How Deepfakes Are Detected, and Why Detection Is Losing
Deepfake detection explained: the visual, audio and metadata artefacts investigators look for, the tools platforms use, and why generation keeps outrunning it.

In short
- Deepfake detection looks for artefacts in three layers: what the eye can see, what the audio reveals, and what the file's metadata and provenance say.
- Visual signs include lighting mismatches, flickering glasses or earrings, blurred teeth and a jawline that lags the voice; they shrink with every model generation.
- Compression on WhatsApp and Instagram destroys many artefacts, so a forwarded clip is harder to analyse than the original. Get the original file.
- Detectors trained on one generator miss the next one. No tool gives a reliable yes or no; a forensic report gives a reasoned opinion.
- For money decisions, replace detection with procedure: call back on a known number, use a safe word, require a second approver.
Who this is for: Victims of deepfake impersonation, journalists, HR and compliance teams, investigators and anyone who has been sent a video and asked "is this real?"
A video of a well-known fund manager recommending a trading app. A clip of a politician saying something he never said. A WhatsApp video call from a "police officer" in uniform. The question in each case is the same, and it is the wrong one: is this real? Deepfake detection can sometimes answer it. More often it can only say "probably", and the probability gets weaker every year.
This explainer sets out how deepfake detection actually works, in the three layers investigators use, what a forensic opinion can and cannot establish, and why the sensible response to a convincing video is a procedure rather than a verdict.
What a deepfake is, mechanically
Most video deepfakes are one of two things. A face swap takes a source face and maps it onto a target performer's head, frame by frame, blending at the edges. A lip-sync or reenactment model keeps the target's face and alters the mouth and expressions to match new audio. Audio deepfakes synthesise speech in a cloned voice from text. Fully generated video, where no performer exists, is improving quickly and blurs these categories.
Each method leaves traces, because each has to invent something: the edge where the swapped face meets the neck, the inside of a mouth that was never filmed, the breath between sentences nobody took. Detection is the search for those inventions.
Deepfake detection, layer by layer
Layer 1: visual artefacts
Trained eyes and trained models look for the same things. Lighting on the face that does not match the room, or a shadow on the nose that disagrees with the shadow on the wall. Reflections in the eyes that differ between left and right. Glasses frames, earrings and hairlines that flicker or shift between frames, because they sit on the blend boundary. Teeth rendered as a blurred white band. Skin that is too smooth across the cheek and too sharp at the jaw. A head that turns while the face stays frontal for a fraction of a second. And lip sync that drifts, most visibly on the consonants p, b and m, where a real mouth closes fully.
Models trained on thousands of fakes pick up patterns no human sees, such as statistical fingerprints left by the generator's upscaling. Those fingerprints are also the first thing the next generator removes.
Layer 2: audio
Cloned voices are often very good on vowels and poor on the texture around them. Analysts listen for unnaturally even pacing, missing breaths, flat emotion on words that should carry stress, and room acoustics that do not change when the speaker supposedly moves. Spectral analysis can show frequency bands that synthesised speech handles differently from a microphone recording. On a compressed phone call, most of this is lost, which is exactly why scammers prefer the phone.
Layer 3: metadata and content provenance
Every file carries metadata: device, software, timestamps, encoding history. A genuine phone video and a clip exported from an editing tool look different inside. But metadata is easily stripped or rewritten, and every platform re-encodes uploads, so its absence proves little. Content provenance is the attempt to fix that from the other side: standards such as C2PA attach a signed record at capture and at each edit, so a viewer can verify origin. It can show that a genuine clip is genuine; it cannot flag a fake produced with tools that never signed anything.
Why detection is losing
Four reasons, and none of them is going away.
The generalisation problem. A detector learns the fingerprints of the generators it was trained on. A new generator, or an old one with a different setting, produces fakes it has never seen, and accuracy drops sharply. Attackers always have the newer tool.
Compression. Indian scam content lives on WhatsApp, Telegram, Instagram and phone calls. Each strips the fine detail detection depends on. A detector that is reliable on a lab-quality file is far weaker on the forwarded version.
Base rates. A tool that is right 95% of the time sounds good until it is applied to millions of genuine videos, where 5% false alarms swamp the real fakes. Platforms therefore tune detectors to be cautious, and cautious detectors miss things.
The liar's dividend. As fakes become plausible, genuine recordings become deniable. A politician caught on a real recording can call it a deepfake, and a scammer can call a victim's real evidence a fabrication. Detection cannot settle that argument either way with certainty, which is why provenance and corroboration matter more than classifiers.
What a forensic analysis can actually say
A competent forensic analysis of a video does not output "fake" or "real". It documents the file's history, examines frames and audio against the signs above, compares the subject with known genuine footage, and gives a reasoned opinion with a confidence level and the limits of the method. That opinion, with a section 63 BSA certificate, is what a court, a platform's grievance officer or an employer's inquiry can act on. A digital evidence and forensic report prepared this way is also what the police forensic lab will recognise. People who want to learn the method can start with digital forensics basics.
What platforms and regulators do
Large platforms run classifiers on uploads and act on reports under their impersonation and sexual-content rules. MeitY's advisories of December 2023 and March 2024 reminded intermediaries of their duties under the IT Rules 2021 to remove reported deepfakes within set timelines, and proposed rules would require synthetic content to be labelled; see the MeitY page and the takedown process for how to use them. Banks and KYC providers add a liveness check to video verification. Europol's IOCTA reports have tracked the shift of deepfakes from novelty to routine criminal tooling, and the wider picture is in our hub on AI and cyber crime.
What to do instead of trusting detection
For any decision that involves money, access or reputation, treat the video as unverified and go to a channel the fake cannot control.
- A relative in distress on a call: hang up and call them back on their saved number. Use the family safe word if you have one.
- A "director" on video approving a payment: confirm by phone on the number in your directory, and require the second approver anyway.
- A celebrity endorsing an investment: assume it is fake. Check the adviser on SEBI's registry; the mechanics are in deepfake investment ad scams.
- An "officer" on video: no Indian agency conducts arrests or interrogations by video call. Hang up and call 1930 if any money has moved.
- A clip of yourself or your company: preserve the original, report on the platform and on cybercrime.gov.in, and get a forensic opinion before any public statement.
Next step: if a specific video is the reason you are reading this, stop forwarding it, download the earliest copy you can find, record where and when you got it, and keep it unaltered. That single act preserves more evidential value than any detector score. Then decide whether you need a platform report, a police complaint or a forensic opinion, in that order.
Frequently asked questions
Is there a free tool that tells me if a video is a deepfake?
There are free and commercial classifiers, and they are useful as one input. None is reliable enough to decide a money transfer, a disciplinary action or a court filing on its own. They give a probability from a model trained on past fakes, and they are weakest on new generators and on heavily compressed clips, which describes most of what circulates in India.
What is a liveness check?
A test used in video KYC and some banking apps to confirm a live person is present: turning the head, blinking on request, reading random words. It defeats a pre-recorded or simple face-swap feed but not every real-time system. It is a useful hurdle, not a guarantee, and regulators treat it that way.
Can a deepfake be proven in court?
A forensic analyst can give a reasoned opinion on authenticity, supported by frame analysis, audio analysis, metadata and comparison with the source. Electronic evidence needs a certificate under section 63 of the Bharatiya Sakshya Adhiniyam 2023. Courts weigh expert opinion alongside the other evidence; a detector score alone is not proof.
Someone made a deepfake of me. What do I do first?
Save the original file and the URL, note the date and account, and do not engage with the poster. Report it on the platform under impersonation or sexual content rules, file on cybercrime.gov.in, and for women the National Commission for Women can help. Content provenance and takedown steps are covered in our takedown guide.
What is content provenance?
A way of attaching a signed record to a photo or video at the moment it is captured or edited, so that a viewer can check where it came from and what was changed. The C2PA standard is the main effort. It helps prove that genuine content is genuine; it cannot flag a fake made with tools that do not sign.
If you want help with this
- Digital evidence and forensic report · Screenshots, chats, emails and recordings only help if they will be accepted. We preserve digital evidence properly, analyse what it shows, and certify it under section 63 of the Bharatiya Sakshya Adhiniyam.
- Digital forensics basics: phones, email, chain of custody, Section 63 BSA · training · A two-day hands-on course for investigators, lawyers and IT teams on acquiring and examining evidence from phones and email, maintaining chain of custody, and producing a report and a Section 63 BSA certificate that stand up in court.
Keep reading
AI and Cyber Crime: What Changed in 2025–26
The scams are the same. The cost of running them has collapsed. Here is what generative AI changed in the way Indians are defrauded, what it...
Deepfake Scams: Celebrity, CEO and Fake Investment Ad Videos
The industrialist on the video never said it. The regulator on the stock-tips clip never recorded it. How deepfake investment ads and cloned...
Voice Clone Scams and the Family Safe Word
A cloned voice can say anything. It cannot say the word your family agreed at dinner last Sunday. Here is how to choose one, who should know...
How to Preserve Digital Evidence for a Cyber Crime Case
Victims delete chats out of shame, "clean" phones before handing them over, and crop screenshots. Each one weakens a case. Here is how to ke...