How to Spot a Deepfake Video: A 5-Step Verification Routine

Your phone buzzes in the family group. A 43-second clip shows a well-known figure saying something shocking, with a message underneath: "Share this before they delete it." Your heart is racing and your thumb is hovering over "Forward".

Before you tap, ask yourself: can I tell by looking that this is fake? The honest answer in this article is that your eyes are no longer enough, and what protects you is a short routine you run before sending. It draws on the methods journalists and fact-checkers use, and on my background in information security. Everything here is current as of October 2026, and every figure or incident has its source listed at the end.

Before you forward: you are part of the channel

When you forward a clip to family or colleagues, you are not just passing it on. You are passing it on in your name. The person who receives it believes it because you sent it, and doesn't ask where it came from. That turns you from a viewer into a link in the chain of spread.

Checking costs a few minutes. Correcting a rumour that has reached dozens of groups may never happen. That trade alone is reason enough to wait.

Why the old visual tells no longer work

You may have read advice like "watch the blinking", "check the lip sync", "look at the fingers and the edges of the face". Those tells were useful when the tools were weak, and they still show up in sloppy fakes. But their absence doesn't make a clip real, and the research supports that caution.

A systematic review and meta-analysis published in 2024 in Computers in Human Behavior Reports pooled 56 papers on how well people spot high-quality deepfakes, with 86,155 participants in all. It found an average accuracy of 55.54%. The 95% confidence interval around that figure, 48.87% to 62.10%, includes 50%, so performance was not statistically better than guessing. For video alone the figure was 57.31%, with an interval of 47.80% to 66.57% that also includes 50. One fairness note: the searches were run in June and October 2024, so the numbers don't reflect the newest generators.

In February 2025 the company iProov announced an experiment with 2,000 consumers in the UK and US who had been primed to look for fakes. Even so, only 0.1% classified every item correctly, real and fake. People were 36% less likely to correctly identify a synthetic video than a synthetic image, and more than 60% stayed confident in their skills despite their poor performance. Keep in mind that iProov sells identity-verification technology and the source is a press release that doesn't detail its methodology, so treat these numbers as an indication, not a verdict.

Experts are no exception. In July 2026 WBUR published a summary of a Here & Now segment on Hany Farid, one of the best-known names in deepfake detection. It says he no longer trusts his own eyes when verifying images and video, and that fakes are now getting past tools he built himself.

Don't ask "does it look real?" Ask who said it, who confirms it, and where it came from.

A case from our region: "leaked" recordings in a cloned voice

This isn't a distant problem. In October 2023 the BBC reported on an anonymous TikTok account that had been posting, since late August, recordings it called "leaked" audio of Sudan's former president Omar al-Bashir. The investigation concluded the voice was fake. The clips drew hundreds of thousands of views. According to The Decoder's account of the investigation, at least four recordings were cut from a blogger's live broadcasts, with no evidence that he took part in the campaign. TikTok suspended the account after the BBC alerted it.

Notice that one thread of the unmasking came from tracing the source material, not from a sharp ear. And notice the second danger, raised by a researcher at Northeastern University's Civic AI Lab: that such recordings can push people to doubt even genuine ones.

The five-step verification routine

One of the simplest verification frameworks is SIFT, published by Mike Caulfield in 2019: four moves, namely stop, investigate the source, find better coverage, and trace claims, quotes and media back to their original context. I kept its skeleton and added a step specific to synthetic media, which makes five. You don't need an expert eye, just a few minutes and a phone.

Tools you can use, and what they can't do

Reverse image search

This is the most useful tool and the simplest. Grab a frame from the clip and upload it to Google Lens, TinEye or Yandex, or use the InVID extension, which splits a video into key frames and lets you search them across several engines. In March 2026 the fact-check unit of Australian Associated Press (AAP) examined clips circulating about a regional military event. A video presented as cars burning "right now" turned out to be real footage from October 2023, found through a Google reverse image search. The problem wasn't artificial intelligence. It was an old clip with a new date.

Here is the limit you should know: reverse search catches recycled footage, not a brand-new fake. A clip a machine generated yesterday has no earlier copy, so finding nothing doesn't mean it's real.

The InVID-WeVerify extension

This is a Chrome extension developed by AFP Medialab R&D, a team at the French news agency AFP, across three EU-funded research projects, and its interface is available in Arabic. According to its Chrome Web Store listing (version 0.93, last updated on 28 July 2026), it splits video into key frames, reads video and image metadata and offers forensic tools for images. It also has a synthetic-image detector and a voice-clone detector, both in beta and reserved for registered users. It is built for journalists and fact-checkers, and a 2019 Amnesty International guide said its forensic tab is not recommended for beginners. So use the parts you understand: splitting video and searching.

SynthID Detector

Google DeepMind's SynthID page (as of October 2026) says anyone can use synthid.com to check whether an image, video or audio file was made with AI from Google or from partners the page names: OpenAI, NVIDIA and Kakao, with Apple "soon". It adds that Gemini, Google Search and Chrome can check too when you upload a file. In the AAP fact-check above, a SynthID mark turned up on images, and a Gemini logo was visible on some of them.

But the limit is plain: the tool looks for one specific watermark. A clip made by a tool that doesn't add it returns "not detected", and that result doesn't mean the clip is genuine.

"AI detector" websites

Be wary of any site that hands you a ready-made percentage such as "92% generated". The Deepfake-Eval-2024 benchmark, an arXiv paper built on media that circulated in 2024 and that users had flagged as possibly fake, found that the performance score (AUC) of open-source detectors fell by about 50% on video, 48% on audio and 45% on images, compared with their results on older academic benchmarks. Commercial models do better, but the researchers say they still fall short of professional forensic analysts. Treat a detector's answer as one more signal, not a ruling.

Provenance signals: what lies behind the picture

The industry's direction is to stop judging the picture and judge its record. The C2PA standard, known as Content Credentials, attaches a signed record to a file showing how it was made and what edits were applied, and its backers compare it to a nutrition label on food. By its own website in October 2026, its steering committee includes companies such as Adobe, Amazon, Google, Meta, Microsoft, OpenAI, Sony and TikTok, alongside the BBC.

How far has it got? In September 2025 Google said Pixel 10 phones are the first to attach this data to every JPEG photo taken with the Pixel Camera, and that Google Photos shows it in the About panel. The Pixel Camera reached Assurance Level 2, the highest security rating defined in C2PA's conformance programme at the time. In May 2024 TikTok announced it reads this data to automatically label AI-generated content from some other platforms, and acknowledged that this depends on the data being present in the file.

Which is the catch. A January 2026 post from the Content Authenticity Initiative says platforms still often strip metadata, which is why it is working on "durable" versions that combine metadata with a watermark and a fingerprint. So a clip with a genuine origin can reach you with no trace of its signature. The rule for you: valid data is strong evidence, and its absence carries no meaning. Google itself says in its Pixel announcement that these credentials are not the only solution, and that labelling only AI content may lead people to assume anything unlabelled is real.

If someone fakes you or someone you love

When your voice or face, or a relative's, is being impersonated, whether in a posted clip or in a call from a "relative" asking for money, the question changes from "should I believe this?" to "how do I protect myself?". The steps are simple:

• Don't trust the voice alone. If a "relative" calls asking for urgent money, hang up and call them on the number you already have. The US Federal Trade Commission advises not to trust the voice, to contact the person directly, and, if you can't reach them, to try another family member or their friends.
• Agree a family code word. In a December 2024 alert the FBI recommends a secret word or phrase among family members to verify identity, and limiting how much of your voice and image you post publicly. It only works if you agree it before the crisis.
• Don't rush to pay. Scammers usually ask for transfers, crypto or gift cards, which the FTC says are hard to get money back from. The FBI warns against sending money, gift cards or crypto to people you haven't met in person, especially anyone you know only from online or phone contact.
• Document before you report. Save the link, account name, posting date and a screenshot, then report to the platform. If there is blackmail or defamation, report it to the cybercrime unit in your country.
• Correct it for those who saw it. Send a short correction through the same channel to everyone who got the clip, and don't wait for it to die down on its own.

The bottom line: move your trust from the look to the origin

The eye is no longer a sufficient witness, but that doesn't mean doubting everything. It means shifting your trust from "it looks real" to "who said it, who confirms it, and where did it come from". Anyone can ask those questions in minutes, and they work on the newest generators as well as the oldest.

It is close to what the book The Blind Manager (Arabic) is about: the difference between seeing and not seeing lies not in sharp eyesight but in insight, in looking past what is put in front of you. The people who need that insight first are the ones who forward clips to their team or family. If you haven't confirmed it, don't send it.