Generative AI and videos: how can we distinguish between what is real and what is fake?

Sora, Veo, Kling, Runway: in just a few months, generative artificial intelligence has profoundly transformed video production. Now, a few words are enough to create strikingly realistic animated images. This technological acceleration is blurring the lines and making it increasingly difficult to identify an authentic video, even for knowledgeable audiences.

On social media, this content blends in with real images, circulates at high speed and shapes perceptions, opinions and reactions. For organisations, media outlets and professional networks, the question is no longer whether we will encounter AI-generated videos, but how to recognise them before they influence our judgements or decisions.

Faced with this avalanche of synthetic content, vigilance is becoming a strategic skill. Observing details, understanding the limitations of detection tools and adopting a critical stance towards images are now essential reflexes. Here are the main benchmarks for assessing the reliability of a video in the age of generative AI.

1. Pay attention to details

Despite rapid advances in generative models, some inconsistencies remain.

Faces and expressions may exhibit more or less subtle anomalies: irregular blinking, blurred or even frozen gaze, absent micro-expressions or micro-expressions that are out of sync with the emotion being expressed. Upon closer inspection, humans seem to sorely lack naturalness and authenticity.

Similarly, hands and body movements remain sensitive areas. Although six-fingered hands have largely disappeared from circulation, it is still common to see poorly proportioned hands and fingers, imprecise gestures or unnatural transitions, with blurring or misalignment. This effect can also be seen in hair textures or animal fur, which are sometimes blurred or, conversely, smoothed out and overly perfect.

Textual elements in an image (logos, inscriptions, signs, etc.) are often key signals for identifying a generated image, especially when they are not at the centre of the scene. Despite its advances, generative artificial intelligence still struggles to produce reliable written language: it draws shapes that resemble writing without actually structuring comprehensible words. The result is visually credible, unless you look closely.

Finally, the physical elements of the environment (shadows, reflections, depth of field) may lack consistency with the natural laws of light and perspective. IAG structures space with dramatic lighting and strong contrasts to give each scene an exaggerated intensity.

This can also be seen in changes to elements related to the sets or characters: it is common to see details disappear or transform several times within the same shot.

However, it should be borne in mind that these benchmarks are evolving: the imperfections visible today tend to disappear over generations of AI, making the detection process increasingly complex.

After this initial sorting by human eyes, which is essential but fallible, the analysis can be reinforced by technical tools capable of providing more objective clues about the origin of the content.

2. Tools that help… but don’t do everything

Faced with the rise of AI-generated videos, several verification tools have emerged (Sensity, Sightengine, Reality Defender, etc.).
Some services offer automated analysis, searching for statistical signatures specific to this type of video: pixel inconsistencies, atypical compression artefacts, or irregularities in facial movements. These tools do not deliver a definitive verdict, but provide a probability index.

Other approaches rely on reverse image search (via Google Image, for example, or TinEye or Yandex). By isolating a key image from the video, it is sometimes possible to identify an earlier version of the content, a different context, or an original source, thus revealing a recent misappropriation or fabrication. This method remains quite effective for videos purportedly related to current events.

At the same time, several technology players have taken steps to ensure source traceability. While some creators are playing the transparency card by indicating their use of AI through hashtags, mentions or labels provided by platforms, this practice is voluntary, hence the importance of additional measures to guarantee the origin of content.

Markers (called watermarks) can be embedded directly into videos created using certain tools to indicate their origin. Google, for example, deploys this type of technology with SynthID, and Adobe does the same with the ‘CR’ for ‘Content Credentials’ label. This approach aims to increase transparency, but it relies heavily on voluntary adoption by creators.

3. Developing a culture of digital vigilance

When a video raises doubts, the analysis must not stop at the image itself. A crucial step is to place the content within its informational ecosystem. A video purporting to document a real event (public event, political statement, crisis situation) usually leaves multiple traces on social media: testimonials, parallel recordings, media coverage. The absence of any corroboration by reliable sources or institutional actors is an important warning sign.

Collective observation also plays a significant role. Comment sections can reveal relevant analyses: inconsistencies mentioned, technical details debated, comparisons with other content. Without relying blindly on these reactions, they can draw attention to elements that went unnoticed during a first viewing.

While no single element can confirm with certainty that a video is fake, the accumulation of clues (visual absurdities, lack of reliable sources, dubious context of dissemination) remains the most reliable way to determine the credibility of a video.

In this context, the challenge becomes cultural: it is now a matter of developing a critical mind, slowing down the reflex to share, and systematically verifying sources before getting emotionally caught up.

In an environment where generation technologies are advancing very rapidly, doubt is a stance that needs to be cultivated. When the authenticity of content cannot be established with a sufficient level of confidence, the precautionary principle remains the best option: suspend sharing, wait for independent confirmation, or rely on more in-depth analysis. In this age of immediacy, rigour, patience and analysis are essential tools for limiting the spread of misleading content.

By Alwine Morel

Sources :

https://news.artnet.com/art-world/adobe-content-credentials-cr-icon-2376669
https://www.lamontagne.fr/paris-75000/actualites/comment-reconnaitre-une-video-generee-par-ia-sur-les-reseaux-sociaux_14816547/
https://www.lemonde.fr/videos/video/2025/12/29/comment-detecter-une-video-generee-par-intelligence-artificielle-comprendre-en-3-minutes_6659693_1669088.html
https://www.blogdumoderateur.com/guide-detecter-video-ia/
https://contentcredentials.org/
https://deepmind.google/models/synthid/
https://www.lemonde.fr/pixels/article/2025/12/15/cinq-conseils-pour-reconnaitre-une-video-generee-par-ia-sur-les-reseaux-sociaux_6657458_4408996.html#:~:text=Cherchez%20les%20%C2%AB%20watermarks%20%C2%BB,sur%20les%20r%C3%A9seaux%20sociaux%20fran%C3%A7ais).
https://sightengine.com/
https://www.ladn.eu/tech-a-suivre/ia-machine-learning-iot/nouveau-job-detecteur-de-deepfakes/
https://sensity.ai/
https://www.realitydefender.com/