Ilia Zavialov on the difference between a deepfake, a voice clone and a synthetic identity
Four words travel together in every conversation about synthetic media, and they are usually treated as interchangeable. A deepfake, a voice clone, a synthetic identity and a generative forgery describe different objects, built in different ways, and each one fails under a different check. When the words blur, the response blurs with them, and a bank, a newsroom or a family ends up applying the wrong test to the wrong thing. This article separates the four terms, marks the border between them, and explains what changes for anyone who has to act once that border is visible.
Ilia Zavialov on the difference between a deepfake, a voice clone and a synthetic identity
Four words, four different objects
The four terms entered ordinary use through headlines, and headlines compress. A video in which a politician appears to say something invented gets called a deepfake. A phone call in a familiar voice gets called a deepfake. A bank account opened in the name of a person who has never existed gets called a deepfake as well. Compression is convenient for a headline and expensive for anyone who has to decide what actually happened.
Each of the four words names a distinct production process. One copies a face and a body in motion, one copies the sound of a specific throat, one assembles a person out of fragments that never belonged to a single human being, and one manufactures an object such as an invoice or a certificate. The tools overlap heavily, because the same families of generative models sit under several of these outputs. The results still behave differently the moment somebody tries to verify them.
Verification is where the distinction earns its keep. A copied face can be checked against the person it was copied from, because an original exists somewhere in the world. An assembled person cannot be checked that way, because there is no original to hold the copy against. The check that catches one of these passes the other cleanly, and that is the practical reason to keep the vocabulary separate.
Deepfake: a copy of a person who exists
A deepfake begins with a real human being and produces media in which that human being appears to do or say something they never did. The original is the defining requirement. Training material comes from a specific person: recorded speech, facial movement, posture, habitual gestures. Without a real subject behind the output there is only invented footage of a stranger, which belongs to a different category entirely.
The label covers video face swaps, full-body puppeteering and lip synchronisation laid over existing footage. It also gets stretched across ordinary editing, where genuine material is cut so that the meaning inverts. That stretched usage weakens the word, because a misleading edit calls for a completely different check: locating the full recording and the minutes around the clip. Reserving the deepfake label for generated media keeps the right check attached to the right word.
The defining vulnerability of a deepfake is comparison. A genuine original exists, which means an investigator can look for the source recording, line up timelines, examine the setting, and ask the depicted person directly whether the event took place. This route is slow and frequently inconclusive, and it still exists, which is more than can be said for the other three categories. Everything about the response to a deepfake follows from the presence of that reference point.
Voice clone: the narrowest term and the hardest channel
A voice clone copies one channel only: the sound of a specific person speaking. By production method it sits inside the deepfake family, and by consequence it deserves its own name, because a voice arrives where the recipient has no picture, no document and almost no time. Current systems build a convincing clone from a sample of roughly three seconds and reach around eighty-five per cent accuracy against the source voice. The sample is trivial to obtain, since voicemail greetings, short video posts and recorded meetings all supply usable audio.
Human detection is the part that surprises people. In controlled listening tests, participants separated a genuine voice from a synthesised one in only 37.5 per cent of attempts, which lands below the result of simple guessing. Hearing evolved to recognize, and recognition is a completely different task from authentication. Familiarity works against the listener, because a familiar voice suppresses the impulse to question anything at all.
Volume followed capability with very little delay. Voice-cloned vishing grew by 1,633 per cent inside a single quarter, which is the signature of a technique moving from demonstration to routine deployment. Cheap generation combined with telephone reach makes the economics obvious for whoever is placing the calls. The countermeasure that survives all of this has nothing to do with listening, since it consists of hanging up and calling back on a number the recipient already holds.
Synthetic identity: a person assembled from fragments
A synthetic identity has no original anywhere. It is constructed from data fragments that may each be genuine, combined into a person who has never existed: a plausible number here, an address there, a date of birth chosen to fit the story. The construct then behaves like a customer. It opens small accounts, pays them on time, accumulates a record, and becomes creditworthy on paper long before anyone asks whether it corresponds to a living human being.
This is exactly why the anti-deepfake toolkit misses it entirely. Comparing a face against a reference photograph assumes that a reference exists and belongs to somebody. A synthetic identity presents a face that matches nothing and contradicts nothing, so the comparison returns silence. Liveness checks can be satisfied by a cooperative accomplice or by a generated stream, and document checks can be satisfied by a well-made artefact.
What catches a synthetic identity is the shape of its history. A real person leaves a long and messy trail across unrelated systems: schooling, employment, medical records, tax filings, family connections, old addresses that overlap awkwardly. A constructed person leaves a short trail that begins abruptly and looks unusually tidy, with no contradictions because nothing accumulated by accident. The verification question moves away from whether the face matches and towards whether the life behind it exists.
Generative forgery: when the artefact carries the authority
The fourth term covers generated objects: invoices, contracts, identity documents, insurance claims, receipts, medical certificates, photographs of damage that never occurred. The subject of the forgery is a thing, and the thing carries authority on its own. A convincing invoice moves money without anybody impersonating anybody, because the recipient trusts the format, the layout and the reference number more than they trust any individual.
Generative forgery scales along a different curve from the other three. A deepfake of a chief executive is expensive in attention, since every target needs research, a plausible pretext and a live operator. A generated document template can be filled ten thousand times by a script and posted into ten thousand claims queues. The check is external in every case: contact the issuing body, look at the record held on their side, and confirm the artefact from the system that created it.
Mixed operations are where the labels collapse in practice. A cloned voice opens the conversation, a generated document supports the story, and a synthetic identity receives the payment at the far end. Calling the whole sequence a deepfake attack hides two of its three components and leaves the corresponding defenses unbuilt. The sequence has to be broken into parts before any part of it can be countered.
Where the borders actually run
Three questions place almost any incident correctly. The first asks whether a real, specific person sits behind the artefact as the source being copied. If the answer is yes, the case belongs to the deepfake family, and the follow-up question is only which channel carried it. If the answer is no, the case belongs to the fabrication family, and the follow-up question is whether the fabricated thing is a person or an object.
The second question concerns what the fake asks the recipient to do. A cloned voice asks for a decision under time pressure, usually a payment or a code. A synthetic identity asks for an account, a credit limit or a set of credentials. A forged document asks for a settlement, a reimbursement or a release of goods. Each request maps onto a different control point, and those control points sit in different departments with different reflexes.
The third question is who is in a position to disprove the thing. For a deepfake, the depicted person and the original recording can. For a voice clone, a callback on an independently known number can. For a synthetic identity, a data trail across independent systems can. For a generative forgery, the body that issued the original document can. Nothing on that list is a detector pointed at the file itself, which explains why detection software on its own settles very little.
Why the statistics rarely name the technique
Reporting systems have only recently started to record whether artificial intelligence was involved at all. The federal complaint center in the United States introduced an AI-related marker for the first time and logged more than 22,000 complaints and roughly 900 million dollars under it. The number is meaningful and fragile at the same time, because the marker depends entirely on the victim recognising what happened and saying so on the form.
The gap shows up inside the same report. Of more than eight billion dollars in investment fraud losses, only 632 million carried the AI marker. Nobody seriously believes that generative tools were absent from the remaining losses. The marker records what was noticed, and noticing requires the person filing to have understood the mechanism, which is a demanding assumption to build a statistic on.
This is the point where loose vocabulary turns into a measurement problem. A victim who calls every synthetic element a deepfake, or who never reaches for any of these words, produces a record that cannot be aggregated with anybody else's. Categories that ordinary people cannot apply correctly return numbers that cannot be trusted at the level of technique. Totals stay useful for scale, while the breakdown by method stays soft and should be read as an approximation.
Impersonation as the container for all four
Whatever the technique, most of these operations sit inside one older and much broader category. Impersonation drew over a million reports and 3.5 billion dollars in losses in a single year, close to one report in every three. The synthetic layer is recent; the request underneath has been the same for decades, since it always comes down to somebody claiming to be a person or an institution the target already trusts. The generative tooling changed the cost of building that claim and left its structure untouched.
Two figures describe the shape of the problem better than any headline total. The typical impersonation report involves a loss of around 700 dollars, so the median event is small and unremarkable. In four impersonation reports out of five there was no monetary loss whatsoever, which means attempts outnumber successes by a wide margin. Large cases exist and they drive the totals, while the daily experience of the system is a constant stream of attempts that fail.
That ratio matters more for anyone building defenses than the spectacular case does. A process tuned exclusively for rare, enormous, deepfake-driven attacks will ignore the steady traffic of ordinary attempts and will be judged by the one case it was designed for. A process tuned for volume, with cheap verification steps applied every single time, catches both kinds. Terminology feeds straight into that choice, because the word an organization uses decides which scenario gets attention and budget.
What precise wording changes in practice
A precise term carries a procedure inside it. Saying voice clone tells the recipient that the medium is audio, the pressure is time, and the answer is an independent callback. Saying synthetic identity tells an onboarding team to examine the depth of a history behind an applicant and to treat a sharp photograph as weak evidence. Saying generative forgery tells a finance function to confirm the record with whoever issued it. A vague word transmits alarm and supplies no route out of it.
The same applies to internal reporting and to explaining an incident afterwards. Precise language lets an organization count episodes by type, notice which channel is under pressure this quarter, and place checks where the pressure actually lands. Vague language merges every episode into a single heap labeled fraud with artificial intelligence. That heap grows steadily and teaches nobody anything, and the next incident arrives against the same defenses as the last one.
The terms will keep drifting, because language always does, and no article stops that. What matters is that the drift stops at the point where a word decides an action. In a headline, deepfake as an umbrella term is tolerable and probably unavoidable. On an incident form, in a bank procedure or in a police report, the difference between a copied voice and a manufactured person is the difference between two entirely separate investigations.
Four terms and the single question that separates each one
Term
What is being copied
The check that settles it
Deepfake
The likeness of a real, identifiable person in video or audiovisual form
Compare with the original recording and ask the person depicted
Voice clone
The timbre and cadence of one specific voice
Call back on a number already held, never on one supplied during the call
Synthetic identity
Nobody. A person assembled from unrelated fragments
Look for depth of history across independent systems
Generative forgery
A document, receipt, certificate or screenshot
Confirm the record with the body that issued it
Edited genuine footage
Nothing is generated, only the context is removed
Find the full source and the minutes surrounding the clip
Mixed operation
Several of the above, arranged in sequence
Split the sequence into components and check each one separately
Questions and answers
What is the difference between a deepfake and a voice clone?
A deepfake copies a real person in video or audiovisual form, while a voice clone copies only the sound of that person speaking. The production methods are closely related, and the practical difference is the channel, because a call gives the recipient no picture, no document and no time to review anything.
Is a synthetic identity the same thing as a deepfake?
No. A deepfake copies somebody who exists, so a genuine original is available for comparison. A synthetic identity is assembled from fragments into a person who has never existed, so there is nothing to compare it against and the check has to examine the history behind the identity.
How much audio is needed to clone a voice?
Current systems produce a convincing clone from a sample of roughly three seconds and reach around eighty-five per cent accuracy against the source voice. Voicemail greetings, short video posts and recorded meetings all supply more material than that.
Can a person tell a cloned voice from a real one?
In controlled listening tests people identified the genuine voice in only 37.5 per cent of attempts. Hearing recognizes familiarity and never authenticates, which is why a callback on an independently known number remains the reliable step.
What counts as a generative forgery?
A generated object that carries authority on its own: an invoice, a certificate, an identity document, a receipt or a screenshot of a transfer. It is verified externally, by confirming the record with the body that supposedly issued it.
Why do fraud statistics rarely say that a deepfake was used?
Because the marker depends on the victim recognising the technique and stating it while reporting. In the first year of an AI-related flag, only 632 million dollars out of more than eight billion in investment losses carried it, which shows that the marker records what was noticed.
Written by Ilia Zavialov, digital security consultant. All glossary terms