experiment · Social Media + Society · la publicación, 1 oct 2026 · gratis
Adultos daneses vieron fotos hechas con IA y apenas acertaron
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 de 18 años o más
- Cuántos
- 2,091
- Dónde
- Dinamarca
- Cuándo
- del 17 al 24 de junio de 2024
- Tipo de estudio
- experiment
- Quién lo hizo
- Universidades de Ámsterdam, Copenhague, Aarhus, Dinamarca del Sur y London School of Economics
- El límite que importa
- Las imágenes eran de baja calidad, hechas con una sola herramienta gratuita, y a los participantes se les avisó de antemano.
Puntería al juzgar imágenes hechas con IA frente a responder 50% a todas
Los analíticos pasaron de 0.30 a 0.26 en la medida del estudio, y el punto de referencia era 0.25
Fue el único grupo que quedó claramente por debajo de ese punto
La gente se sintió más segura con las de personas, pero esa confianza no se tradujo en mejor puntería
En promedio, la gente no fue mucho mejor que el azar para distinguir una imagen hecha por IA de una foto auténtica, y eso pasó sin importar el tipo de imagen que vieran.
En Dinamarca, 2,091 adultos vieron siete fotos cada uno, tomadas al azar de un grupo de 52, y trataron de decir cuáles eran hechas con inteligencia artificial y cuáles eran reales. Los investigadores les avisaron antes que habría imágenes falsas. Aun así, el resultado quedó apenas por encima de adivinar al azar, y por debajo de responder 50% a todas. Los jóvenes de 18 a 24 años fueron el único grupo que lo hizo mejor que responder 50% a todas. Las imágenes eran de baja calidad, hechas con una herramienta gratuita, así que en redes sociales, sin aviso, la gente probablemente distingue menos.
Las fotos mostraban eventos cotidianos, crisis, temas políticos y personas, en su mayoría conocidas. El estudio no nombra ninguna página donde usted pueda verificar una imagen por su cuenta.
La ley danesa no aparece en el artículo, ni una resolución de la autoridad electoral. Los autores proponen marcas de agua y reglas de transparencia, pero no las probaron.
Qué significa para usted
Para usted, que va a votar, el estudio danés deja una advertencia concreta: si una foto le mueve el ánimo, no la tome como prueba de lo que pasó. Los autores proponen marcas de agua y reglas de transparencia, pero no las probaron, así que por ahora no hay una herramienta casera que le diga qué imagen es falsa.
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, la autoría ni la publicación de este artículo; la recogida de datos se realizó en el marco de una subvención de la Agencia de Cultura y Palacios y la Agencia de Gobierno Digital de Dinamarca, y el artículo no indica si esos financiadores tuvieron alguna influencia.
Versión detalladaLos pasajes copiados del artículo, las ilustraciones y cada fuente con cuánto leímos de ella · 1079 palabras · unos 5 minLeerla →Cerrar
La regla y la evidencia: unas fotos falsas parecían reales y casi nadie las distinguió
### La regla, tal como está escrita
Esta semana no hay una resolución, una sentencia ni una auditoría que citar. El documento que sí llegó es un estudio científico, y su regla es otra: una prueba con 2,091 adultos daneses, mayores de 18 años, a quienes se les mostraron siete imágenes al azar de un conjunto de 52. La prueba se hizo del 17 al 24 de junio de 2024. No hay autoridad que la aplique ni plazo que venza. Lo que hay es un resultado que conviene leer con calma.
### El caso que responde
Esta semana no se encontró ningún caso real que contar con nombre, lugar, fecha y segunda fuente. Así que esta parte se queda en una sola frase: no hubo caso esta semana.
### Lo que se midió
Imagine la escena: usted está en el teléfono, con prisa, y ve una foto. Puede ser una vaca en un pastizal, un incendio en California, una bandera de Ucrania sobre la puerta de Brandeburgo o una mujer mayor recibiendo una vacuna. Algunas de esas fotos las hizo una persona. Otras las generó un programa gratuito llamado OpenArt.ai, con el que cualquiera puede escribir una descripción y esperar entre cinco y quince minutos a que salga una imagen parecida a la real1. A cada persona del estudio se le pidió decir, en una escala de 0 a 100, qué tan probable era que la imagen fuera hecha por inteligencia artificial.
El resultado central es incómodo. En promedio, la gente no fue mucho mejor que el azar para distinguir una imagen hecha por IA de una foto auténtica, y eso pasó sin importar el tipo de imagen que vieran2. Hubo una diferencia pequeña a favor de las personas frente a respuestas totalmente al azar, pero resultó que habrían hecho mejor si le hubieran puesto 50% a todas las imágenes sin mirarlas3. Es decir: el esfuerzo de mirar y decidir no les rindió.
Ningún tipo de imagen resultó más fácil que otro. Las imágenes de personas, de crisis, de política y de la vida diaria dieron resultados parecidos, y ninguna categoría se separó de las demás de forma clara4. La gente se sintió más segura al juzgar imágenes de personas que al juzgar imágenes de crisis, pero esa confianza no se tradujo en mejor puntería56. La seguridad de más no sirvió de mucho.
Los grupos que sí mostraron alguna ventaja fueron pocos. Quienes piensan de forma más analítica tuvieron un puntaje mejor, aunque la diferencia no los sacó del problema: pasaron de 0.30 a 0.26 en la medida que usó el estudio, y el punto de referencia era 0.257. El único grupo que quedó claramente por debajo de ese punto fueron los jóvenes de 18 a 24 años8. Fuera de ellos, ningún grupo demográfico superó de forma sustancial el nivel del azar9.
El estudio tiene límites que sus propios autores señalan. Las imágenes eran de calidad más bien baja, hechas con una sola herramienta gratuita, así que el desempeño de la gente en esta prueba puede tomarse como un techo, no como un piso, frente a campañas de desinformación más sofisticadas10. Además, a los participantes se les avisó de antemano que verían imágenes creadas por IA, lo que probablemente hace que el resultado sea más conservador que en la vida real. Y el estudio se hizo en Dinamarca, un país con alto nivel de alfabetización digital, así que si allí no pudieron, es probable que en otros países tampoco11. Eso no significa que los números se trasladen tal cual a su país; significa que la dirección del resultado es preocupante.
El propio artículo menciona trabajos anteriores con resultados mixtos: algunos encontraron que la gente sí puede identificar imágenes manipuladas con IA, mientras que otros hallaron que las personas apenas superan el azar. Eso no desmiente este estudio, pero sí advierte que la habilidad de distinguir no es la misma en todos lados ni con todas las imágenes. Es una advertencia sobre trasladar resultados de un lugar a otro, no sobre lo que aquí se midió.
### La respuesta de otra región al mismo asunto
Esta semana no encontramos ninguna corte, ley o autoridad de otra región que se haya pronunciado sobre este mismo método. No hay nada que contar en esta parte.
### La lectura de la biblioteca
La propaganda siempre ha hecho lo mismo: toma un miedo que ya existe en la gente, lo alimenta y lo exagera hasta convertirlo en certeza. Casi siempre viene de alguien de confianza, y se apoya en fotos alteradas, documentos fabricados o material sacado de contexto. Lo nuevo no es el engaño; es el costo. Antes hacía falta un taller, un impresor, un equipo con recursos. Ahora basta un teléfono y unos minutos.
Lo que un público puede aprender a revisar es la fuente: quién publicó primero, si hay una fecha, si algún medio o verificador lo confirmó. Lo que no puede aprender a revisar con la sola vista es si la imagen fue fabricada, porque las señales técnicas que antes delataban el engaño se están borrando. Esta lectura se equivocaría si resulta que la familiaridad con la imagen no aumenta su credibilidad, o si la gente distingue lo fabricado por rasgos técnicos aunque le resulte conocido. Antes de reenviar una imagen que "prueba" algo, pregunte quién la publicó primero y si alguna autoridad, medio o verificador la confirmó. Si desconfía de la institución, exija su informe o auditoría en vez de sustituirla por una foto.
### Lo que no se sabe
No se sabe si ver una imagen fabricada cambia un voto, una creencia o una decisión de compartir. El estudio midió si la gente distingue, no qué hace después con lo que ve. Tampoco se sabe cómo le iría a la gente con imágenes de herramientas más avanzadas, ni con modelos distintos al que se usó aquí. No se sabe si las medidas que los autores proponen, como marcas de agua o mayor vigilancia de las plataformas, funcionarían, porque no se probaron12. Y no se sabe si en su país la gente se comportaría igual que en Dinamarca.
La pregunta que le queda a usted antes de votar no es si puede confiar en sus ojos. Es a quién le va a pedir la verificación cuando sus ojos le digan una cosa y su desconfianza le diga otra. Esa pregunta no la resuelve una foto.
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: "The AI-generated images were created using OpenArt.ai, a free and publicly available tool. Creating each AI-generated image involved entering multiple prompts and selecting the most suitable image from the generated results, taking approximately 5 to 15 min per 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: "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: "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: "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: "Participants were more confident in their judgments when evaluating images of people compared to images displaying crises. This effect was not hypothesized or preregistered and should therefore be considered exploratory."
- 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, despite this higher confidence, actual discernment was not substantively better than chance. This overconfidence may reflect a tendency to detect patterns where none exist, especially when people expect themselves to be able to discern."
- 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: "Compared to individuals with lower levels of analytical thinking, who have an average Brier score of 0.30, the highly analytical-thinking individuals have an average 0.04 lower Brier score, bringing this group of individuals down to an average Brier score of 0.26. However, this still does not exceed 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: "The only parameter where this is the case is age, where young people (18–24-yearolds) have an average Brier score statistically significantly below 0.25."
- 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: "Therefore, participants’ discernment abilities in this experiment can be considered an upper bound compared to more advanced disinformation efforts."
- 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: "As a result, Denmark can be considered a more likely case to find individuals who can discern AI-generated from authentic images, meaning that if Danes cannot do well in the test, we would expect that to hold true for most countries worldwide."
- 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."
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, la autoría ni la publicación de este artículo; la recogida de datos se realizó en el marco de una subvención de la Agencia de Cultura y Palacios y la Agencia de Gobierno Digital de Dinamarca, y el artículo no indica si esos financiadores tuvieron alguna influencia.
experiment · Social Media + Society · the paper, 1 Oct 2026 · free
Most people shown AI images in Denmark could not reliably tell them from real photos, and the test was easier than social media
Short version · the longer version follows, about 7 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 aged 18 and older
- How many
- 2,091 respondents
- Where
- Denmark
- When
- June 2024
- Kind of study
- experiment
- Who did it
- University of Amsterdam, University of Copenhagen, Aarhus University, University of Southern Denmark, LSE
- The limit that matters
- Only one free AI tool was used, and every image showed an event that really happened.
How well people did at telling AI images from real photos
The high group had a Brier score 0.04 lower, meaning more accurate answers, but still did not beat answering 50 for every image.
Young people had an average Brier score below 0.25, the only group to beat the 50% mark.
On average they did only slightly better than random guessing, and worse than simply answering 50 for every image.
In June 2024, researchers asked 2,091 adults in Denmark to judge seven images each, drawn at random from 52 photos of real events. Half the photos were made with a free AI tool, and people rated each one from 0 to 100 for how likely it was AI-generated. On average they did only slightly better than random guessing, and worse than simply answering 50 for every image. Only people aged 18 to 24 beat that 50 mark.
The article names no way for a voter to check this at home. It proposes watermarking by AI tools, transparency rules and platform monitoring, but tested none of them.
It also reports no rule, ruling or audit from any electoral authority. Denmark is not named as having one on this point, and no other country's authority is mentioned. The article also says news media and fact-checking groups should be strengthened to spot AI-generated content.
What this means for you
For now, watch how you judge a photo before you vote, and ask whether the source can be checked rather than guessed. The test ran on 2,091 adult Danes and found that people could not reliably tell AI-made pictures from real ones, so a picture that looks real is not proof. Ask your electoral authority what it has published on AI images and the count, and treat any image without a verifiable source as a question, not an answer.
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 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 whether those funders had any say.
The longer versionThe passages copied from the paper, the pictures, and every source with how much of it we read · 1492 words · about 7 minRead it →Close
The rule and the evidence: Most people shown AI images in Denmark could not reliably tell them from real photos, and the test was easier than social media
### THE ACT, quoted exactly and linked
There is no act this week. The document is a study, not a law, a ruling, an audit or a bill. No resolution is in force, no authority enforces anything, and no date is fixed by law for a person to act. The document records a test and what came of it, and nothing more.
What the study did set out to find is stated plainly: "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"1. That is the whole of the finding at its widest, and it is a measurement, not a rule.
### THE CASE IT ANSWERS
No real instance was found this week.
### WHAT WAS MEASURED
Consider what the test actually asked of a person. Two thousand and ninety-one adult Danes, eighteen and older, sat with a screen and were shown seven images, one at a time, drawn at random from a pool of fifty-two. Half the pool was made by a machine, half was real, and each real event had a machine-made twin. After each image the respondent slid a marker between 0 and 100 to say how likely it was that a machine made it. The median time spent looking was seven seconds.
The scoring rewards a cautious answer over a confident one. A person who says "fifty-fifty" every time and is wrong half the time earns a certain middling score. A person who says "certainly fake" and is wrong earns a worse one. On that scale, the Danes did slightly better than if they had picked a random number for every image: the gap was 0.05, and it was statistically significant. But they did worse than the person who simply answers fifty-fifty every time: that gap was 0.034, also statistically significant2. Put plainly, the average respondent would have scored better by refusing to judge at all.
The kind of picture made no difference. Everyday scenes, crowds, disasters, political moments: the scores did not separate in any way that changes the reading, and for three of the four kinds the result was no better than random3. The hope that faces would be easier, because people look at faces all day, did not hold. The difference between pictures of people and ordinary pictures went the expected direction but was small and not statistically significant.
Some groups did better. People who score high on a test of careful reasoning had a Brier score 0.04 lower than people who score low4, which means more accurate answers. Yet that still left them at about 0.26, and 0.25 is what you get by answering fifty-fifty every time5. So the most careful reasoners in the sample were still, on average, no better than a coin that always lands on its edge. Only one group beat that line: people aged eighteen to twenty-four6.
The images themselves were made with a free tool anyone can open, OpenArt.ai, at roughly five to fifteen minutes per picture7. The authors say this mimics the small operator hunting for clicks, not a state campaign, and that the images are therefore on the low-quality end. That cuts in the respondents' favour. The pictures were easy compared with what a well-funded forger could make, so the scores are an upper bound8. And because the respondents were warned in advance that fakes were coming, which real life does not do, they would be expected to do better in the test than on social media, which makes the results more cautious still.
No passage in this document measures whether seeing such an image changes a vote, a belief, a share or a turnout. The study measures discernment alone. It was done in Denmark, and the authors argue Denmark is a hard test to fail: near-universal internet use, high digital literacy, most people on social media, most people getting news online. If Danes cannot do well, the reasoning goes, most countries would do no better9. That is an argument about carrying over, not a measurement made anywhere else.
### THE OTHER REGION'S ANSWER TO THE SAME THING
Other work this week approaches the same problem from the machine's side rather than the voter's. One team reports that standard detection tools look at single frames and are blind to the flicker between frames, the odd blink, the seam where motion breaks10, and that compression on social media blunts the traces they hunt for11. Another reports that automated detectors can reach high accuracy on a benchmark set of videos12, and a third reports similar gains on its own tests13. None of these is a law, a court ruling or an electoral authority's decision, and none was tested on the kind of still images a voter scrolls past.
### THE LIBRARY'S READING
Here is weeklyAI's reading. A false story about a candidate has always travelled, and it has always worked by making a person feel that the question was settled somewhere they could not see. What is new is not the lie. What is new is that the proof can now be manufactured to order, in minutes, by anyone, and that the habit of treating a picture as proof has not been switched off. A reader who has spent a lifetime learning that a photograph shows what happened will keep applying that lesson to a photograph that shows nothing of the kind.
weeklyAI expects this to bite hardest where the picture feels familiar. A face you know, a street you have walked, a disaster that looks like the last one: the sense of recognition arrives before any checking does, and the study's finding that confidence runs ahead of accuracy points the same way. This reading would be wrong if readers, told plainly that a picture may be manufactured, stopped citing it as proof and began asking what else confirms the event. The thing to do with it is small and repeatable. Before you pass a picture on, ask what else tells you the event happened: a statement from someone with a name, a second outlet's own photograph, a place and a date you can look up. Treat a familiar face as a reason to look harder, not as the reason to stop looking.
weeklyAI also read other work this week that points the other way, and it belongs here. One study found that people could generally identify images that had been manipulated by AI, which is not the same as spotting images made from scratch. That is a caution about carrying a result from one set of people and one set of pictures to another. It does not undo the Danish numbers, and it does not tell you how you would do.
### WHAT IS NOT KNOWN
What is not known is large. The study cannot say whether a person who cannot spot a fake would still be moved by one; a picture can fail the test and still do its work. It cannot say how these respondents would fare against pictures from a better tool, because only one free tool was used. It cannot say how they would do on a fabricated event, since every image here showed something that really happened, which removes the memory cues but also removes the ordinary experience of meeting a lie.
The study does discuss the widely reported case of AI-generated images of Taylor Swift falsely shown supporting Donald Trump, which the candidate reposted, and notes that a fact check of the image circulated widely, including from Swift herself. It is a study from one country, and reading it as though it were measured here would be a mistake the authors themselves warn against. It is a study from one country, and reading it as though it were measured here would be a mistake the authors themselves warn against.
The authors do propose things: watermarks on the tools, transparency rules, stricter monitoring of platforms14. None of these was tested here. They also warn that teaching people to hunt for visual flaws may backfire, leaving them more sure and no more right, and that loud warnings about fakes can make people doubt true things too. Those are cautions, not findings.
So what does a voter do with this before a date fixed by law? Not learn to spot the flaw, since the flaws are going and the confidence is not. The usable move is to stop treating the picture as the evidence and start asking what stands behind it. When a photograph of a candidate or a crisis arrives, ask who took it, where it was published first, and whether anyone with a name will stand behind it. If the only answer is the picture itself, you have learned what you needed to know.
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: "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: "On average, people with a high level of analytical thinking have a 0.04 lower Brier score than those with a low level of analytical thinking (p < .001, 95% CI = [−0.051, −0.021])."
- 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: "Compared to individuals with lower levels of analytical thinking, who have an average Brier score of 0.30, the highly analytical-thinking individuals have an average 0.04 lower Brier score, bringing this group of individuals down to an average Brier score of 0.26. However, this still does not exceed 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: "The only parameter where this is the case is age, where young people (18–24-yearolds) have an average Brier score statistically significantly below 0.25."
- 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: "The AI-generated images were created using OpenArt.ai, a free and publicly available tool. Creating each AI-generated image involved entering multiple prompts and selecting the most suitable image from the generated results, taking approximately 5 to 15 min per 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: "Therefore, participants’ discernment abilities in this experiment can be considered an upper bound compared to more advanced disinformation efforts."
- 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: "As a result, Denmark can be considered a more likely case to find individuals who can discern AI-generated from authentic images, meaning that if Danes cannot do well in the test, we would expect that to hold true for most countries worldwide."
- Deshpande S, Lajurkar S, Sarak P, Jain PPA. (2026). VeriMedia: A Spatial-Temporal Deep Learning Platform for Video Deepfake Detection and Forensic Verification. IJARCCE. https://doi.org/10.17148/ijarcce.2026.1511115 — only the abstract - the full text could not be fetched — the passage: "Standard digital forensic workflows rely on static two-dimensional Convolutional Neural Networks (2D-CNNs) that are inspecting isolated frames; in fact, yet, frame-by-frame classifiers are fundamentally blind to temporal boundary flickers, physiological inconsistencies (such as irregular eye blinking flow and micro-expression dynamics), and motion discontinuities."
- Deshpande S, Lajurkar S, Sarak P, Jain PPA. (2026). VeriMedia: A Spatial-Temporal Deep Learning Platform for Video Deepfake Detection and Forensic Verification. IJARCCE. https://doi.org/10.17148/ijarcce.2026.1511115 — only the abstract - the full text could not be fetched — the passage: "Along these dimensions, suffering from acute vulnerability to lossy social media compression (such as H.264/H.265 codecs at high quantization parameters c23 and c40)"
- Liu C, Wang X, Zhang Y, Liu J, Ma Z. (2026). A spatiotemporal defect-integrated deepfake video detection and forgery algorithm attribution model. Scientific Reports. 10.1038/s41598-026-55334-x — only the abstract - the full text could not be fetched — the passage: "Comprehensive experiments on the FaceForensics++ dataset demonstrate that the proposed method achieves 97.86±0.18% detection accuracy and 99.81±0.11% AUC, as well as 98.42±0.15% accuracy for forgery algorithm attribution, outperforming most state-of-the-art approaches on this dataset."
- Hussain M, Saeed F, Aldera S. (2026). Bioinspired Adaptive-Depth Neural Growth for Deepfake Video Forensics: An Entropy-Guided State-Space Framework. Biomimetics. 10.3390/biomimetics11090653 — only the abstract - the full text could not be fetched — the passage: "Evaluated using Celeb-DF v2, AEGIS-FD achieves a test AUC of 0.9600 and a validation AUC of 0.9607, above the performance of a single-layer Mamba SSM baseline (AUC = 0.8355)."
- 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."
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 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 whether those funders had any say.

