Video used to feel like a witness you could replay. If the clip existed, the argument seemed settled: the camera had been there, so what more was there to discuss?

Quite a lot, as it turns out. A convincing clip may be synthetic, edited or taken from another event. A genuine recording can now be dismissed with two words: “That’s AI.”

Video has not become worthless. It has become less self-explanatory. Seeing is a lead, not a verdict.

Why this matters⌗

The trouble is not only that computers can make fake people speak. We ask one small file to answer several questions at once: Is it genuine? Did the event happen? Is this the original? What was cut? Does it prove the claim attached to it?

Those are different questions. A real clip can have a false caption or omit the moment that explains it. A synthetic video can also be clearly labelled and cause no deception.

This matters in elections, journalism, courts, war reporting, scams and family arguments. The stakes vary, but the question is the same: what does this evidence actually earn us the right to say?

Brazil is writing rules for a problem technology cannot settle⌗

Brazil makes the problem concrete because its electoral authority has had to define it in legal rather than purely technical terms.

On 1 September 2026, Brazil’s Superior Electoral Court, the TSE, set criteria for deepfake content in the 2026 election context. Its definition covers realistic or plausible synthetic or AI-manipulated material that creates, reproduces or changes the image, voice or expression of a living, dead or fictional person.

The legal qualification matters: the TSE said the prohibition requires the material to be characterised as electoral propaganda. “Made with AI” is not the whole legal answer.

The decision concerned a video shown during a party convention that recreated the image and voice of Jair Bolsonaro, Brazil’s former president. The TSE rejected a request for a fine against Flávio Bolsonaro, the presidential candidate whose campaign the video supported, considering the closed setting and the absence of an explicit request for votes.

The TSE’s April guidance also covers disclosure of synthetic multimedia in campaign propaganda. These are Brazilian electoral rules, not a global definition of truth. They show why regulation has to consider audience, context, purpose and timing, not just pixels.

The United States has a patchwork, not one answer⌗

The National Conference of State Legislatures (NCSL), which tracks state-level policy, published a snapshot on 23 June 2026 saying 31 U.S. states had enacted laws regulating deepfakes in political messaging. It describes prohibitions and disclosures: Minnesota and Texas prohibit certain deepfakes during defined pre-election periods, while Maryland prohibits deceptive election-related deepfakes without the same time limit. The other 28 states in the snapshot use disclosure requirements, meaning the political material must tell viewers when synthetic media has been used.

Definitions, timing, remedies and complaint routes differ. “The U.S. has banned deepfakes” is therefore about as useful as saying “the U.S. has a rule about sandwiches”. It gestures towards something real while leaving out what matters.

Four questions hidden inside “Is this video real?”⌗

1. What is the file?⌗

Is it an original camera file, a screen recording, a cropped repost, an edited export or a synthetic creation?

“Authentic” does not mean untouched. Newsrooms edit, platforms compress and people add captions. Those changes may be ordinary; the question is whether they alter the claim. Find the earliest accountable version rather than treating the most popular copy as the source.

2. Where did it come from?⌗

Provenance is the trail behind the file: who recorded it, who published it, how it travelled and whether anyone can explain the chain.

An anonymous account may share a real video; a famous account may share a misleading one. Look for the date, place, original uploader and reporting that existed before the clip went viral. If the caption is doing all the explaining, the footage may be genuine while the story around it is not.

3. What changed or was left out?⌗

Cuts, crops, dubbing, speed changes, overlays and missing audio can all affect meaning. A video does not have to be entirely fabricated to mislead.

A six-second clip of someone shouting may omit the question, the sentence that followed or the longer exchange that changes the tone. Editing is not the same as lying; it is a reason to ask what became easier to miss.

In journalism and war reporting, date, location and sequence still need checking. Repeated copies are not independent corroboration. Ten accounts repeating the same caption are one unverified claim wearing ten hats.

4. What does it actually prove?⌗

A camera records an image and sound. It does not automatically record motive, background, causation or the entire event.

A genuine video may prove that a person said a sentence in a room. It may not prove why they said it, what happened before, or what the caption claims they intended. The question is whether the claim is bigger than the footage.

Provenance can help without becoming a truth machine⌗

Content Credentials, built around the C2PA open standard, are a way of attaching provenance details to digital media. In plain English, they can record where an asset came from and what happened to it, including work on live video in C2PA’s 2.3 material.

That is useful, but provenance is not truth. Credentials can support a chain of custody without proving the event happened as described. Missing credentials do not prove fabrication: metadata can be stripped, lost during upload or never added.

Think of provenance as a label on the evidence box, not the judge’s ruling about what is inside it.

The liar’s dividend: when “AI” becomes an escape hatch⌗

Researchers use “liar’s dividend” for the advantage someone can gain from a world full of convincing synthetic media. Once people know fake video exists, someone confronted with genuine evidence can claim it is fake. The denial only needs to create enough doubt for the audience to disengage.

Analysis from the Brennan Center and the Brookings Institution—research organisations, not election-enforcement bodies—describes this danger, including the possibility that public figures benefit from dismissing real scandals as deepfakes. That is research about a social effect, not proof that a particular denial is dishonest. The excuse works because the technology is real.

A better response is to ask who is making the claim, what independent evidence exists and whether the dismissal supplies anything beyond the word “AI”.

Courts, journalism and war footage still need a chain of reasoning⌗

Courts already separate two questions: is this item what its proponent says it is, and how persuasive is it? In the United States, Federal Rule of Evidence 901 addresses the first question. Authentication is not an automatic ruling that the evidence proves the whole case.

The U.S. Advisory Committee on Evidence Rules—the committee that studies possible changes to the federal evidence rules—discussed a possible Rule 901(c) response to deepfake challenges in November 2025. It was a proposal under discussion, not an adopted new rule.

Journalists face the same practical distinction outside a courtroom. They need to identify the source, test the date and location, compare the footage with other reporting and say what remains unresolved. War footage deserves more caution, not less, because the cost of a wrong caption can be measured in fear, propaganda and real-world danger.

A two-minute pause before sharing⌗

Before forwarding a dramatic clip, try this short pause:

  1. Stop: dramatic is not the same as verified.
  2. Find the source: look for the earliest accountable uploader or publisher.
  3. Check time and place: search for older versions, different events and distinctive words from the caption.
  4. Look at the whole sequence: note cuts, crops, dubbing, overlays, speed changes and missing context.
  5. Find independent support: seek separate footage, documents, witnesses or reputable reporting.
  6. Shrink the claim: say only what the evidence supports. “Unverified” is sometimes the most accurate sentence available.

Do not outsource the decision entirely to an AI detector. A score can be a clue, not a court of appeal. Nor should you rely on folklore about blinking, fingers, teeth or a face looking odd after compression.

Dadbot take⌗

We do not need to choose between gullibility and permanent cynicism.

The old habit was: “There is a video, so it must be true.” The new habit should not become: “There is a video, so it might all be fake.” Both are shortcuts. One trusts the pixels too much; the other lets anyone escape the evidence by saying “AI”.

The more useful habit is smaller and less dramatic. Treat video as a lead. Check where it came from. Work out what changed. Compare it with independent evidence. Then make the claim no bigger than the footage can carry.

Final thought⌗

Seeing is still valuable. It can show us something that would otherwise be invisible, preserve a moment and prompt questions that matter.

It just cannot do all the reasoning for us anymore.

When someone says, “But there’s video,” the best answer is no longer “Then that settles it.” It is: “What does the video actually prove?”

That question will not make the internet peaceful. It may stop it becoming quite so easy to fool.

Sources and caveats⌗

Research checked 8 September 2026. Brazil’s TSE rules, the NCSL state-law count and U.S. evidence material are jurisdiction-specific and date-stamped. The article does not treat Content Credentials or AI detectors as conclusive proof.