A Photo Is No Longer Worth a Thousand Words
For decades, the photographic image enjoyed an implicit credibility: seeing is believing. Photographs were evidence. Videos were testimony. That era is coming to an end.
In 2026, artificial intelligence models generate images, voices, and videos of real or entirely fictitious people with a fidelity that, in many cases, exceeds the human ability to tell real from fake. This is not a science-fiction hypothesis: it is the state of the world as it stands.
The question is no longer merely technical — how do you detect a deepfake? — but deeply social and political: what happens when no one can trust what they see?
Deepfakes: From Curiosity to Disinformation Infrastructure
The term deepfake first appeared in 2017, on obscure online forums, to describe pornographic videos featuring fabricated celebrity faces. In under a decade, the technology moved from the internet's basement into mainstream consumer tools, election campaigns, financial fraud, and military conflicts.
In 2024, an employee at a Hong Kong company transferred the equivalent of $25 million following a video conference with what he believed were his colleagues — all of them real-time deepfakes. That same year, several countries saw fabricated videos circulate showing political leaders announcing decisions of war or surrender.
This is no longer fringe manipulation. It is a disinformation infrastructure, available for a few euros a month.
What AI Cannot (Yet) Simulate
Faced with this surge in capabilities, researchers have sought biological signals that AI cannot easily replicate. Among the most promising: remote photoplethysmography, or rPPG.
The principle is subtle but powerful. When the heart beats, blood pulses beneath the skin and slightly alters the color of facial pixels — a variation imperceptible to the naked eye, but measurable algorithmically. A real person captured on video displays this regular signal. A deepfake, generated frame by frame without modeling human physiology, ignores it entirely or reproduces it inconsistently.
In the same vein, researchers are studying micro-facial expressions — involuntary facial movements lasting less than 0.01 seconds. Eye blinks, subtle muscle contractions, reflex flushing: signals that diffusion models struggle to generate because they reason frame by frame, without fine temporal coherence.
These biological approaches are not foolproof. But they illustrate a key principle: the most robust detection relies on what life produces and what computation cannot yet feign.
The Arms Race Between Generation and Detection
The fundamental problem with AI detection is its reactive nature. Tools like Hive Moderation, Illuminarty, or the internal classifiers used by major platforms are trained on images generated by known models. As soon as a new model is released — Midjourney v6, Sora, an open-source tool — its specific artifacts are not yet recognized.
This is an asymmetric arms race: generating is becoming ever easier and faster; detecting demands time, data, and must start over with each new model generation.
Some researchers speak of spectral drift: each architecture generates characteristic artifacts in an image's high-frequency ranges, visible via Fourier transform. These signatures can sometimes identify the source model. But newer architectures are reducing them — and users are learning to erase them with post-processing tools.
C2PA: Provenance Rather Than Content
Faced with the impossibility of winning the race on the content front, a different approach is emerging: certifying the chain of provenance rather than analyzing pixels.
This is the ambition of the C2PA standard (Coalition for Content Provenance and Authenticity), backed by Adobe, Microsoft, Intel, the BBC, and others. The concept: every image or video file embeds a cryptographically signed manifest recording its origin (which device, which software) and all successive modifications.
Leica launched a camera (the M11-P) in 2023 with C2PA built in at the hardware level. Adobe Lightroom and Photoshop automatically sign exports. Certain broadcast cameras are beginning to adopt the standard.
C2PA is not a silver bullet. It is a trust infrastructure that works when every link in the chain adopts it — which is far from the case today. But it is one of the few approaches that does not depend on a technical race between generation and detection.
The Liar's Dividend: Deepfakes' Unexpected Gift to Liars
There is a paradox at the heart of the deepfake crisis that few anticipated: the spread of fakes also benefits authentic liars.
This phenomenon, theorized by researchers Bobby Chesney and Danielle Keats Citron under the name Liar's Dividend, is straightforward to understand. If everyone knows that perfectly realistic videos can be fabricated in minutes, then every video becomes suspect — including real ones.
A politician filmed committing a crime can now cry "deepfake!" — and plant enough doubt to avoid consequences. An authentic audio recording can be declared synthetic. A real photograph can be dismissed as AI-generated.
The Liar's Dividend may be the most serious risk posed by deepfake technology — not the fake that passes as real, but the real that can be made to pass as fake. It threatens the very foundations of journalistic and judicial evidence.
The Legal Framework: Between Transparency and Prohibition
Legislators have begun to respond. The European Union adopted the Artificial Intelligence Regulation (EU AI Act) in 2024, the world's first comprehensive legal framework on the subject.
For deepfakes and synthetic content, the law distinguishes several levels:
- Unacceptable risk: AI used to manipulate people without their knowledge or exploit vulnerabilities — prohibited.
- Limited risk: deepfakes and non-artistic synthetic content — permitted under a condition of transparency. The end user must be clearly informed that they are viewing AI-generated content.
- Exception: art, satire, and fiction, provided they are explicitly identified as such.
The AI Act applies to any company marketing AI products in the EU, regardless of where they are based. Its penalties can reach €35 million or 7% of global turnover.
In the United States, legislation is fragmented: some states have passed specific laws on electoral or pornographic deepfakes, but no unified federal framework exists as of 2026.
What We Can Do Today
Faced with this reality, three habits have become essential:
- Question the source, not just the image. Where was this video first published? Who is sharing it, and with what motive? Viral content with no verifiable source always warrants caution.
- Look for corroboration. A real event is typically covered by multiple independent sources. An AI-generated image is not. The context around an image matters as much as the image itself.
- Use verification tools. Reverse image search (Google Images, TinEye), AI detection tools (Hive, Illuminarty), and Adobe Content Credentials can often clarify the origin of a piece of content.
Education: The Only Scalable Defense
Detection technologies will evolve. Laws will be imperfectly enforced. C2PA will take years to become widespread. In the meantime, the most effective and durable defense is education in the broadest sense — not memorizing a checklist of technical signals, but cultivating a systematically critical posture toward online content.
This is what we call synthetic literacy: the ability to navigate a world where a growing fraction of visual and audio content is generated, altered, or optimized by algorithms — without falling into paranoia or naivety.
This literacy is not acquired in a single reading. It is cultivated, practiced, and tested. That is why tools like Fake or Real exist: not to frighten you, but to train you to doubt intelligently.
Conclusion: Trust as a Civilizational Challenge
The problem of digital post-truth is not, at its core, a technological problem. It is a problem of trust — in institutions, in media, in images, in one another.
The tools to address it are technological (C2PA, rPPG, forensic detection), legal (AI Act, national regulations), and cultural (media literacy, fact-checking journalism). None will suffice alone.
What we can do, individually, is refuse two symmetrical temptations: believing everything because it is comfortable, and rejecting everything because it is simpler. Truth exists. It has simply become harder to reach — and that is precisely why it deserves more effort.