How to Spot AI Images and Deepfakes
Updated: Oct 31
In short:
To spot an AI image or a deepfake, first find where it came from: a reverse image search often leads to the original or a fact-check. Then check for content credentials, and only then look for errors in hands, text and shadows, or, in video, lips, blinking and voice.
Check an image now
Find where it came from. Upload the image to Google Lens, Bing Visual Search or TinEye to see where it appeared first and whether fact-checkers have already written about it.
Look for content credentials. Upload it to Content Credentials Verify. Most images don't carry credentials yet, so finding none proves nothing.
Then check the details in the checklist below: hands, text and shadows, or, in video, lips, blinking and voice.
Do AI detectors help? Only as one clue. In our test of four free detectors, "two caught all 10 AI images, one caught only 4, and two called a real photo AI-made" (Do AI Image Detectors Work? We Tested Them).
Also known as: AI image detection; deepfake detection; synthetic media; AI-generated content.
Key facts
In a 2025 Microsoft study with more than 12,500 participants, people identified AI-generated images correctly only 62% of the time, slightly above chance.
On real-world deepfakes collected in 2024, open-source detectors performed about half as well as on older test sets, and no commercial detector tested reached 90% accuracy.
The FBI warned in May 2025 that AI-generated content "has advanced to the point that it is often difficult to identify".
From 2 August 2026, the EU AI Act requires anyone who publishes a deepfake to disclose that the content has been artificially generated or manipulated.
Google lets users check for its SynthID watermark in the Gemini app (since November 2025) and in Search (since May 2026), but this detects only content made with Google's own AI tools.
Debunk.org analysis of a deepfake scam campaign, January–February 2025: 278 Facebook and Instagram ads using AI-generated news presenters were viewed 1,088,829 times.
What is a deepfake? A deepfake is an image, audio, or video generated or manipulated with AI so that it "would falsely appear to a person to be authentic or truthful", in the words of the EU AI Act. An AI image is any picture made by an image generator, whether it imitates a real person or not.
Quick checklist: how to spot a deepfake or AI image
Check the source first and the pixels last: where an image comes from tells you more than how it looks.
Reverse-search it (Google Lens, Bing Visual Search): is there an original, an older copy or a fact-check?
Who posted it first? A real outlet, a named person, or a new account with no history?
Look for a label or credentials: "AI-generated" labels, Content Credentials, a SynthID check.
Check hands, text and objects: extra or merged fingers, garbled letters, jewellery that melts into skin.
Check light and physics: shadows in different directions, impossible reflections, floating objects.
For video: lips out of sync, odd blinking, a voice that doesn't match, a studio with the wrong logo.
Ask what it wants from you: urgency, money, or outrage are red flags, no matter what the image looks like.
Found a fake? Report it on the platform where you saw it; Facebook explains how in its guide to reporting content.
Find the source
The most reliable way to check an image is to find its original, because AI fakes and old photos reused out of context both fall apart once you see where they came from.
Full Fact, the UK fact-checker, notes that "fact checking usually starts with an attempt to find the original". Upload the image or paste its address into Google Lens or Bing Visual Search. Look for the earliest copy, a credible publisher, or a fact-check. No results at all for a dramatic "news" photo is itself a warning sign.
Check for content credentials and AI watermarks
Some images carry a digital record of how they were made; checking for it takes seconds, but its absence proves nothing.
Content Credentials (C2PA): a standard for attaching information about how an image was made and edited, backed by companies including Adobe, Google, Meta, Microsoft, OpenAI, Sony and the BBC. Google shows this information in "About this image" where it exists.
SynthID: Google's invisible watermark for content made with its AI. You can ask the Gemini app whether an image was "generated with or edited by Google AI", and since May 2026, the check is available in Google Search. But "Gemini can currently only recognise content created by Google AI tools".
Platform labels: social networks label some AI content, but labels are easily lost when an image is cropped, screenshotted or re-uploaded.
How to identify AI images: look for visual errors
AI images often contain errors in anatomy, text, objects and physics, but newer models make fewer of them.
A 2024 guide by researchers at Northwestern University groups the tell-tale signs into five types: "anatomical, stylistic, functional, violations of physics, and sociocultural". Look for:
Hands and bodies: "missing fingers, extra fingers, merged fingers, nonexistent fingernails"; teeth and ears that don't match.
Text: garbled letters on signs and labels. But text "is an area that is rapidly improving", and real text can be added to an AI image afterwards.
Accessories: the FBI lists "unrealistic accessories such as glasses or jewelry" and "inaccurate shadows".
Physics: "shadows cast in inconsistent directions", reflections that don't match, objects that float or merge.
Faces: eyes with "misaligned pupils, pupils that aren't circular, unnaturally glossy eyes"; "an unlikely alignment of teeth, numbers of teeth, or an overlapping of teeth and lips"; ears, nose and mouth whose proportions don't match; hair that looks "excessively soft, fine-grained and windswept".
Style: skin that is too smooth, a glossy "perfect" look, repeated patterns in crowds or backgrounds.
These signs help, but don't rely on them alone: the FBI warned in May 2025 that AI-generated content is "often difficult to identify".
How to identify deepfakes in video and audio
Video deepfakes give themselves away through mismatched lips, blinking and voice, and through context such as the wrong studio or logo.
The MIT Media Lab's Detect Fakes project suggests asking: "Do the lip movements look natural?" and "Does the person blink enough or too much?" The UK's Advertising Standards Authority warns of "unnatural facial movements, jerkiness or unusual-sounding speech". For voice calls, the FBI points to "voice call lag time, voice matching, and unnatural movements", and recommends families agree on a secret word to check a caller's identity.
On a live video call, Hong Kong's Investor and Financial Education Council suggests asking the caller "to do specific gestures in front of the camera and watch carefully for any anomalies".
Also check the context: is the person really on that channel, does the studio match the logo, and would they really promote a product or an investment?
Examples of fake AI images that went viral
Some of the best-known AI fakes fooled millions of people before fact-checkers caught them.
Fake | What happened | Fact-check |
Pope Francis in a white puffer jacket (March 2023) | Went viral on 25 March 2023; first posted on Reddit's Midjourney forum; "a deepfake image created via artificial intelligence" | |
An "explosion" near the Pentagon (May 2023) | A fake image spread on social media; the local fire department said "There is NO explosion or incident"; the S&P 500 "briefly dropped a modest 0.3%" | |
Donald Trump being "arrested" (March 2023) | Made with Midjourney by a researcher "while waiting for Trump's arrest", then shared as if real |
The lesson: each fake was obvious to someone checking the source, but not to people scrolling past it.
How to detect deepfakes: do detectors work?
Detectors can help, but they are often wrong on real-world fakes, so treat their score as one clue, not a verdict.
On deepfakes collected from social media in 2024, open-source detectors' performance (AUC) fell by about 45–50% compared with older benchmarks, and many scored "close to 0.5, the same as random guessing".
The best commercial tools tested reached 0.82 accuracy for images, 0.78 for video and 0.89 for audio; "No commercial models that we evaluated had an accuracy of 90% or above".
In a 2023 Bellingcat test, a popular detector "falsely identified seven of ten images as real" once they were compressed, and flagged six of 20 real prize-winning photos as AI.
People do little better: in a 2025 Microsoft study of more than 12,500 participants, the "overall success rate" was "only 62%… slightly above chance".
Our FIMI 101 course, a 10-hour, self-paced online course, covers the skills and tools to recognise, analyse and counter AI-driven disinformation.
Real deepfake scam ads we caught
Debunk.org tracks scam ads on social media as an EU trusted flagger; deepfakes of news presenters and politicians are a recurring tool.
Case | What the deepfake showed | What gave it away |
Fake "Interpol compensation" ads, January–February 2025 | A deepfaked news reporter in a real TV studio, urging "scam victims" to claim money; other videos used deepfakes of public figures | "The branding on the screen has always been mismatched"; the AI segment was "typically under a minute", followed by "a static banner"; the ad text was rated "98.7% likely to be AI-generated" by a detection tool. 278 ads, 1,088,829 views |
A minister "advertising" foot-fungus cream, November 2024 | A real interview with a Montenegrin minister, "with an altered voiceover" | The product name changed within the ad; he "speaks" a language that didn't match him; the ad came from a new Facebook page with no content, run from abroad |
The Interpol ads were later removed for breaking Meta's rules. More: how to find out who owns a website shows how we traced the Interpol scam's domain.
What the law says (EU)
Under the EU AI Act, deepfakes must be disclosed from 2 August 2026, and AI tools must mark their output in a machine-readable way.
People who publish deepfakes "shall disclose that the content has been artificially generated or manipulated" (Article 50(4)); for evidently artistic, satirical or fictional work, a lighter disclosure applies.
Providers of AI tools must ensure output is "marked in a machine-readable format and detectable as artificially generated or manipulated" (Article 50(2)). Tools placed on the market before 2 August 2026 have until 2 December 2026 to comply.
Timing: the disclosure must come "at the latest at the time of the first interaction or exposure".
The European Commission published a voluntary Code of Practice on marking and labelling AI-generated content on 10 June 2026.
Scammers don't follow these rules, so the checks above still matter. This is general information, not legal advice.
Deepfakes can also be used in foreign information manipulation and interference (FIMI): what FIMI is.
FAQ
How can you tell if an image is AI-generated? Find the original with a reverse image search, check for content credentials or a SynthID watermark, then look for errors in hands, text, shadows and backgrounds. Use several checks together.
Is there an app to identify AI images? Detection apps exist, but tests show they often misidentify real-world images. Google's SynthID check works only on content made with Google's AI.
How do you spot a deepfake video? Watch the lips, blinking and voice, and check the context: the channel, the studio, the logo and whether the person would really say this.
What can I do if someone made a deepfake of me? Keep evidence (screenshots with links and dates), report the content to the platform, and contact the police. For fake intimate images, StopNCII.org lets adults create a "hash" (a digital fingerprint) on their own device so participating platforms can block the image without uploading it, and Google accepts removal requests when "the content is fake and falsely depicts you nude or in a sexually explicit situation". An EU directive makes creating and sharing such deepfakes without consent a crime, and countries must adopt it by 14 June 2027.
Can I ask an AI chatbot if an image is AI-generated? You can, but don't rely on the answer. General chatbots are not built as detectors, and even Google's Gemini, which can check for Google's SynthID watermark, "can currently only recognise content created by Google AI tools". Use the checks above instead.
Are deepfakes illegal? Is it illegal to watch them? Watching a deepfake is generally not a crime in itself, but sharing one can be: for example, for fraud, or sexual deepfakes of real people made or shared without consent, which an EU directive requires countries to criminalise by June 2027. Any sexual image of a child is different: EU law covers "realistic images" too, and knowingly accessing such material online is a crime across the EU. This is general information, not legal advice.
What should I do if I see a deepfake scam ad? Don't click or pay; report the ad to the platform. If you lost money, contact your bank and the police.
Get our latest analysis by email
Subscribers are notified when we publish new articles, with a short summary, and hear about upcoming Debunk.org events.
