AI-generated videos are getting harder to identify by sight alone. Faces can appear natural, voices can sound familiar and synthetic clips are increasingly showing up in scams, impersonation attempts and misleading posts. As the old visual clues become less reliable, verification is moving beyond simply asking whether a video “looks fake”.
There was a time when spotting an AI-generated video seemed relatively straightforward. Look closely at the hands. Watch whether the person blinks naturally. Check if the mouth matches the words. Pause when the face turns sideways and see whether something suddenly looks wrong.
That checklist is becoming less useful.
Generative AI systems can now produce increasingly convincing faces, voices, movements and environments. A video appearing to show a chief executive announcing an investment opportunity, a public figure making a controversial statement or even a familiar person asking for money may no longer look obviously synthetic.
The risks are no longer theoretical. In July 2026, the Federal Bureau of Investigation warned that criminals were using AI-generated videos to create convincing depictions of public figures, law-enforcement personnel and company executives, including during private communications and real-time interactions.
The FBI’s 2025 Internet Crime Report, released in 2026, recorded 22,364 complaints containing AI-related information, with associated reported losses of $893.3 million. The figures cover AI-enabled cybercrime broadly rather than deepfake video alone, but they point to the wider fraud environment in which synthetic media is being used.
Biometric verification company iProov has also reported a rise in attacks involving synthetic identities. Its 2026 threat report found that injection attacks targeting iOS devices increased 1,151% in the second half of 2025 compared with the corresponding period a year earlier. The figures come from iProov’s own security operations and are not representative of all online fraud, but they show how synthetic media is becoming part of identity-based attacks.
“Identity is becoming the new battleground in cybersecurity,” iProov Chief Scientific Officer Andrew Newell said.
For internet users, marketers and newsrooms, this creates a difficult question. If simply looking at a video is no longer enough, what should people actually check?
There is no single test that can reliably authenticate every video. Instead, verification increasingly depends on combining several pieces of evidence.
1. Check who posted it before checking the person’s face
The first question should not be whether the face looks fake. It should be: where did this video come from?
Imagine receiving a WhatsApp video in which the CEO of a well-known company announces an investment opportunity promising unusually high returns. The face looks convincing. The voice sounds familiar. The company’s logo appears in the background.
None of those things establishes that the video is authentic.
Search for the announcement on the company’s official website and established social-media accounts. Check whether credible news organisations have reported it. If the video supposedly comes from a government official, look for the statement on official government channels.
Account details can offer clues too. Slightly misspelt usernames, newly created profiles and domains that resemble legitimate websites can accompany impersonation attempts.
The FBI advises users to verify surprising videos and images through reputable news sources or known official channels. It also recommends examining account names, email addresses, telephone numbers and URLs for subtle variations.
The principle becomes particularly important when a clip is designed to provoke an immediate emotional response.
UC Berkeley digital-forensics researcher Hany Farid has pointed to confirmation bias as another challenge in assessing manipulated media. “When I send you something that conforms to your worldview, you want to believe it,” he has said while discussing how people respond to synthetic content.
The opposite can happen as well. Genuine material may be dismissed as AI-generated simply because viewers dislike or distrust what it shows.
Source verification therefore protects against two errors: accepting fabricated content as genuine and incorrectly labelling genuine material as fake.
The FBI’s advice is simple: “Be wary of online content that elicits strong emotion.”
Before studying someone’s eyes or hands, find out whether the event itself has an independently verifiable existence.
2. Pause the video and reverse-search important frames
Videos move quickly. Verification often becomes easier when they stop.
Instead of watching a suspicious clip repeatedly, pause it at several distinctive moments and capture frames showing useful details such as a person, building, road sign, vehicle, logo, stage, product or identifiable location.
Those images can then be reverse-searched using tools such as Google Lens. Verification platforms including InVID can also help users and journalists extract frames and examine video material more closely.
This matters because misinformation does not always involve a completely AI-generated video.
A genuine clip from an old event can be reposted as breaking news. Footage filmed in one country can be presented as something happening somewhere else. A real video can be edited, have its audio replaced or contain only a few AI-generated elements.
Reverse-searching a frame may reveal that a supposedly new video has existed online for years.
Consider footage circulating during a breaking event. A post may claim that it shows an explosion, protest or natural disaster occurring that morning. Searching recognisable frames could uncover the same footage in an older article or social-media post from a different location.
The relevant question is therefore broader than “Was this generated by AI?”
Users should also ask: Does this video actually show what the accompanying post claims it shows?
Recent research is beginning to reflect the same shift. A 2026 review published through the Association for Computational Linguistics examined how AI-video detection is moving beyond low-level visual artefacts towards “factual fidelity”, including whether the people, events and physical processes depicted in a video are consistent with real-world information.
For ordinary users, the principle is less technical.
Search the event, not only the face.
A major speech, company announcement, public incident or celebrity appearance will often leave other digital traces. If the only evidence is one viral clip uploaded by an unknown account, that should invite further checking.
3. Look for Content Credentials, labels and watermarks
A third layer of verification may increasingly come from information attached to the content itself.
Technology companies and media organisations are developing provenance systems designed to record how digital content was created or modified.
One approach is Content Credentials, based on standards developed by the Coalition for Content Provenance and Authenticity, or C2PA. Cryptographically signed provenance information can provide details about an asset’s creation and editing history and, where available, indicate whether AI tools were involved.
It is better understood as a digital history than a simple “real” or “fake” badge.
Google has taken a related approach with SynthID, an invisible watermark designed to identify content created or modified using supported Google AI systems. In May 2026, the company said SynthID had been applied to more than 100 billion images and videos and 60,000 years of audio. Google also reported that its AI-content verification capability had been used 50 million times globally.
Those are company-reported figures, but they illustrate the scale at which provenance technology is beginning to be deployed.
Some verification tools can check supported media for SynthID signals or display available Content Credentials, potentially revealing information about editing, composition and AI use.
India is also moving towards greater disclosure of synthetic content. Amendments to the IT Rules announced in February 2026 introduced requirements around the labelling and traceable metadata of permissible synthetically generated information.
However, there is an important limitation: no label does not mean no AI.
Provenance information can be lost when content is downloaded, edited, compressed, screen-recorded or uploaded through services that do not preserve the relevant metadata. Disclosure systems may also depend on platforms, creators and AI providers supporting compatible standards.
Research from UK communications regulator Ofcom illustrates the gap. Among 2,143 UK internet users surveyed, 85% considered it important for platforms to label AI-generated content, but only 34% said they had encountered such a label.
Ofcom has argued that responsibility cannot fall entirely on viewers, stating: “Users should not be left to identify deepfakes on their own.”
A Content Credential or watermark can therefore be valuable evidence when it exists. Its absence should not be treated as proof of authenticity.
4. Watch whether the entire scene behaves consistently
Visual inspection is still useful. What has changed is how much weight it should carry.
Rather than relying exclusively on familiar clues such as strange fingers or unusual blinking, watch how details behave throughout the video.
Pause at different moments.
Does writing on a wall or sign change between frames? Do earrings or spectacles disappear when the person turns? Does the reflection in a mirror match what is happening in the room? Does someone’s hand interact naturally with an object? Do shadows follow the movement of people? Does the person’s mouth remain synchronised with the audio during faster speech?
The FBI continues to advise users to look for irregular faces, distorted extremities, unrealistic accessories, inaccurate shadows, lip or voice mismatches, lag and unnatural movements.
But none of these should deliver the verdict by itself.
AI-generated video is improving partly by eliminating the artefacts people have been taught to spot. The 2026 ACL review of AI-video detection research found that traditional approaches centred on visible generation artefacts are becoming less sufficient as synthetic videos become more realistic.
At the same time, genuine videos can contain apparently suspicious features.
Low lighting, digital stabilisation, beauty filters, video-call glitches, frame interpolation and aggressive social-media compression can distort faces and movements. A genuine hand caught between frames may look unusual. Audio can fall slightly out of sync after a video has been repeatedly edited or re-uploaded.
This is why visual clues work better cumulatively than individually.
One strange-looking frame is weak evidence. Changing background text, inconsistent reflections, unnatural object interactions and a lack of independent evidence that the event occurred together make a stronger reason to investigate.
5. Use AI detectors, but do not let them make the final call
The obvious response to AI-generated content is to use AI to detect it.
A user uploads a suspicious clip, receives a score saying it is “87% likely AI-generated” and assumes the problem has been solved.
Current research suggests it is not that simple.
A 2026 Northwestern University preprint compared 200 human participants with 95 AI detectors across two deepfake video datasets. On the more realistic dataset involving everyday, lower-quality mobile videos, the automated detectors averaged 53.7% accuracy, while human participants averaged 78.4%.
The study does not establish that people are always better at identifying deepfakes. It examined particular detectors, participants and datasets, while detection systems continue to develop.
What it does show is the danger of treating one detector score as definitive.
Detection tools may struggle when they encounter material produced by generators they were not trained to recognise. Videos can also be compressed, cropped, edited or screen-recorded before reaching a detector.
A better approach is to combine machine detection with other verification signals.
A detector might identify an unusual region of a video. A provenance checker might find a watermark. Reverse search could reveal that the background footage came from an older event. Credible reporting could establish whether the supposed incident actually happened.
For high-stakes private requests, verification may need to move outside the video altogether.
If a video call appears to show a senior executive asking an employee to transfer money urgently, the employee can contact that executive through a separately known number or internal communication channel rather than trusting the call itself.
The same principle applies personally. If a video or voice message from a family member suddenly asks for money, contacting that person independently may be more useful than spending ten minutes analysing their facial movements.
Verification is becoming more important than detection
AI-generated video is changing an assumption that has shaped the internet for decades: if there is video evidence, something probably happened.
That assumption can no longer operate without qualification.
It does not mean every strange video is synthetic or that internet users should distrust everything they watch. It means the process of establishing authenticity is becoming broader.
Check who uploaded the video. Reverse-search its important frames. Look for provenance information and labels. Examine whether the scene behaves consistently. Use detection technology, but compare its assessment with independent evidence.
No single method works reliably in every situation.
The more useful question may therefore no longer be simply, “Does this look fake?”
It is increasingly, “What evidence do I have that this is real?”
As AI-generated video becomes better at looking ordinary, that distinction could become one of the most important digital habits for anyone scrolling through a feed.
Disclaimer: All data points and statistics are attributed to published research studies and verified market research. All quotes are either sourced directly or attributed to public statements.