To spot an AI-generated image, start with where it came from, not how it looks. Run a reverse image search, find the earliest copy, check for Content Credentials or a watermark, and only then look for visual signs, which fade as the tools improve. For video, watch what happens over time; for voices, do not trust your ear: hang up and call back on a number you already have.
The old giveaways, like extra fingers and garbled signs, are largely solved. Check physics, check time, and above all check where a piece of media came from.
- In images, look for consistency. Shadows and reflections that disagree with the light, backgrounds that stop making sense, focus that is too even.
- In video, watch what happens over time. Objects that drift between shots, hair and cloth that move without weight, a face that stays suspiciously the same.
- With voices, verify instead of listening. A clone may be impossible to hear, so rely on a family code word and calling back on a saved number.
- Provenance outlasts the visual tells. Look for the earliest copy, a second angle, and the source's own verified channel.
- Detection tools give signals, not proof. They struggle with compressed or re-shared media, and a score should not decide what happens to a person.
Count the fingers. Check the teeth. Look for garbled text on signs. Look for the weird ear. These were real failures of 2023-era image models and they are largely solved. Advice built on them is worse than useless now, because it hands you a confident all-clear on a fake that simply does not have that flaw. If a guide's main advice is finger-counting, it was written for a world that no longer exists.
Step-by-step check for any image
The plain point: check the picture’s history before you study its pixels. Four steps, in this order, catch more fakes than staring does.
- Run a reverse image search. Upload the picture, or paste its link, into an image search such as Google’s “Search by image”, and see where else it appears.1
- Find the earliest copy. Among the results, look for the first time the image was posted and who posted it. A real event usually has more than one witness: a second angle, or the same moment from another person. A picture that appears from nowhere, with no source and no second copy, deserves the most doubt.
- Check for Content Credentials or a watermark. Content Credentials are a signed record of how a file was made, which C2PA, the group behind the standard, likens to “a nutrition label for digital content”; a small Content Credentials pin on an image shows one is attached.23 For images made with Google’s AI, you can upload the picture to the Gemini app and ask “Is this AI-generated?”; it looks for SynthID, Google’s invisible watermark, which Google says is designed to survive cropping, filters and compression.45 A missing label proves nothing: credentials can be separated from a file, and metadata is often stripped when pictures are shared.67
- Only then, look for visual signs. Use the signs below to decide whether to dig further, not to reach a verdict.
Treat a suspicious picture like a rumour: before you judge it, ask who said it first and whether anyone else saw it.
Search for the picture before you study it: earliest copy first, labels second, pixels last.
Images: look at physics, not anatomy
Generative image models learned what scenes look like without learning the rules that produce them. So the durable signals are about consistency rather than craft.
Light that disagrees with itself
Find the light source, then check every shadow in the frame points away from it, and that reflections in glass, water, eyes and metal show what should actually be there. Multiple objects lit from subtly different directions is a strong signal — it is a rule the model was never given.
Background that stops making sense
The subject gets the model's attention; the background gets plausibility. Look for architecture that cannot be built, repeated patterns in crowds, objects merging into each other, and text that is confidently shaped but meaningless at a glance.
Uniform focus and flawless surfaces
Real lenses produce depth: some things sharp, others not, in a way governed by distance. Generated images often render everything with similar clarity, and skin, fabric and stone can be too clean — no dust, no wear, no pores where pores should be.
Scale and count errors
Hands slightly too large for the arm, a chair the wrong size for the table, five posts in one shot and six in another. Models learn appearance, not measurement.
Stop counting fingers. Check whether the light, shadows, reflections and sizes in the frame agree with each other.
Video: look at time, not frames
The same scene rules apply, and continuity adds new ones — because a video model has to keep a world consistent across time, which is much harder than making one convincing frame.
Objects that drift between shots
A logo changes slightly, a necklace moves to the other side, a room gains a window. Cross-shot consistency comes from reference material, and reference material has gaps.
Motion without weight
Hair, cloth, water and dust betray generated video, because they should carry inertia — continuing after the body stops, settling, catching. Synthetic motion often moves with the subject and stops when it stops.
A face that is too consistent
Counter-intuitive, and one of the strongest signals. Production pipelines use character reference sheets to hold a face steady across shots, which produces identical bone structure, hair behaviour and wear from every angle. Real filming drifts: hair moves, make-up degrades, fabric creases differently. Uncanny sameness is itself the tell. See how AI films are actually made for why.
Short clips and convenient cuts
Coherence degrades with length, so generated footage tends to arrive in short pieces with cuts falling at suspiciously regular intervals. A long unbroken take of a person doing something complicated is still expensive to fake.
Sound that does not match the room
Audio is frequently replaced after the fact. If the voice keeps identical texture and cadence while the visible space changes — indoors to outdoors, small room to large — the audio and video probably did not come from the same event.
Judge a clip over time: whether objects stay put between shots, whether loose things carry weight, and whether a face is too unchanged from every angle.
Voice: stop listening, start verifying
This is the section where honest advice contradicts instinct. You will probably not hear it. Voice cloning requires only a few seconds of sample audio — Microsoft researchers demonstrated reproduction from roughly three seconds in 2023 — and a short phone call gives your ear very little to work with.8
Which is why the defence has to be procedural, not perceptual. The habits that work regardless of how convincing the voice sounds: a family code word agreed in advance, and a callback rule — you hang up and dial the number already saved in your phone, never one the caller supplies. Treat urgency and secrecy as the finding rather than the context. The full set, with the documented cases, is on the deepfake defence page.
Where you do have a recording to examine, the useful signals are environmental: breathing that does not fit the sentence rhythm, room tone that stays constant while the speaker supposedly moves, and emotional delivery that is appropriate but never quite spontaneous — grief with correct words and even pacing.
Every signal above will weaken. The one that does not is provenance: where did this come from, and does it exist anywhere else? A real event has more than one witness. Search for the earliest copy, look for a second angle, check whether the person or outlet posted it on their own verified channel. Media with no traceable origin and no corroboration is the actual red flag — regardless of how it looks or sounds.
Do not stake a decision on your ear or your eye. For a call, hang up and ring a saved number; for media, find where it first appeared and whether anyone else captured it.
Do AI image detectors work?
The plain point: not well enough to settle anything. A detector gives a guess, and a guess should never decide what happens to a person. Do AI detectors work? sets out the evidence for text, images and video.
Treat software claiming to identify deepfakes with extra caution on compressed, re-encoded, screenshotted or re-filmed media — the form many shared clips take — because each of those steps throws away detail. That is reasoning: this page cites no measurement of how much accuracy such tools lose. The same limitation applies here as with text detectors: a percentage is not a verdict, and nobody should face consequences on one. Watermarking and provenance standards embedded at the point of creation are a more promising direction than after-the-fact detection, because they carry a claim rather than guessing at one.
Treat a detector's score as one signal to weigh, not as grounds for acting against a person.
Synthetic images and video are not inherently deceptive — the illustrations on this site are generated, and labelled. The harm is unlabelled synthetic media used to make someone appear to do or say something they did not. As the tells disappear, the burden shifts from your eyes to the record: verified channels, corroboration, provenance. Learn the signals, but do not stake anything important on them alone.
QUICK CHECK · 3 QUESTIONS
Pick an answer to see if you got it. Nothing is recorded.
1. According to the page, what should you check first when an image looks suspicious?
- Count the fingers and check the teeth
- Where it came from, using a reverse image search and the earliest copy
- Whether the colours look too bright
Show the answer
Answer: Where it came from, using a reverse image search and the earliest copy. The page says to check a picture’s history before its pixels: reverse search, earliest copy, labels, and only then visual signs.
2. What does the page suggest if a call in a loved one’s voice asks for urgent help?
- Hang up and call back on a number already saved in your phone
- Listen closely for robotic sounds in the voice
- Ask the caller for a number to ring back on
Show the answer
Answer: Hang up and call back on a number already saved in your phone. The page says a cloned voice may be impossible to hear, so rely on a family code word and a callback to a number you already have.
3. What does the page say an AI detector’s score is worth?
- Enough to settle whether media is fake
- Nothing, so it should be ignored
- One signal to weigh, not grounds for acting against a person
Show the answer
Answer: One signal to weigh, not grounds for acting against a person. The page says detectors give signals, not proof, struggle with compressed or re-shared media, and a score should not decide what happens to a person.
Related on this site: Old Photo Restoration Prompts · Is This Image or Video Real? · Deepfake vs AI-Generated Image
- 1Google Search Help, Search with an image on Google, read at source 5 Oct 2026“Click Search by image and then Upload a file”; or “In ‘Paste image link,’ paste the URL”. ↑
- 2C2PA, c2pa.org (home page), read at source 1 Oct 2026“Content Credentials function like a nutrition label for digital content”. ↑
- 3Content Credentials, contentcredentials.org (home page), read at source 1 Oct 2026“The Content Credentials pin signals that the content contains information about its provenance.” ↑
- 4Google, How we’re bringing AI image verification to the Gemini app, 20 Nov 2025, read at source 5 Oct 2026upload it “to the Gemini app and ask a question such as: ‘Was this created with Google AI?’ or ‘Is this AI-generated?’” ↑
- 5Google DeepMind, SynthID, read at source 5 Oct 2026“SynthID embeds digital watermarks directly into AI-generated images, audio, text or video”; it is “designed to stand up to modifications like cropping, adding filters, changing frame rates, or lossy compression.” ↑
- 6C2PA, FAQs, read at source 1 Oct 2026“While C2PA Manifests are typically embedded in the asset, they can be separated.” ↑
- 7NIST, NIST AI 100-4, November 2024, read at source 1 Oct 2026, comparing watermarks with metadata“unlike metadata, which is often stripped when content is disseminated, covert watermarks are designed to be persistent”. ↑
- 8Source: Microsoft Research, VALL-E — “Neural Codec Language Models are Zero-Shot Text to Speech Synthesizers”, January 2023, read at source 10 Sep 2026“only a 3-second enrolled recording of an unseen speaker”. Microsoft's own description: high-quality personalised speech from “only a 3-second enrolled recording of an unseen speaker”. Trained on 60,000 hours of speech from over 7,000 speakers; code not released. A 2024 successor claimed human parity. ↑
Free printable AI cheat sheets, and AI news every week.
Sixteen one-page AI guides and checklists in one free PDF, no sign-up. And if you want to keep up, the newsletter covers what changed in AI each week, in plain English, with the sources.
SUBSCRIBE FREE ↗ GET THE CHEAT SHEET PACK