A detector flagged fake faces in a lab test, not in your feed.

experiment · Cybernetics and Information Technologies · la publicación, 1 jun 2026 · gratis

Un detector automático de caras falsas acertó 95 de cada 100 veces en un conjunto de imágenes fijas; no se probó con video.

Versión breve · la versión detallada sigue, unos 5 min

Pregunte a weeklyAI

Pregúnteme por este estudio: a quiénes se estudió, qué encontró y qué no dice.

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: cinco modelos automáticos de detección de caras falsas (un Vision Transformer y cuatro redes convolucionales)
Cuántos: 140,000 imágenes de caras: 70,000 reales y 70,000 falsas
Dónde: no lo dice el pasaje
Cuándo: no lo dice el pasaje
Tipo de estudio: experiment
Quién lo hizo: no lo dice el pasaje
El límite que importa: Solo se probaron rostros quietos de un único conjunto; no se probó con video ni audio.

Precisión de cada modelo al distinguir caras reales de falsas

Vision Transformer (ViT-S/16)0.95accuracy

MobileNet-V20.9accuracy

Inception-V30.86accuracy

VGG-160.84accuracy

ResNet-500.61accuracy

Precisión de cada modelo sobre un único conjunto de 140,000 imágenes fijas de caras; no se probó con video ni audio.

Qué regiones de la cara usó cada tipo de modelo para decidir

Modelos convolucionales (VGG-16, ResNet50, InceptionV3, MobileNetV2)frente aVision Transformer (ViT-S/16)

Los convolucionales se fijaron en ojos, nariz, boca y contornos del cabello; el Transformer repartió su atención por toda la imagen.

Un detector automático de caras falsas acertó 95 de cada 100 veces en un conjunto de imágenes fijas; no se probó con video.
Lectura de weeklyAI

Así podría verse · ilustración generada por weeklyAI.watch, no es una fotografía

Cinco modelos fueron entrenados para distinguir caras reales de falsas usando 140,000 imágenes: 70,000 reales y 70,000 falsas. El conjunto de prueba tuvo 20,000 imágenes, mitad reales y mitad falsas. El modelo Vision Transformer (ViT-S/16) logró 0.95 de exactitud general y 0.94 de sensibilidad para la clase falsa. MobileNet-V2 obtuvo 0.90; Inception-V3, 0.86; VGG-16, 0.84; y ResNet-50, 0.61. Solo se usaron fotografías fijas de un único conjunto de datos.

El artículo no nombra ninguna forma en que usted pueda verificar esto por su cuenta.

El artículo no menciona ninguna autoridad electoral, ninguna regla en vigor ni ninguna auditoría. Solo reporta resultados de laboratorio sobre imágenes fijas de un conjunto de datos.

Qué significa para usted

Cuando vea una cara sospechosa en un video de campaña, recuerde que este estudio solo evaluó fotografías fijas, no material audiovisual, y que la exactitud de 0.95 corresponde a un conjunto de imágenes de laboratorio, no a lo que circula en redes ni a lo que cuenta su autoridad electoral. Ningún resultado aquí le dice si el conteo de votos fue alterado, así que pida los informes y auditorías de su organismo electoral antes de concluir algo.

Koudad, Z., Bekkouche, A., Benahmed, H. et al. (2026). Trustworthy Deepfake Detection: Explainable LIME Method of ViT and CNN Architectures. Cybernetics and Information Technologies. https://doi.org/10.2478/cait-2026-0014

Quién pagó: El artículo no dice quién financió el estudio, si los financiadores tuvieron alguna influencia, ni quién prestó equipos o software.

Versión detalladaLos pasajes copiados del artículo, las ilustraciones y cada fuente con cuánto leímos de ella · 1081 palabras · unos 5 minLeerla →Cerrar

La regla y la evidencia: un detector automático de caras falsas acertó 95 de cada 100 veces en un conjunto de imágenes fijas; no se probó con video

Un estudio entrena cinco sistemas para distinguir rostros reales de falsos y les pide que señalen qué parte de la cara los convenció. La pregunta del votante es otra: ¿quién audita esa herramienta y con qué caras se entrenó?

### EL ACTO

El documento que tenemos este viernes no es una resolución ni una sentencia ni una auditoría electoral: es un artículo de investigación aceptado el 5 de marzo de 2026 y publicado con fecha de recepción del 16 de diciembre de 2025¹. No está en vigor para nadie, no obliga a ninguna autoridad y no fija ninguna fecha límite para ningún votante¹.

Lo que el artículo sí hace es describir su propio procedimiento: cinco modelos de detección de rostros falsos fueron entrenados sobre el conjunto de 140,000 imágenes de caras reales y falsas, y luego explicados con un método llamado LIME². El texto define así su objeto: "A deepfake refers to visual or audio content that is generated or altered using advanced artificial intelligence techniques, particularly deep learning, to make a person appear to say or do something they never actually said or did"³.

Lo que un lector puede hacer con esto es limitado y conviene decirlo sin adornos: puede leer el método, puede pedir a quien le presente un detector automático que le muestre con qué caras fue entrenado y con qué método se explica, y puede negarse a tratar el resultado de una máquina como prueba mientras nadie se lo haya auditado². No hay aquí ninguna obligación para ninguna autoridad electoral, ni un plazo, ni un recurso que un votante pueda presentar¹.

### EL CASO QUE RESPONDE

Esta semana no encontramos ningún caso real, con nombre, lugar, fecha y segunda fuente, que pueda contarse como el asunto que este estudio responde; no lo inventamos.

### LO QUE SE MIDIÓ

El estudio trabajó sobre rostros quietos, no sobre video ni audio: 140,000 imágenes de caras, 70,000 reales y 70,000 falsas, repartidas en 100,000 para entrenamiento, 20,000 para validación y 20,000 para prueba, con la prueba dividida en partes iguales entre reales y falsas⁴. El conjunto se describe como especialmente difícil porque las diferencias entre una cara real y una falsa "no son fácilmente distinguibles al ojo humano"⁵.

Sobre esa base, el modelo más preciso fue una variante pequeña de Vision Transformer, con una precisión de 0,95 y un acierto de 0,94 sobre la clase falsa⁶. Para que se entienda la magnitud: de cada cien rostros falsos que se le presentaron, acertó en noventa y cuatro, y de cada cien rostros reales, identificó correctamente noventa y cinco coma tres cuatro, mientras que seis coma dos siete de cada cien caras falsas pasaron como reales⁴.

Los demás modelos quedaron por debajo: uno alcanzó 0,90 de precisión, otro 0,86, otro 0,84 y el último apenas 0,61⁷. El propio artículo advierte que el sobreajuste comenzó a aparecer hacia la época 18 del entrenamiento y que se guardó el modelo anterior a ese punto⁸.

El estudio es una evaluación de modelos sobre un solo conjunto de imágenes, y el propio texto señala que la dificultad del conjunto está en que las diferencias entre una cara real y una falsa no son fácilmente distinguibles al ojo humano⁵. No hay en el documento ningún intervalo de confianza, ninguna prueba de significación y ninguna validación cruzada entre conjuntos distintos⁸.

### LA RESPUESTA DE LA OTRA REGIÓN AL MISMO ASUNTO

Sobre el mismo método, esta semana no encontramos ninguna corte, ley o autoridad de otro país que se pronuncie en los pasajes que tenemos; lo decimos en una sola frase y no lo rellenamos con lo que creemos recordar.

### LA LECTURA DE LA BIBLIOTECA

Así lo leemos nosotros. Un rumor falso sobre un candidato siempre ha hecho lo mismo: llega con la cara de alguien de confianza, se apoya en una foto retocada, en un dato sacado de contexto o en un documento que existió pero decía otra cosa, y crece porque toca una grieta que ya estaba abierta en la comunidad. Lo nuevo no es la mentira: es la facilidad para fabricar la cara que la sostiene y la velocidad con que esa cara viaja sin que nadie haya visto el original. Un público puede aprender a desconfiar de un titular, a buscar la fuente, a comparar dos versiones; lo que no puede verificar solo, mirando, es si una imagen fue alterada.

De ahí lo que esperamos en las casas como la suya: que una herramienta automática que marca un video de campaña como falso o como auténtico empiece a circular como si fuera un veredicto, y que casi nadie pregunte con qué rostros se entrenó, quién la pagó ni si la autoridad electoral la revisó antes de usarla. Nos equivocaríamos si los votantes, los docentes y los periodistas exigieran ver las caras del entrenamiento, los errores por grupo y el acta de la auditoría antes de aceptar o rechazar un video. Lo que usted puede hacer con esto es concreto: cuando alguien le muestre un video marcado por una máquina, pida el nombre de quien la entrenó, las caras que usó y el documento de la autoridad que la revisó. Si no existen, trátelo como una pista y no como una prueba, y fíjese si los videos señalados se concentran siempre en los mismos candidatos o en las mismas comunidades.

### LO QUE NO SE SABE

No se sabe si estos modelos funcionarían sobre un video de campaña, un audio o un clip de redes sociales, porque solo se probaron rostros quietos de un único conjunto⁵.

El artículo advierte que el sobreajuste comenzó a aparecer hacia la época 18 del entrenamiento y que se guardó el modelo anterior a ese punto⁸.

El artículo describe el procedimiento general de LIME y, para la comparación entre modelos, seleccionó una imagen de prueba de la clase falsa que todos clasificaron correctamente².

El conjunto de imágenes está equilibrado: contiene el mismo número de caras falsas que de reales⁴.

No se sabe nada sobre un deepfake que se haya reportado en todas partes y no haya llegado a casi nadie, porque esta semana no encontramos ningún caso así.

Para la próxima vez que alguien le muestre un video marcado por una máquina, pregunte en voz alta: ¿quién entrenó esto, con qué caras, y quién lo auditó antes de que yo lo creyera?

De dónde sale cada dato de contexto, y cuánto leímos de cada documento

Koudad, Z., Bekkouche, A., Benahmed, H. et al. (2026). Trustworthy Deepfake Detection: Explainable LIME Method of ViT and CNN Architectures. Cybernetics and Information Technologies. https://doi.org/10.2478/cait-2026-0014 - el artículo del que trata esta nota — el artículo completo — el pasaje: "With the remarkable progress of generative AI, deepfakes have become highly realistic and increasingly difficult to detect, which has facilitated the spread of propaganda for political purposes and poses a serious threat to social security."
Koudad, Z., Bekkouche, A., Benahmed, H. et al. (2026). Trustworthy Deepfake Detection: Explainable LIME Method of ViT and CNN Architectures. Cybernetics and Information Technologies. https://doi.org/10.2478/cait-2026-0014 - el artículo del que trata esta nota — el artículo completo — el pasaje: "For each face image, LIME highlights the specific regions that had the greatest influence on the model’s prediction, providing a visual and interpretable explanation of its reasoning."
Koudad, Z., Bekkouche, A., Benahmed, H. et al. (2026). Trustworthy Deepfake Detection: Explainable LIME Method of ViT and CNN Architectures. Cybernetics and Information Technologies. https://doi.org/10.2478/cait-2026-0014 - el artículo del que trata esta nota — el artículo completo — el pasaje: "A deepfake refers to visual or audio content that is generated or altered using advanced artificial intelligence techniques, particularly deep learning, to make a person appear to say or do something they never actually said or did"
Koudad, Z., Bekkouche, A., Benahmed, H. et al. (2026). Trustworthy Deepfake Detection: Explainable LIME Method of ViT and CNN Architectures. Cybernetics and Information Technologies. https://doi.org/10.2478/cait-2026-0014 - el artículo del que trata esta nota — el artículo completo — el pasaje: "The model correctly identifies 95.34% of real faces and 93.73% of fake faces. However, 6.27% of fake faces are misclassified as real, and 4.66% of real faces are misclassified as fake."
Koudad, Z., Bekkouche, A., Benahmed, H. et al. (2026). Trustworthy Deepfake Detection: Explainable LIME Method of ViT and CNN Architectures. Cybernetics and Information Technologies. https://doi.org/10.2478/cait-2026-0014 - el artículo del que trata esta nota — el artículo completo — el pasaje: "The main advantage, or rather, the challenge posed by the 140k Real and Fake Faces dataset lies in the subtle differences between real and fake images, which are not easily distinguishable to the human eye."
Koudad, Z., Bekkouche, A., Benahmed, H. et al. (2026). Trustworthy Deepfake Detection: Explainable LIME Method of ViT and CNN Architectures. Cybernetics and Information Technologies. https://doi.org/10.2478/cait-2026-0014 - el artículo del que trata esta nota — el artículo completo — el pasaje: "Among these models, the Vision Transformer achieved the highest accuracy (0.95) and a high recall of 0.94 for the fake class, which is particularly important because misclassifying a real face as fake leads to immediate rejection, but failing to detect a fake face represents a far more serious threat"
Koudad, Z., Bekkouche, A., Benahmed, H. et al. (2026). Trustworthy Deepfake Detection: Explainable LIME Method of ViT and CNN Architectures. Cybernetics and Information Technologies. https://doi.org/10.2478/cait-2026-0014 - el artículo del que trata esta nota — el artículo completo — el pasaje: "The remaining models ranked as follows: MobileNet-V2 (accuracy 0.90), Inception-V3 (0.86), VGG-16 (0.84), and ResNet-50 (0.61)."
Koudad, Z., Bekkouche, A., Benahmed, H. et al. (2026). Trustworthy Deepfake Detection: Explainable LIME Method of ViT and CNN Architectures. Cybernetics and Information Technologies. https://doi.org/10.2478/cait-2026-0014 - el artículo del que trata esta nota — el artículo completo — el pasaje: "Naturally, at some point, overfitting begins to appear, which in our case occurs around epoch 18, where the validation curves start to diverge from the training curves. The best model was therefore selected as the one saved just before the onset of overfitting."

Quién pagó: El artículo no dice quién financió el estudio, si los financiadores tuvieron alguna influencia, ni quién prestó equipos o software.

experiment · Cybernetics and Information Technologies · the paper, 1 Jun 2026 · free

A detector flagged fake faces in a lab test, not in your feed.

Short version · the longer version follows, about 5 min

Ask weeklyAI

Ask me about this study: who was studied, what it found, and what it does not say.

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: five deepfake face detection models (one Vision Transformer, four CNNs)
How many: 140,000 face images (70,000 real, 70,000 fake)
Where: Tlemcen, Algeria
When: accepted on 05.03.2026
Kind of study: experiment
Who did it: institutions in Tlemcen, Algeria
The limit that matters: Tested on still face images from one dataset, not in an election or on social media.

Accuracy of five deepfake face detection models

Vision Transformer0.95accuracy

MobileNet-V20.9accuracy

Inception-V30.86accuracy

VGG-160.84accuracy

ResNet-500.61accuracy

Accuracy of each model on the test set of the 140k Real and Fake Faces dataset; these are lab results on still face images, not in an election or on social media.

Model accuracy compared with human accuracy at telling deepfake faces from real ones

Vision Transformer modelagainsthumans

model reached 0.95 accuracy; humans performed above chance at 56.1%

Vision Transformer modelagainsthumans in experiments designed to improve detection

model reached 0.95 accuracy; those humans reached 62.2%

A machine that scores in the nineties and a person who scores in the fifties are not rivals; they are two different instruments, and only one of them can show you its reasoning.
weeklyAI's reading

How it could look · illustration generated by weeklyAI.watch, not a photograph

Researchers trained five detection models — one Vision Transformer and four CNN architectures — on 140,000 face images (70,000 real, 70,000 fake), split into 100,000 for training, 20,000 for validation and 20,000 for testing. The best model reached 0.95 accuracy and 0.94 recall for the fake class. The article does not say where the work was performed.

The paper names no way for a voter to check a specific video or ad.

The paper states no electoral rule, no ruling and no audit; it is a model evaluation, not a finding by any electoral authority.

What this means for you

What this means for you as you weigh what you see before voting is narrow: five detection models trained on 140,000 face images, with the best reaching 0.95 accuracy and 0.94 recall for fake faces, and visual explanations of model decisions. The study did not test any real feed, ad or count, so watch for the electoral authority's own resolutions, audits and observation reports rather than treating a detector's score as proof about the vote.

Who paid: The article does not say who funded the study, whether funders had any say, or who lent equipment or software.

The longer versionThe passages copied from the paper, the pictures, and every source with how much of it we read · 1054 words · about 5 minRead it →Close

The rule and the evidence: A detector flagged fake faces in a lab test, not in your feed

Researchers taught five systems to tell real faces from generated ones. The best one catches most fakes—and points to the exact features behind each call.

### THE ACT

The document is not a law, a ruling or an audit. It is a research article, accepted on 05.03.2026, that develops and evaluates five deepfake face detection architectures—a customized small Vision Transformer (ViT-S/16) and four CNN-based architectures built with transfer learning from VGG-16, ResNet-50, Inception-V3, and MobileNet-V2—and integrates the LIME interpretability framework to explain their decisions¹.

Nothing here is in force for anyone, and no authority enforces it. What a person can take from it is a method and a measured result, not a right or a deadline.

### THE CASE IT ANSWERS

No real instance was found this week.

### WHAT WAS MEASURED

Five models were trained on the 140k Real and Fake Faces dataset—140,000 face images, 70,000 real and 70,000 fake, split into 100,000 for training, 20,000 for validation and 20,000 for testing—and the test set held 10,000 real and 10,000 fake faces, evenly balanced.

On that test set, the Vision Transformer reached the highest accuracy, 0.95, with a recall of 0.94 for the fake class; the remaining models ranked as MobileNet-V2 (0.90), Inception-V3 (0.86), VGG-16 (0.84) and ResNet-50 (0.61)². In the Vision Transformer’s confusion matrix, the model correctly identified 95.34% of real faces and 93.73% of fake faces, while 6.27% of fake faces were misclassified as real and 4.66% of real faces were misclassified as fake³.

The authors chose the fake-class recall as the metric that matters most, on the reasoning that misclassifying a real face as fake leads to immediate rejection, while failing to detect a fake face is the more serious threat⁴.

The models were not tested in an election, a court, a newsroom or on social media. They were tested on still face images from one dataset.

The difficulty of that dataset is the point: its real and fake faces are, in the article’s words, “not easily distinguishable to the human eye”⁵, and “extremely difficult for humans to tell real and fake faces apart”⁶.

A separate meta-analysis, read only as an abstract, pooled 36 studies covering 51 experiments and 13,197 participants and found that people performed above chance—56.1%—at telling deepfake faces from real ones, with higher accuracy for real faces than deepfake ones, and that experiments designed to improve detection reached 62.2%⁷⁸⁹. The same abstract warns that the results varied widely between studies and that the practical relevance of those accuracy figures should be treated with caution¹⁰¹¹.

Here is how we read it. A machine that scores in the nineties and a person who scores in the fifties are not rivals; they are two different instruments, and only one of them can show you its reasoning. What the detector offers is not a verdict you must accept but a place to look.

### THE OTHER REGION’S ANSWER TO THE SAME THING

We found none this week. No passage gives another country’s court, law or authority ruling on this detection method.

### THE LIBRARY’S READING

A false story about a candidate has always worked before the ballot is cast, and it has always worked the same way: it plants a suspicion that is hard to wash off, and it spreads because people enjoy passing it along. What is new is not the lie but the evidence. A manufactured face or voice arrives looking like proof, and the old defense—ask around, wait for the correction—moves slower than the thing it is meant to answer.

What a public can learn to check together is provenance: who made this, where did it first appear, has anyone independent looked at it. What no single voter can check alone is whether the tool doing the checking was built and tested on faces like theirs. That is a question for an authority, not a kitchen table.

We also read, in work on this kind of finding, that the results have not been shown to carry over to real campaign settings—to video, to audio, to the compressed clips people actually share. That is about carrying over, not about this result. And we read that systems of this class carry a known risk: trained on one kind of face, they can be less reliable on faces from other groups, and their errors do not fall evenly. That is something to watch for, not a finding of this study.

### WHAT IS NOT KNOWN

The models were trained on 100,000 images, validated on 20,000 and tested on a separate 20,000-image test set.

The LIME explanations were generated for a single correctly classified fake image, with 1,000 perturbed samples and the eight most influential regions kept.

Overfitting appeared around epoch 18, and the authors selected the model saved just before it began¹². The dataset is balanced half real and half fake.

The claim that LIME is most effective rests on prior work, not on this study’s own experiments¹³.

Here is how we read it. The pattern is simple: we treat what we see as true, and we rarely stop to ask how an image was made. That habit is strongest when the content moves us and when it arrives through a channel we already trust. So expect the flag to matter less than the feeling—unless the flag appears inside the app where the video is already playing, next to the face, at the moment of watching. You would prove us wrong if you saw people routinely seek out a detector’s explanation before forming an opinion, or revise a belief when a detector pointed to the specific features behind its call.

What you can do with it: when a political video moves you, pause and ask who made it, where it first appeared, and whether anyone independent has checked it—before you pass it on. If a check points to particular parts of a face, look at those parts yourself and treat the explanation as one more piece of evidence, not as a ruling. And ask your electoral authority whether any detection tool is in use, what audit was done, and whether the error rates were broken down by group.

What would you need to see, before you believed a machine’s word about a face?

Where each piece of context comes from, and how much of it we read

Koudad, Z., Bekkouche, A., Benahmed, H. et al. (2026). Trustworthy Deepfake Detection: Explainable LIME Method of ViT and CNN Architectures. Cybernetics and Information Technologies. https://doi.org/10.2478/cait-2026-0014 - the article this story is about — the whole article — the passage: "For each face image, LIME highlights the specific regions that had the greatest influence on the model’s prediction, providing a visual and interpretable explanation of its reasoning."
Koudad, Z., Bekkouche, A., Benahmed, H. et al. (2026). Trustworthy Deepfake Detection: Explainable LIME Method of ViT and CNN Architectures. Cybernetics and Information Technologies. https://doi.org/10.2478/cait-2026-0014 - the article this story is about — the whole article — the passage: "The remaining models ranked as follows: MobileNet-V2 (accuracy 0.90), Inception-V3 (0.86), VGG-16 (0.84), and ResNet-50 (0.61)."
Koudad, Z., Bekkouche, A., Benahmed, H. et al. (2026). Trustworthy Deepfake Detection: Explainable LIME Method of ViT and CNN Architectures. Cybernetics and Information Technologies. https://doi.org/10.2478/cait-2026-0014 - the article this story is about — the whole article — the passage: "The model correctly identifies 95.34% of real faces and 93.73% of fake faces. However, 6.27% of fake faces are misclassified as real, and 4.66% of real faces are misclassified as fake."
Koudad, Z., Bekkouche, A., Benahmed, H. et al. (2026). Trustworthy Deepfake Detection: Explainable LIME Method of ViT and CNN Architectures. Cybernetics and Information Technologies. https://doi.org/10.2478/cait-2026-0014 - the article this story is about — the whole article — the passage: "Among these models, the Vision Transformer achieved the highest accuracy (0.95) and a high recall of 0.94 for the fake class, which is particularly important because misclassifying a real face as fake leads to immediate rejection, but failing to detect a fake face represents a far more serious threat"
Koudad, Z., Bekkouche, A., Benahmed, H. et al. (2026). Trustworthy Deepfake Detection: Explainable LIME Method of ViT and CNN Architectures. Cybernetics and Information Technologies. https://doi.org/10.2478/cait-2026-0014 - the article this story is about — the whole article — the passage: "The main advantage, or rather, the challenge posed by the 140k Real and Fake Faces dataset lies in the subtle differences between real and fake images, which are not easily distinguishable to the human eye."
Koudad, Z., Bekkouche, A., Benahmed, H. et al. (2026). Trustworthy Deepfake Detection: Explainable LIME Method of ViT and CNN Architectures. Cybernetics and Information Technologies. https://doi.org/10.2478/cait-2026-0014 - the article this story is about — the whole article — the passage: "The challenge of the 140k Real and Fake Faces dataset is that it is extremely difficult for humans to tell real and fake faces apart, making the task even more challenging for CNN or Transformer-based models."
Stockner M, Convertino G, Cambedda S, Mazzoni G. (2026). Are humans able to discriminate between real and deepfake faces? A systematic review and meta-analysis. Computers in Human Behavior: Artificial Humans. 10.1016/j.chbah.2026.100332 — only the abstract - the full text could not be fetched — the passage: "Including 36 studies (k = 51 experiments), totaling 13197 participants, we tested if humans perform above chance-level in face detection tasks and if accuracy is higher for real (vs. deepfake) stimuli, alongside moderator analyses including a series of methodological variables (i.e., publication year, study quality, deepfake type and response modality)."
Stockner M, Convertino G, Cambedda S, Mazzoni G. (2026). Are humans able to discriminate between real and deepfake faces? A systematic review and meta-analysis. Computers in Human Behavior: Artificial Humans. 10.1016/j.chbah.2026.100332 — only the abstract - the full text could not be fetched — the passage: "Overall, our results provide evidence that humans perform above chance-level (56.1%), and confirm significantly higher accuracy for real vs. deepfake face stimuli."
Stockner M, Convertino G, Cambedda S, Mazzoni G. (2026). Are humans able to discriminate between real and deepfake faces? A systematic review and meta-analysis. Computers in Human Behavior: Artificial Humans. 10.1016/j.chbah.2026.100332 — only the abstract - the full text could not be fetched — the passage: "Finally, sub-group analyses on experiments aiming to improve deepfake detection indeed reached considerably higher accuracy (62.2%)."
Stockner M, Convertino G, Cambedda S, Mazzoni G. (2026). Are humans able to discriminate between real and deepfake faces? A systematic review and meta-analysis. Computers in Human Behavior: Artificial Humans. 10.1016/j.chbah.2026.100332 — only the abstract - the full text could not be fetched — the passage: "However, significant heterogeneity observed suggests high variability in the results obtained by individual studies and both effects are shown to be moderated by the assessed variables."
Stockner M, Convertino G, Cambedda S, Mazzoni G. (2026). Are humans able to discriminate between real and deepfake faces? A systematic review and meta-analysis. Computers in Human Behavior: Artificial Humans. 10.1016/j.chbah.2026.100332 — only the abstract - the full text could not be fetched — the passage: "However, practical relevance of the encountered accuracy should be considered with caution."
Koudad, Z., Bekkouche, A., Benahmed, H. et al. (2026). Trustworthy Deepfake Detection: Explainable LIME Method of ViT and CNN Architectures. Cybernetics and Information Technologies. https://doi.org/10.2478/cait-2026-0014 - the article this story is about — the whole article — the passage: "Naturally, at some point, overfitting begins to appear, which in our case occurs around epoch 18, where the validation curves start to diverge from the training curves. The best model was therefore selected as the one saved just before the onset of overfitting."
Koudad, Z., Bekkouche, A., Benahmed, H. et al. (2026). Trustworthy Deepfake Detection: Explainable LIME Method of ViT and CNN Architectures. Cybernetics and Information Technologies. https://doi.org/10.2478/cait-2026-0014 - the article this story is about — the whole article — the passage: "T s i g o s et al. [20] compared several explainability methods and concluded that the perturbation-based method LIME [21-23] was the most effective for highlighting the regions that influence the model’s decisions. This motivated the direct adoption of LIME as the chosen method for model explainability."

Who paid: The article does not say who funded the study, whether funders had any say, or who lent equipment or software.

The findings of other studies mentioned here are known to us through this document, which is the one we read; we did not open each of those studies.