 min
Ask the weeklyAI chat
Ask me about AI in your life: I point you to the stories weeklyAI has published on it, this week and before.
Conversations are saved for as long as weeklyAI exists, to improve the publication. Answers come in the language you write in.
- The study at a glance
- Who
- adult Danes
- How many
- 2,091
- Where
- Denmark
- When
- June 17 to June 24, 2024
- Kind of study
- experiment
- Who did it
- University of Amsterdam, University of Copenhagen, Aarhus University, University of Southern Denmark, London School of Economics
- The limit that matters
- Images were low quality, made with a free tool, and respondents were warned to look for AI.
Average participants against the simplest strategy: answering 50 for every image
People did worse; the gap was 0.034
In plain words: a person who shrugged at every picture would have scored better than the average participant who tried.
In June 2024, researchers showed 2,091 adult Danes seven images each, drawn at random from 52 pictures of real events. Half the pictures were made with a free AI tool. People rated how likely each image was AI-made. Their scores were barely above random guessing, and worse than simply answering 50% every time. Only those aged 18 to 24 did better than that. The images were low quality and made with a free tool, so real AI pictures may be harder to spot. Participants were also told in advance that some images were AI and were looking for them, which makes the test easier than scrolling social media.
The article names no way for a voter to check an image's origin tonight.
The article names no rule from any electoral authority.
What this means for you
For your own feed before an election, that Danish result offers one plain lesson: a picture that looks real tells you little about whether it is. The study cannot say whether any training or check would help, and it did not test pictures from other countries, so watch for a label or a source before you share.
Hove, M. F., Hjorth, F., Golovchenko, Y. et al. (2026). Seeing Is Believing? Most People Struggle to Detect AI-Generated Images. Social Media + Society. https://doi.org/10.1177/20563051261483376
Who paid: The authors received no financial support for the research, authorship, and/or publication of this article; data collection was conducted as part of a grant from the Agency for Culture and Palaces and the Agency for Digital Government in Denmark, and the article does not say the funders had any role.
The longer versionThe passages copied from the paper, the pictures, and every source with how much of it we read · 1118 words · about 6 minRead it →Close
The rule and the evidence: Most people shown AI-made images in a Danish test could barely tell them from real photos
### THE ACT, quoted exactly and linked
There is no rule here to quote. This week's document is not a resolution, a ruling, an audit or a bill, and no passage carries one. What it carries is a measured test of the eye: a preregistered survey experiment in which each respondent was shown seven images drawn at random from a pool of 52 and asked how likely each was made with AI1.
The test was run on 2,091 adult Danes, aged 18 and older, between June 17 and June 24, 2024. It was published in the October–December 2026 issue of Social Media + Society. Nobody enforces anything in it. The only thing in force is a finding, and its limits are as binding as its numbers1.
The authors say plainly that the images were made with a free, publicly available tool, and that they are likely lower in quality than what a well-funded operation could produce. The authors also note that the setup makes the results more conservative, because respondents were told in advance that some images were AI and were expected to do better in the experiment than they would on social media2.
### THE CASE IT ANSWERS
No real instance was found this week.
### WHAT WAS MEASURED
Picture the test from the respondent's chair. Seven images arrive one at a time, each with a short line saying what it shows and where the event is from. The respondent slides a marker between 0 and 100: how likely is this made by a machine? The median time spent looking at an image was seven seconds. Then the next image comes.
The scoring punishes confident mistakes hardest. A respondent who says 100 percent AI on a real photograph, or 0 percent on a generated one, is marked down more than a respondent who hedges at 50. The measure is called a Brier score, and lower is better3.
On average, the Danes were slightly better than random guessing1. Then came the finding that matters. Against the simplest possible strategy — writing 50 for every image and thinking no further — respondents did worse. The difference was 0.034, with a 95 percent confidence interval from 0.028 to 0.0393. In plain words: a person who shrugged at every picture would have scored better than the average participant who tried.
Confidence ran ahead of accuracy. Respondents were more confident judging images of people than images of crises, while being no better at telling the two apart. Familiarity made them sure, not right.
The categories did not rescue anyone. Regular scenes, people, crises, political events: no category differed significantly from another4. One group cleared the bar: young people under 25 performed substantially better than chance5. Analytical thinkers did better than less analytical ones.
The study measures discernment and confidence. It does not measure whether a false image changes a vote, a belief or a turnout. Nobody in this experiment cast a ballot.
### THE OTHER REGION'S ANSWER TO THE SAME THING
None was found this week: no court, law or authority outside Denmark, in the passages read, has ruled on or regulated the discernment of AI-generated images.
### THE LIBRARY'S READING
Here is weeklyAI's reading. A false rumor about a candidate has always done one thing well: it leaves a residue. The charge is denied, the denial is printed, and something stays behind — the sense that where there was talk, there may have been something. What is new is not the appetite for the rumor. It is the cost of making the evidence. A convincing picture of an event that never happened used to require a forger, a darkroom and time. Now it requires a prompt and a few minutes.
This reading leads weeklyAI to expect something in homes like the reader's. When a striking image of a candidate or a disaster arrives, the judgment will lean on how familiar and how plausible the scene feels, and on how many others are already reacting to it, rather than on any check of the image itself. The feeling of certainty will arrive before any verification does. This reading would be wrong if people who felt most sure turned out to be the ones who judged best, or if heavy circulation made a picture seem less real rather than more. Neither happened in the Danish test: confidence and accuracy came apart.
Other work read for this story gives the reader a caution. One study reports that individuals generally can identify AI-manipulated images, the opposite of the near-chance result here, which is a reminder about carrying a finding from one test to another, not a doubt about what happened in this one. Another line of argument warns that piling on labels and warnings can add to the work of judging without ever closing the question, so that people give up the effort while keeping the ability.
What the reader can do with this is small and concrete. Before you forward a picture that moves you, ask where it first appeared and who else is showing it. Treat your own certainty as a warning light rather than a certificate. And when a post is going viral, separate two questions that the feed fuses into one: is this spreading, and is this real. Only the second one is worth your vote.
### WHAT IS NOT KNOWN
The experiment cannot say what happens in a feed. Participants were primed, warned and watching for a trick; the authors themselves say the results are more conservative than real life, and more conservative means the real gap is likely wider, not narrower2.
It cannot say whether seeing a false image changes a vote, because no vote was measured. It used a single AI model, made images with a free tool, and cannot speak to what a better-funded operation would produce. It is a snapshot from one country, taken in June 2024, among 2,091 adult Danes.
No rule with enforcement behind it appears in this document. The authors suggest watermarks, transparency standards and monitoring duties for platforms, and they note the risk that raising alarm can make people doubt true things too, or let a dishonest actor dismiss real evidence as machine-made6. None of that is in force.
The record is silent on the count. Nothing in these passages touches a register, a tally sheet, a ballot or an electoral authority, and no audit of any count appears here.
So the voter is left with a question that fits in one breath before the screen: before I believe this picture, who else is showing it, and where did it start?
Where each piece of context comes from, and how much of it we read
- Hove, M. F., Hjorth, F., Golovchenko, Y. et al. (2026). Seeing Is Believing? Most People Struggle to Detect AI-Generated Images. Social Media + Society. https://doi.org/10.1177/20563051261483376 - the article this story is about — the whole article — the passage: "In a preregistered survey experiment, we find that, on average, people are not substantially better than chance at discerning AI-generated images from authentic images, regardless of the context of the image."
- Hove, M. F., Hjorth, F., Golovchenko, Y. et al. (2026). Seeing Is Believing? Most People Struggle to Detect AI-Generated Images. Social Media + Society. https://doi.org/10.1177/20563051261483376 - the article this story is about — the whole article — the passage: "However, this concern primarily makes the results more conservative as respondents are expected to perform better in the experiment than they would on social media."
- Hove, M. F., Hjorth, F., Golovchenko, Y. et al. (2026). Seeing Is Believing? Most People Struggle to Detect AI-Generated Images. Social Media + Society. https://doi.org/10.1177/20563051261483376 - the article this story is about — the whole article — the passage: "However, the difference in means between individuals and the consistent 50% probability evaluation is 0.034 and statistically significant (p < .01, 95% CI = [0.028, 0.039]). This means that while individuals are better than random evaluations, they would, on average, have been better off by taking the 50% probability evaluation."
- Hove, M. F., Hjorth, F., Golovchenko, Y. et al. (2026). Seeing Is Believing? Most People Struggle to Detect AI-Generated Images. Social Media + Society. https://doi.org/10.1177/20563051261483376 - the article this story is about — the whole article — the passage: "This indicates that individuals’ discernment abilities are not different for different types of images, and for three out of four categories, not in any way that would change the interpretation of H1."
- Hove, M. F., Hjorth, F., Golovchenko, Y. et al. (2026). Seeing Is Believing? Most People Struggle to Detect AI-Generated Images. Social Media + Society. https://doi.org/10.1177/20563051261483376 - the article this story is about — the whole article — the passage: "While younger people, highly analytical-thinking individuals, and experienced users of generative AI have better discernment abilities, no demographic group other than young individuals (<25) performs substantially better than chance."
- Hove, M. F., Hjorth, F., Golovchenko, Y. et al. (2026). Seeing Is Believing? Most People Struggle to Detect AI-Generated Images. Social Media + Society. https://doi.org/10.1177/20563051261483376 - the article this story is about — the whole article — the passage: "Policy interventions could require generative AI tools to include watermarks, enforce transparency standards, and impose stricter monitoring obligations on social media platforms. Achieving this will require substantial investment in AI detection tools and verification methods."
Hove, M. F., Hjorth, F., Golovchenko, Y. et al. (2026). Seeing Is Believing? Most People Struggle to Detect AI-Generated Images. Social Media + Society. https://doi.org/10.1177/20563051261483376
Who paid: The authors received no financial support for the research, authorship, and/or publication of this article; data collection was conducted as part of a grant from the Agency for Culture and Palaces and the Agency for Digital Government in Denmark, and the article does not say the funders had any role.
The findings of other studies mentioned here come from this document; those studies were not opened one by one.

