Deepfakes and misleading clips: how to check a political clip before you share it (method, tools, law)
In 2025 the EU’s diplomatic service counted 540 foreign information-manipulation incidents, France the second most targeted country, and 27 % of them used AI. Most fake political clips are cruder than that: a real sentence cut short, a subtitle rewritten, a date moved. Here is how to check a clip in five minutes, which tools help and where they stop, and what the law says in France, as of 24 October 2026.
Contents
Deepfake, cheap fake, out of context: three ways a clip lies
The word deepfake hides the simplest cases. The guide the French government published for campaign teams in December 2025 defines it as an audio or video manipulation made from existing material, using artificial intelligence to change a person’s face in a video or reproduce their voice. That is one family. The second contains no AI at all: a real video slowed down, cut before the end of a sentence, overdubbed or re-subtitled, what specialists call a “cheap fake”. The third is a real, untouched clip presented with a false source, place, date or meaning. X’s rules name all three in one definition: “inauthentic media, including, manipulated, or out-of-context media that may result in widespread confusion on public issues”, and list “modified subtitles” among the edits that fundamentally change “the understanding, meaning, or context of the media”.
How common is it? The only official counts are European. In its fourth threat report, published in March 2026, the European External Action Service detected 540 foreign information manipulation incidents in 2025, Ukraine first and France second with 107; “27% of the incidents detected by the EEAS involved AI-related TTPs”, up from 41 cases in 2024 to 147, and “the use of synthetic audio is no longer experimental”. The same report adds the nuance that matters for your feed: “much of the AI-generated material in 2025 remains low-quality”, with “limited overall impact as organic engagement remains low”. The French monitoring service Viginum, in its February 2025 report on AI and information manipulation, relays the Swedish Psychological Defence Agency’s warning about the “liar’s dividend”: when fakes are everywhere, even authentic content can be dismissed as fake. And for scale: in 2022, the commission that supervised the presidential campaign reported no deepfake at all; its final report mentions false polls and “infox”, nothing about manipulated video. The subject is new for 2027.
The five-minute method before you share
- Find the source eventA political clip comes from somewhere: a debate, a rally, an interview, a parliamentary session. Search for the full replay on the broadcaster’s or the candidate’s official channel and find the passage. If you cannot find the event at all, that is your first signal. How these extracts are produced and spread is in our piece on debate and rally clips.
- Look at the cutWhere does the clip start and stop? A sentence cut before its “but”, a question removed, a reaction placed after a different statement: compare the clip’s first and last words with the replay. Twenty seconds of context often reverse a meaning.
- Read the subtitles against the audioBurned-in subtitles are the cheapest manipulation: mute the clip and read, then listen with the subtitles hidden. A changed number, a dropped negation, a swapped name. X’s policy explicitly lists “modified subtitles” among manipulations.
- Inspect image and soundLips against the words, hands and jewellery, text in the background, the room’s sound against the voice, a voice that never breathes. The InVID-WeVerify browser plugin, designed and maintained by AFP Medialab, extracts keyframes from a video and sends them to reverse image search engines, reads metadata and applies forensic filters. Google’s “About this image” shows when Google first saw an image, how other sites use it and, when the information exists, whether it was “captured with a camera or generated with AI”; it is “currently only available in some regions”.
- Check provenance and watermarksDrag the file into the Content Credentials Verify tool (the C2PA standard, backed by more than 500 companies): if the file carries credentials, you see who made it and how it changed over time. The tool itself warns that credentials “are still rolling out”, and platforms often strip metadata, so their absence proves nothing. Google’s SynthID Detector, open to everyone since 7 October 2026, finds the watermark that Google and its partners (OpenAI, NVIDIA, Kakao, soon Apple) embed in their AI output: a negative result means “no SynthID watermark found”, not “authentic”.
- Decide, then report rather than argueIf two checks fail, or one is impossible, do not share. If the clip is manipulated, report it on the platform (the steps are below) and, if you are a newsroom or a campaign, document the checks: source link, timecodes, screenshots.
Four tools, and the line each one stops at
| Tool | What it tells you | Where it stops |
|---|---|---|
| Content Credentials Verify (C2PA) | The provenance and edit history embedded in a file; 16 formats including MP4, MOV, JPEG, PNG, WAV and MP3 | Credentials are “still rolling out”: most files have none, and no credentials is not a verdict |
| SynthID Detector (Google DeepMind) | Whether the file carries a SynthID watermark, with the portions most likely watermarked highlighted | Only content from Google and its partners; a negative result is not proof of authenticity |
| About this image (Google) | When Google first indexed the image, how other sites use it and, for some images, whether it was captured or AI-generated | Images, not video; “only available in some regions” |
| InVID-WeVerify plugin (AFP Medialab) | Keyframes from a video, reverse image search on several engines, metadata, forensic filters | A toolbox that needs a reader: it surfaces clues, it does not rule |
What French law says (this is not legal advice)
Two texts matter for a shared clip, and a third is on its way.
- Undisclosed montages and AI content: article 226-8 of the Criminal Code. Rewritten by the law of 21 May 2024, it punishes with one year in prison and a 15,000 € fine the publication of a montage made with a person’s words or image without their consent, “if it is not obvious that it is a montage or if it is not expressly mentioned” (our translation), and treats the same way visual or audio content “generated by algorithmic processing” that depicts a person’s image or words without consent, unless obviously or expressly labelled; the penalties rise to two years and 45,000 € when an online service is used. A labelled parody is therefore outside the offence; an unlabelled one is not. Sexual deepfakes have their own article, 226-8-1: up to three years and 75,000 € online.
- False information before an election: article L. 163-2 of the Electoral Code. During the three months before the first day of the month of a general election, and until the vote is acquired, when inaccurate or misleading allegations of a fact likely to alter the sincerity of the upcoming vote are spread online “in a deliberate, artificial or automated and massive manner”, the interim judge can order platforms to stop the spread; the judge rules within forty-eight hours, and only the Paris court hears these cases. The law of 1962 makes the article applicable to the presidential election. For a first round on 18 April 2027, the window opens on 1 January 2027: that date is our calculation from the text, not a published one.
- A bill under way. On 20 October 2026 the Senate adopted, at first reading, a bill against foreign interference in democratic life that repeals L. 163-2 and replaces it with an article L. 48-3, creates a referral procedure outside election periods for false information plausibly spread by or for a foreign power, and raises the penalties of article L. 97 to three years and 45,000 €. The National Assembly still has to examine it; nothing in it applies yet.
European rules add a labelling duty: in a written answer published on 5 February 2026, the ministry in charge of AI reminded that under the AI Act, deployers must indicate that content has been generated or manipulated by AI, a requirement applicable from 2 August 2026. The same answer points campaign teams to the Viginum guide of December 2025 on protecting public debate online.
Reporting a manipulated clip on each platform
- YouTube. Below the video, click More, then Report, and select the reason that best fits. YouTube says reported videos are reviewed around the clock, and that a report does not take a video down automatically.
- TikTok. Go to the post, tap Share or press and hold the post, tap Report, choose a reason and a subtopic if asked, then Submit; on the web, hover over the “More options” button at the top of the video.
- Instagram. Click Options at the top right of the post, select Report and follow the instructions; “False information” and “Scam, fraud or spam” are among the reasons. Reports are anonymous, except for intellectual property claims.
- X. Open the post, tap the icon at its top, select “Report post” and follow the steps. X’s Authenticity policy prohibits media “depicting a real person that has been fabricated or simulated, especially through use of algorithms or broader artificial intelligence”, and media “presented with false or misleading context surrounding the source, location, time, or authenticity of the media”.
If you make political clips yourself
The same rules, read in reverse. Keep the cut on complete sentences and keep the question with the answer; subtitle from the audio, never from what you wish had been said; label any edit that is not a plain cut; keep the source link and the timecode in the description, so that anyone can run step one of the method on your clip. Clipping tools can help or hurt here: at flysub, every clip starts and ends on a complete sentence and the subtitles are transcribed from the audio word by word, so what is on screen is what was said.
Frequently asked questions
How can I tell whether a political video is a deepfake?
Start with the source: find the full replay and the passage. Then look at the cut, read the subtitles against the audio, inspect lips, hands, background text and room sound, and run the tools: Content Credentials Verify for provenance, SynthID Detector for Google-family watermarks, About this image and the InVID-WeVerify plugin for reverse search. Most fakes fail the first two checks before any tool is needed.
Is there a tool that detects deepfakes with certainty?
No. Each tool covers a perimeter: Verify reads credentials that most files do not carry yet, SynthID Detector only finds the watermark of Google and its partners, About this image works on images in some regions, InVID-WeVerify surfaces clues for a human reader. A negative result from any of them is not a proof of authenticity.
Is sharing a deepfake illegal in France?
Article 226-8 of the Criminal Code, rewritten in May 2024, punishes publishing a montage or AI-generated content depicting a person’s image or words without consent, unless the manipulation is obvious or expressly mentioned: one year and 15,000 €, two years and 45,000 € online. Sexual deepfakes fall under article 226-8-1, up to three years and 75,000 € online. This is a summary, not legal advice.
What is the 48-hour electoral referral?
Article L. 163-2 of the Electoral Code: in the three months before the first day of the month of a general election, when false information likely to alter the vote is spread online in a deliberate, artificial or automated and massive manner, the Paris interim judge can order platforms to stop it within forty-eight hours. A bill adopted by the Senate on 20 October 2026 would replace it with a new article L. 48-3.
How do I report a fake political clip?
On YouTube: More, then Report under the video. On TikTok: Share or long press, then Report. On Instagram: Options at the top right, then Report. On X: the icon at the top of the post, then Report post. Each platform asks for a reason; X’s rules explicitly cover fabricated or out-of-context media.
Cover photo: Thomas Bresson, CC BY 4.0, Wikimedia Commons.
