Deepfakes before the judge: how a video is challenged —and authenticated— when any image can be a lie
For much of the last century, a photograph or a video arrived in a legal proceeding with an extraordinary psychological advantage: if it was recorded, it happened.
Generative artificial intelligence has put an end to that comfort.
Today a false piece of evidence can look authentic, and at the same time, a genuine piece of evidence can be discredited simply by claiming it was generated or manipulated using AI. This second phenomenon is known as the liar's dividend.
AI doesn't just make it easier to fabricate false evidence. It also makes it easier to deny genuine evidence.
The LEC already had the tools
The Civil Procedure Act (LEC) has regulated means of reproducing words, sound, and images since 2000.
Article 299.2 recognises them as means of evidence, and Articles 382 to 384 set out their regime.
The party proposing a recording may accompany it with expert opinions and instrumental means, and the opposing party may submit its own when it disputes authenticity or accuracy. The court weighs the evidence according to its sound judicial discretion (sana crítica).
The procedural architecture was already there. What has changed is the technical burden of sustaining it.
Authenticity has not legally stopped being presumed
There is no new rule under which every photograph must be considered false until an expert report proves otherwise.
But in practice, it is no longer prudent to rely solely on the appearance of authenticity.
The more important the medium, the greater our ability must be to explain who obtained it, how, which file is the original, what has happened to it since, and what external elements corroborate what it shows.
In the age of the deepfake, the provenance of the image has itself become evidence.
Four distinct concepts
Provenance: where does the file come from? Who captured it? With what device? Under what circumstances?
Integrity: is the file submitted the same one that was originally preserved, or has it undergone changes?
Authenticity: is there sufficient evidence to support that it corresponds to the fact, people, moment, and context claimed?
Evidentiary truthfulness: does what appears in it actually prove the legal fact we are trying to establish?
A file can be perfectly intact and still contain a staged scene. And an authentic file can prove far less than the party submitting it claims.
First line of defence: documenting the origin
Here, detective evidence has a structural advantage.
Behind the camera is an identified person. A licensed professional who can appear in court and say: "I was there."
They can identify the place, time, position, device, conduct before and after, continuity of observation, and material obtained.
Compare that with a video forwarded through a messaging app from an unknown source. Visually, they may look the same. As evidence, they are not.
Preserving the original file
The native file is preserved.
No one works on the sole original. It is not replaced by the version sent over messaging. Only the interesting excerpt is not the only thing kept.
The file coming from the device is preserved, and working copies are generated.
Every transformation can strip out metadata, encoding information, container structure, or technical data useful for a later analysis.
The hash: useful, but not magic
A cryptographic digest produces a mathematical fingerprint of the file. If a single bit changes, the fingerprint will be different.
The hash can therefore prove that a given file has not been modified since the moment that fingerprint was calculated.
But it does not prove that the content is true.
If someone generates a deepfake and immediately calculates its hash, we end up with a perfectly intact deepfake.
The hash answers "has this file changed?"; on its own, it does not answer "did what it shows actually happen?"
Traceability and the chain of provenance
It must be possible to explain when the file was incorporated, who obtained it, who received it, where it was stored, what copies were produced, who had access, and what transformations were carried out.
But in the synthetic era we need an earlier question: how was this file born?
Who created it. What device was involved. What software. What transformations. What original exists. What relationship there is between the file and the physical fact.
To the chain of custody we must add a genuine chain of provenance.
Corroborating the image against reality
Suppose a seemingly perfect video surfaces.
The question should not only be "does it look manipulated?" but also "what do we know independently of this video?"
Where was that person? Who else was present? Is there another camera? Are there witnesses? A receipt? A vehicle? A contemporaneous document?
A deepfake can fabricate a convincing image. What is far harder is fabricating, retrospectively, the entire reality that should surround it.
How to challenge a suspicious video
First, determine what file is being submitted: original, export, social-media copy, screen recording, forward.
Second, carry out forensic analysis when it is central to the litigation: metadata, structure, compression, discontinuities, audio, synchronisation, signs of editing, and other indicators.
Third, investigate the context: locations, witnesses, activity, timelines, and other independent evidence.
The expert report asks "what is this file?" The investigation asks "did what this file claims to show actually happen?"
No detector should be an oracle
Deepfake-detection systems are competing against generators that evolve rapidly.
That is why a serious expert report should not promise an infallible machine that says "true" or "false."
Often the rigorous conclusion will be: "No indicators consistent with certain forms of manipulation have been identified," or conversely: "Inconsistencies have been detected that do not match the claimed provenance."
The best authentication will be multi-layered: technical analysis + provenance + custody + testimony + external corroboration.
The European AI Act
As of 2 August 2026, the transparency obligations of Regulation (EU) 2024/1689 apply generally.
Article 50 addresses synthetic content: it provides for machine-readable identification mechanisms for certain generated or manipulated content and disclosure duties in relation to deepfakes, along with their conditions and exceptions.
But the AI Act does not create a procedural presumption whereby "no label = true" or "no label = false."
The judge will still weigh the evidence, and we will still have to prove its provenance.
The detective's advantage in the synthetic era
When a detective properly documents a fact, there is something an anonymous image will never have: a witness to the birth of the evidence.
They know where they were, what they saw, when the recording began, what happened before and after, what device they used, and how they preserved the file.
Artificial intelligence can fabricate plenty of pixels. What it cannot fabricate retrospectively with the same ease is a complete, coherent, and verifiable chain of reality.
Conclusion
For decades we asked "what does the video show?"
AI forces us to start with a different question: "why should we believe this video comes from where it is claimed to come from?"
We need origin, an original, integrity, traceability, context, corroboration, and, when necessary, expert analysis.
In a world where a perfect image can be created in seconds, the most valuable product will no longer be the spectacular photograph.
It will be evidence whose complete history can be explained and proven.
Related reading: La IA: Asedio a la privacidad — Editorial Tirant Lo Blanch
Truth has a method.