Last updated 24 August 2026 · legal position reviewed for the EU, UK, US and India
Check a photo, or any frame from a video, for the traces an AI image generator leaves behind. It is free, it takes about two seconds, and you do not need an account. Then read the part most detector pages skip: what the score can and cannot tell you, the manual checks that catch what software misses, and how to trace where a file actually came from.
Voice runs on a different engine
Cloned and synthetic speech is not detected by looking at pixels, so audio does not run through the checker above. It has its own anti‑spoofing model, trained on the artefacts a voice cloner leaves in a waveform, and it lives on its own page.
Take it a suspicious voice note, a recording of a call, or a clip pulled from a video, and it returns the same kind of 0–100 score. It accepts MP3, WAV, M4A, OGG and FLAC up to 25 MB. WhatsApp voice notes are Opus and need converting to MP3 or WAV first.
Your file is sent to our API to be scored and is not published, shared or shown to anyone else — the privacy policy sets out what we retain. Videos are never uploaded at all: the frame is cut in your browser. Signed-in free accounts get 5 checks a day, Starter 25 and Pro 100, with Business and Enterprise unlimited — see the plans. Every score is a probability, never a verdict.
Most pages that rank for "deepfake detection" will tell you they catch everything at ninety-something percent. We are going to be specific instead, because the difference decides whether the number above is any use to you.
What runs here. The image checker looks at the pixel statistics of a still and scores how strongly they resemble the output of an AI image generator. Generators leave behind consistent traces — in the frequency domain, in noise distribution, in the way fine texture is reconstructed — that differ from what a camera sensor produces. That is what the 0–100 number measures. The video tab is the same engine applied to one frame you choose: your browser decodes the video locally, draws the frame you scrub to onto a canvas, and only that still is sent to be scored. The video file itself never leaves your device.
What it does not do. It does not do temporal analysis, so it cannot see the things that only exist across frames: a blink rate that is subtly wrong, lip movements a few milliseconds out from the audio, a head that shifts without the neck following. It does not read metadata or Content Credentials. It does not tell you which generator made an image, and it will not name a model, because the engine does not return that and we are not going to invent it. It does not analyse audio — that is a separate engine on the voice detector.
Where that leaves face swaps. This is the honest limitation and it matters. A fully synthetic image or a fully generated video — the kind that comes out of a text prompt — is generated pixel by pixel, so the traces are everywhere in the frame and a still-frame check has a lot to work with. A classic face swap is different: a real camera recorded the room, the clothes, the lighting and the background, and only the face region was replaced. Most of that frame is genuinely photographic. A whole-image score can be diluted by all that real content and come back low. So a low score on a video frame is meaningfully weaker evidence than a low score on a portrait-sized crop of the face.
What to do about it. Crop to the face and check the crop. Zoom to the head, screenshot just that, and run it through the image tab. You are giving the detector a frame where the suspect pixels are most of the picture rather than five percent of it. Then use the manual checks further down, which are specifically good at what the software is bad at.
One person's face mapped onto another person's recorded body and scene. The background, the lighting and the body language are real; the identity is not. This is the original deepfake and still the most common in harassment and fraud.
The real person's own face, puppeteered. Their expressions and head movements are driven by an actor so they appear to say or do something they never did. Nothing is swapped, which makes it harder to spot.
Genuine footage with the mouth region rewritten to match new audio, usually a cloned voice. Everything outside the mouth is authentic, so the tell is timing and the small area around the lips and jaw.
An image or video with no camera involved at all, generated from a text prompt. There is no original to compare against and no real scene. This is the category a pixel-statistics detector handles best.
A few seconds of someone's recorded speech turned into a model that can say anything in their voice. Increasingly the whole attack on its own, over a phone call, with no video at all.
Real media, misrepresented. Old footage relabelled as today's, a clip cut to remove the sentence before, a photo from another country. No AI at all, so no detector will ever flag it — only reverse search will.
That last card is not a throwaway. A large share of what circulates as a "deepfake" is ordinary footage with a false caption, and it is the category people are most often fooled by, because the media survives every forensic test — it is genuine. If your file passes the checker and still feels wrong, the question to ask next is not was this generated but where has this appeared before.
There is a great deal of unsourced statistics circulating about deepfake fraud, much of it published by companies selling deepfake detection. We have left all of it out. What follows are documented cases, each linked to its primary reporting, because the shape of a real incident is more instructive than a percentage anyway.
A finance worker in Hong Kong paid out about US$25.6 million after a video call. Every other participant on that call — including the firm's UK chief financial officer, whose face and voice he recognised — was synthetic. He made fifteen transfers before the deception came to light. What is worth noticing is that the fake did not need to be flawless: it needed to survive a routine call among familiar colleagues, with a plausible reason to hurry. Reported by CNN.
Voice alone was enough, years before the video was convincing. In 2019 the chief executive of a UK energy firm transferred roughly €220,000 after a call from what he believed was his German parent-company boss, complete with the right accent and speech rhythm. Reported by Forbes, from the Wall Street Journal. Voice cloning has become dramatically cheaper and better since.
One still image moved a stock index. In May 2023 an AI-generated photo of an explosion near the Pentagon spread through verified accounts; the S&P 500 dipped briefly before officials confirmed nothing had happened. Reported by NPR. It was debunked largely by open-source researchers noticing that the building frontage was wrong and that no one else had photographed it.
Hiring is now a target. In July 2026, eleven governments issued a joint alert describing operatives using real-time deepfake video during live job interviews to work under stolen identities. If you screen remote candidates, this is no longer hypothetical.
The pattern across all four: the deepfake was not caught by looking harder at the pixels. It was caught — or would have been — by process. A call-back on a known number. A second source for a breaking image. An interview step that a real-time filter cannot survive.
There are four families of technique in use, and each one is good at something the others miss. Knowing which one produced a number is most of understanding what the number is worth.
The approach behind the checker on this page. A model looks at a single frame and asks whether the pixel statistics look like a sensor recorded them or like a network drew them. Generators tend to leave characteristic patterns: over-smooth skin texture, unnatural high-frequency detail, telltale periodic artefacts from upsampling. It is fast, it works on a single still, and it needs no reference copy. Its weakness is that it degrades when the image has been re-encoded, screenshotted, resized or passed through a platform's compression — which is to say, when it has been shared, which is how you got it.
Watching the sequence rather than the frame. Real faces blink at fairly regular intervals, the head and neck move as one connected object, the jaw and the audio agree to within a few tens of milliseconds, and there is a faint colour pulse across real skin as blood flows. Manipulated video often breaks one of these, especially over long clips. This is what catches face swaps and lip syncs that a still-frame check waves through. It is also expensive to run and needs the whole file.
A separate discipline entirely. Anti-spoofing models look for the artefacts a vocoder leaves in a waveform: unnaturally clean breath and silence, spectral detail that stops abruptly at a frequency ceiling, prosody that is a shade too even. Our voice engine works this way, and one of its more useful behaviours is that it refuses to answer when a clip is too short, too quiet or has no speech in it, rather than guessing. Read more about how we talk about audio accuracy.
The one that does not analyse the media at all. Instead of asking "does this look fake", provenance asks "what does this file say about itself, and is that record intact". C2PA Content Credentials attach a signed, tamper-evident history to a file at the point of capture or generation. Several camera makers and most large AI image tools now write them. When the record survives, it is far stronger than any detector score, because it is cryptographic rather than statistical. The catch is that most social platforms strip metadata on upload, so absence of a credential proves nothing at all.
No serious workflow relies on one family. The reason this page pairs a detector with a manual checklist and a provenance step is that the failure modes are different: compression that blinds the pixel model leaves eye reflections untouched, and a caption that lies is invisible to every model but obvious to a reverse image search.
Software is not better than you at all of this. Human vision is unusually good at faces and at physical plausibility, and generators still get a predictable set of things wrong. Open the image at full size, and if there is a video, pause it and step through frame by frame rather than watching it play — motion hides everything. Work down this list.
Two cautions on all of this. First, every tell on this list is a tell that the next model generation fixes, and hands and text have already improved dramatically. Absence of an artefact is not evidence of authenticity. Second, and more importantly: a bad photograph looks like a fake. Low light, aggressive phone beautification, heavy compression and screenshots-of-screenshots all destroy the same fine detail a generator fails to produce. Do not convict a real image for being blurry.
No uploaded-file detector helps you here, and this is where the largest losses happen. Real-time face swaps and voice clones now run well enough on ordinary hardware to survive a routine meeting. The defence is not better looking — it is interaction, because a live model has to render whatever you ask for, in the moment, and it fails in specific ways.
Do not announce that you are testing them. Ask the way you would ask anything else in a call, and watch the edges of the face rather than the middle.
If the call involves money, access, credentials or urgency, verification by appearance is the wrong control entirely, because it can only ever be beaten by a better model. Hang up and call back on a number you already had — not one you were just given, not a number in the meeting invitation. Confirm on a second channel the other party did not choose. Treat urgency plus secrecy plus a changed payment detail as the signal, regardless of whose face is on screen. Both of the multi-million-dollar cases above would have failed against a single call-back.
If you have a recording of the call, the audio is checkable after the fact with the voice detector. Teams that deal with this regularly may find the notes for fraud and finance teams and for recruiters more directly useful than this page.
This is the step that most often actually resolves the question, and no detector can do it for you. You are trying to find the earliest appearance of the file, because the origin usually settles the matter faster than any forensic signal.
Run the image through more than one engine, because they index different corners of the web and disagree constantly. Google Lens is strongest on products, places and mainstream news. TinEye is the one that sorts by oldest first, which is exactly what you want when you are establishing which copy came first. Yandex is unusually good at faces and at finding visually similar rather than identical images. Bing Visual Search catches a slightly different slice again. For a face on a dating or hiring profile, this is usually the whole investigation: the same photo appearing on a stock library, or on a stranger's Instagram from four years ago, ends it.
Two free tools do this properly rather than one engine at a time. Bellingcat's Search by Image extension fires the same picture at Google Lens, Bing, Yandex and TinEye together. The InVID-WeVerify plugin — open-source, EU-funded, and the standard kit for fact-checkers — adds a magnifier for pixel-level inspection, keyframe extraction from video, and metadata reading, then pushes any frame straight into reverse search. Dedicated face-search engines also exist and are effective; we deliberately do not point people at them, because searching the web for a private individual's face carries consequences of its own.
If the whole image finds nothing, crop to a distinctive element — the face alone, a logo, a building, a tattoo — and search that. Reverse search matches on composition, so a stolen photo that has been mirrored, recoloured or had a border added often fails as a whole image but matches instantly on a crop.
Take screenshots at several points, especially just after cuts, and reverse-search those stills. Older footage relabelled as a current event is one of the most common fakes there is, and one screenshot of the right frame is usually enough to find the original upload with its real date.
Drop the file into the official C2PA verification page at contentcredentials.org/verify. If the file carries an intact credential, you get a signed history: what device or tool made it and what was done to it since. This is the strongest single piece of evidence available to a non-specialist. Remember the asymmetry, though — a present, valid credential tells you a great deal, while a missing one tells you almost nothing, because platforms routinely strip this data when you upload.
EXIF can carry the camera make and model, the capture time and sometimes the editing software. When it survives, it is a useful corroborating detail. It is also trivially editable and almost always destroyed by messaging apps, so treat anything you find as a hint rather than a fact.
Who posted it first, and what else does that account post? Does any independent source have the same event from a different angle? Does the weather, the season, the foliage or the light match the date claimed? Verification professionals resolve most cases here rather than in a forensics tool, and so will you.
Run one file through several deepfake detectors and you will frequently get conflicting answers. That is not a scandal, it is the shape of the problem, and understanding why keeps you from over-trusting any single number — including ours.
They were trained on different fakes. A detector learns the fingerprints of the generators it was shown. Point it at output from a model that came out after its training set closed and accuracy drops, sometimes sharply. This generalisation gap is the central unsolved problem in the field and it is why benchmark scores quoted in marketing rarely survive contact with real files from the open internet.
Compression eats the evidence. The traces these models rely on live in fine, high-frequency detail — precisely what JPEG compression, resizing and video re-encoding throw away first. A file that has been posted, screenshotted, forwarded and re-posted has been through that grinder repeatedly. It is entirely normal for the same image to score very differently before and after a trip through a social platform.
Real photographs get flagged. False positives are not rare and they are not evenly distributed. Heavy phone beautification, night-mode computational photography, aggressive denoising and AI upscaling all leave statistical signatures that resemble generated content, because in a literal sense they are partly generated. Professionally retouched portraits do the same. Anyone whose photos routinely go through those pipelines will see more false alarms than average.
A partial fake dilutes the signal. As above: the score is computed over the whole image. Change ten percent of the pixels and the other ninety percent are still authentically photographic. Crop to the region you suspect.
Thresholds are a choice, not a fact. The bands on this page — strong signal, uncertain, no strong signal — are cut-offs somebody selected. Move them and the same score becomes a different verdict. Any tool presenting a bare "REAL" or "FAKE" stamp has made that choice for you and hidden it.
You will see 99% on a lot of deepfake detection pages. Independent measurement keeps landing somewhere else, and it is worth knowing the numbers before you trust anyone's, including ours.
Read those together and the conclusion is not "detection is useless." It is that a single score, from a person or a model, is a weak instrument, and that combining independent signals — a detector, a provenance check, a reverse image search and the manual inspection above — is how you get to a defensible answer. Each catches what the others miss.
This is also why we publish no headline accuracy percentage. A single number, quoted without the dataset it was measured on, tells you nothing about how a tool behaves on your file, and every vendor picks the dataset that flatters them. We would rather describe the failure modes plainly. The same reasoning is set out for text in our accuracy methodology, and for audio in how accurate is voice detection.
The rule that follows from all of this: a detector score is one input, not a decision. It should never be the sole basis for a disciplinary hearing, a firing, a payment, a police report or a public accusation. Use it to decide how hard to look — not to decide what happened.
Reverse image search first, detector second. A romance or investment scammer usually reuses photos that exist elsewhere, and finding the real owner ends it in a minute — faster than any forensic score. If the photos are unique and still feel off, check a crop of the face here, and ask for a live video call with a specific unscripted action.
Live face filters and pre-recorded loops both show up in hiring now. Ask the candidate to turn their head fully to profile and to pass a hand in front of their face — both break most real-time swaps. Screenshot and check the frame here. See voice checks for recruiters for the audio side.
Treat urgency plus secrecy plus a new account number as the signal, not the voice. Hang up and call back on a number you already had. Then check any recording with the voice tools built for finance teams. A cloned voice needs only seconds of source audio, so familiarity is no longer authentication.
Establish provenance before analysis. Reverse-search keyframes, find the earliest upload, check whether any independent outlet has the same event from another angle. Our verification notes for journalists cover the audio workflow in more depth.
Preserve first: screenshot the post with its URL, account name, date and any caption, and save the file itself before it disappears. Do not edit the original. Then report to the platform under its synthetic-media or non-consensual imagery policy, which is usually faster than any other route.
Different problem, same anxiety. For writing rather than images, use the AI text detector — and if you have already been accused, how to prove you didn't use AI is the practical guide.
The detector is the least useful thing on this page in that situation, so here is the order that actually helps.
If you need something in writing about what our tools do and do not measure, get in touch — we would rather give you an accurate description of the limits than have a score of ours misrepresented in a proceeding.
"Is this even illegal?" is one of the most common questions people arrive with, and the answer changed a great deal in 2026. What follows is a short orientation with links to the primary sources. It is general information, not legal advice, and it is not a substitute for talking to someone qualified in your jurisdiction.
European Union. Article 50 of the AI Act requires anyone deploying a system that generates or manipulates image, audio or video content constituting a deepfake to disclose that the content is artificially generated or manipulated. The obligation does not depend on intent to deceive. It applies from 2 August 2026. Article 50 text.
United Kingdom. Creating — not merely sharing — a purported intimate image of someone without their consent became a standalone criminal offence on 6 February 2026, under section 138 of the Data (Use and Access) Act 2025, which inserted new offences into the Sexual Offences Act 2003. The provision. Sharing or threatening to share such images was already a priority offence under the Online Safety Act 2023, which obliges platforms to act against it.
United States. There is now a federal statute covering non-consensual intimate imagery including synthetic material, alongside a fast-growing patchwork of state law — by mid-2026, 33 states had enacted laws addressing deepfakes in political communications. That patchwork is genuinely contested rather than settled: California's election-deepfake law was permanently enjoined in 2025 on First Amendment grounds, and Hawaii's met the same fate in early 2026.
India. In October 2025 the Ministry of Electronics and Information Technology announced amendments to the intermediary rules requiring platforms to label synthetically generated content, with unusually specific requirements: a visual label covering at least a tenth of the display area, or an audio announcement across the first tenth of the duration, neither removable by users or intermediaries. The MeitY notification.
Two practical consequences follow. First, if you publish or advertise with synthetic media, disclosure is becoming a compliance obligation in several large markets rather than a courtesy. Second, if you are the subject of a deepfake, the fastest route is still usually the platform's own reporting flow — the legal route matters, but it runs on a slower clock than the one the content is spreading on.
| You have | Use | What it accepts | Free without an account |
|---|---|---|---|
| A photo | The checker on this page, or the AI image detector | JPG, PNG, WebP, BMP · 10 MB | 3 a day |
| A video | The Video frame tab above | Anything your browser can decode; only the chosen frame is sent | 3 a day |
| A voice note or call recording | Voice detector | MP3, WAV, M4A, OGG, FLAC · 25 MB | 3 a day |
| Written text | AI detector | Pasted text | 3 a day |
| A PDF or Word file | Document detector | PDF, DOCX | Shares the text allowance |
Signed-in free accounts get more room on each of these (5 a day), and paid plans raise the daily ceiling further — Starter 25, Pro 100, Business and Enterprise unlimited — the ladder is on the pricing page. The image and voice engines are separate products with separate allowances, so using one does not spend the other.
Every number this page returns is a statistical probability produced by a model, not a finding of fact about a file. It can be wrong in both directions: it can flag a genuine photograph, and it can pass a manipulated one. Nothing here is forensic authentication, expert analysis, or a certification of authenticity.
Do not rely on a score from this page as the basis for:
If a matter may reach a court, a tribunal, a regulator or a formal internal process, instruct a qualified forensic examiner. A properly conducted examination documents its method, its chain of custody and its error rates, and can be tested by the other side. Output from an automated web tool does none of that and should not be presented as though it does.
The description of legislation on this page is general information, current at the date shown at the top of the page, and is not legal advice. It is not a complete statement of any law, it creates no lawyer–client relationship, and this area is changing quickly and differs by jurisdiction. Take advice from a qualified lawyer where you are before acting on any of it.
We claim only what the tool does: our image engine scores a single still for statistical signs of AI generation. It does not perform temporal video analysis, read metadata or Content Credentials, and does not identify which generator produced a file — and we do not display capabilities it does not have. We publish no accuracy percentage, because a figure quoted without the dataset it was measured on would not be a meaningful claim. We give no warranty, express or implied, that any result is accurate, complete or fit for a particular purpose.
Use of this page is governed by our Terms of Service, and what happens to a file you submit is set out in our Privacy Policy. Please upload only material you have the right to upload.
TextSight builds AI-detection tools for text, images and voice. Every score we return is a probability produced by a statistical model, not a determination of fact, and it can be wrong in both directions. Our image engine scores a single still for signs of AI generation; it does not perform temporal video analysis, read metadata, or identify which generator produced a file, and we do not display capabilities the engine does not return. No detector output, ours included, is forensic evidence, and none of it should be the sole basis for a disciplinary, employment, financial or legal decision.
Three checks a day without an account, more on a free one. Same score, same explanation, on every plan — and we tell you where the number is weak.