SecurityDeepfakeVerification2026

How to Tell a Deepfake Voice or Video From a Real One

The artifact-hunting method — squinting for the glitch, the wrong shadow, the uncanny blink — is dying, because synthetic media is improving faster than detection and a targeted attack may have no visible flaws. The honest defense is not "can I tell this is fake?" but "can I confirm this through a channel the attacker does not control?" Here is the detection-eroding fast, the callback rule, and why the caller ID and the context are the tells, not the pixels.

General information only — not security advice, and not a reliable deepfake-detection method. The honest premise of this page is that artifact-hunting is dying: synthetic media is improving faster than detection, and a targeted attack may have no visible flaw. The defense this page teaches is channel verification — hang up, call back on a known number — not content inspection. Do not rely on spotting a fake; rely on confirming through a path the attacker does not control.

A perfectly still dark pool reflecting a bare winter tree so precisely that the reflection and the real tree are indistinguishable — The still pool and its flawless reflection are the eroding fast made physical: a generation good enough that you cannot tell the copy from the source by looking. The honest defense is not to squint harder at the water, but to step away and check the tree from another angle — to confirm through a channel the reflection does not control.

Updated September 2026By the AI Help Guide editorial team9 min read

The Eroding Fast

Squinting for the Glitch Is a Losing Game

The popular image of deepfake detection is a forensic expert finding the artifact — the extra finger, the wrong shadow, the uncanny blink, the lip that does not quite match the word. That image is becoming obsolete, and the page that pretends otherwise is selling you a defence that will fail on the day you need it. Generation quality is outpacing detection, and a targeted synthetic call or clip — the kind aimed at one person with one ask — may have no artifact a human can find, because the attacker only needs it to fool you for the thirty seconds it takes you to act.

This page is the media half of a trio; its companion, the phishing guide, is the same problem in text, and the account-recovery guide is the same problem in the process. The shared defense is channel verification: confirm through a path the attacker does not control, rather than inspecting the content they sent you. What follows is the eroding fast stated honestly, the callback rule that survives it, and the inversion — the caller ID and the context are the tells, not the pixels — that the phishing guide already taught you to apply to a message.

The dying method: inspect the content for artifacts — the glitch, the wrong shadow, the uncanny blink — and decide real-or-fake by looking harder.

The surviving defense: do not decide real-or-fake at all. Confirm the claim through a second channel the attacker does not control, and act on the confirmation, not on the media.

The shift is from a perceptual question (can I see the fake?) to a procedural one (can I verify the ask?). The perceptual question is eroding; the procedural question is not, because a real person on a known channel confirms or denies in a sentence regardless of how good the synthetic media was.

The Callback Rule

Hang Up, Then Call Back on a Number You Already Have

The single maneuver that survives the eroding fast is the callback, and it is the channel-break from the phishing guide applied to voice. The rule is short and it is absolute: hang up, then call back on a number you already have — from your contacts, a known card, the official site — never a number the caller gave you, because a number in the attacker's message is part of the attacker's channel.

The callback, in three steps

End the inbound channel. Hang up, or close the call, without acting on the ask. The act of ending the call is the act of leaving the attacker's channel; nothing you do inside it is safe, because the pressure lives inside it.
Re-establish on a known one. Dial the number you already have for that person or organisation — your contacts, the back of your card, the official site you typed yourself. Not a number from the message, not a number the caller read out, not a callback on the inbound line.
Confirm the claim, not the voice. Ask the real question — "did you call me about X?" — and act on the answer. The real person confirms or denies in a sentence, and the synthetic voice is irrelevant to the verification, because you are no longer listening to it.

The callback defeats both the deepfake and the account-recovery attack, because both depend on you acting inside the channel the attacker chose. Step outside it and the attack has no surface left to work on.

The Context Tells

The Caller ID and the Ask Are the Signals, Not the Pixels

If content-inspection is eroding, what is left to look at? The metadata around the media — how they reached you, what they know, what they are asking, and how fast they want it — and these tells are not eroding, because they live in the attacker's behaviour rather than in the synthesiser's output. A perfect synthetic voice still has to make an ask, and the ask, the channel, and the pressure are the signals.

Tell

How they reached youAn unexpected inbound call or video from a number or account you do not recognise, especially one that bypasses your usual channels, is the first signal. The channel the attacker chose is part of the attack.

Tell

What they knowPersonalization is the attack, not the proof — the same inversion the phishing guide names. A targeted synthetic call uses breached or scraped information to sound credible; knowing your name is not authentication.

Tell

What they are askingThe ask is the tell that never erodes. Money, a code, a credential, a transfer, "do this now" — a synthetic voice that asks for nothing is a prank; one that asks for something is an attack, and the ask is what you verify.

Tell

How fast they want itReal-time pressure is the deepfake's enabler, because it is designed to prevent the verification step. The urgency is not a property of the situation; it is a property of the attack.

The signal	What it actually tells you	Why content-inspection fails here
The caller ID / channel	Whether the route is one you recognise and control	A perfect voice can arrive on a spoofed or unexpected channel; the channel is checkable, the voice is not
What they know	Whether the personalization is research, not relationship	Breached data sounds intimate; knowing your name proves nothing about the speaker
What they are asking	The ask is the thing to verify, regardless of the medium	The artifact hunt inspects the delivery; the attack lives in the ask, which no glitch reveals
The real-time pressure	Urgency is the enabler that blocks verification	Pressure is behavioural; a flawless synthetic voice under pressure still has to rush you, and the rush is the tell
The refusal to verify	A real person welcomes a callback; an attacker resists one	Content-inspection cannot see this; the callback test reveals it instantly, because the fake cannot survive a second channel

The table's last row is the cleanest signal and the one that requires no detection skill at all: propose the verification — "let me call you back on your usual number" — and watch the response. A real person welcomes it; an attacker, whose entire method depends on you staying inside their channel, resists it, delays it, or escalates the pressure to stop it. The refusal to be verified is itself the confirmation, and it is a tell that does not erode with generation quality, because it is a property of the attack's structure rather than the media's pixels.

The real-time pressure — the deepfake's enabler, and the reason the callback matters more, not less. A synthetic voice or video is most dangerous in the live moment, because the attacker's goal is not to fool a forensic examination next week; it is to fool you for the thirty seconds it takes you to act before you think.

The pressure — "I cannot talk long," "do this now," "I am in trouble" — is not a property of the emergency; it is a property of the attack, engineered to collapse the verification step that would otherwise defeat it. This is why the callback rule is absolute even when the voice sounds distressed and the situation sounds urgent: the more the moment pressures you to skip verification, the more it is the attack working, and the more the verification is the defence. A real emergency survives a callback; a synthetic one does not.

The trust anchor: change the question from "is this real?" to "can I confirm this through a channel the attacker does not control?"

The eroding question: can I look at this media and tell whether it is fake?

The surviving question: can I confirm the claim through a path the attacker does not control, and act on the confirmation rather than the media?

The anchor is the page's thesis in one swap. The eroding question puts you in a perceptual arms race you are losing. The surviving question puts you in a procedural step you always control, because you choose the second channel. Hold the anchor and the eroding fast stops mattering: a flawless synthetic voice is only as dangerous as your willingness to act on it without confirming, and confirmation is the one thing the attacker cannot synthesise for you.

A worked example: the voice that sounded exactly like my son

Concretely, the eroding fast, the callback rule, and the pressure enabler meeting the call that terrifies most people. A parent receives a phone call from what sounds, unmistakably, like their teenage son: distressed, in trouble, begging her not to tell his father yet, and needing money sent to a number he reads out right now. The voice is perfect. The personalization is perfect — he uses the family nickname, references a real trip. The pressure is perfect — "please, quickly, I cannot talk long."

Running the context tells instead of the artifact hunt reveals the structure. The channel: an unexpected inbound call on a number she does not recognise, bypassing the son's usual contact route. What they know: the nickname and the trip are not intimacy; they are scraped or breached data, the same personalization paradox the phishing guide names, now in voice. The ask: money, to a number read out live, the tell that never erodes. The pressure: "quickly, cannot talk long" — the enabler, engineered to stop her verifying. The refusal test: when she says "let me call you back on your mobile," the voice panics, insists, escalates — and that resistance to verification is itself the confirmation.

The callback rule is what saves her, and it requires no detection at all. She hangs up — leaving the attacker's channel — and calls her son's actual number from her contacts. He is fine, at a friend's house, unaware of any call. The perfect synthetic voice, the real nickname, the real trip, the urgent pressure: all of it collapses in one sentence on a known channel, because the real person confirms or denies instantly and the media is irrelevant to the verification. Had she squinted for the glitch instead, she might have found none — the voice was flawless — and acted anyway. The defence was never in the pixels; it was in the refusal to stay inside the channel the attacker chose.

The Check

A Sensible Default

Hold the eroding fast: artifact-hunting is dying; do not rely on spotting a fake, rely on confirming the claim.
Apply the callback rule absolutely: hang up, call back on a number you already have, never one the caller gave.
Read the context tells — channel, what they know, the ask, the pressure — because they do not erode with generation quality.
Remember the personalization inversion: what they know is research, not relationship; the ask is the signal.
Respect the pressure enabler: the more the moment rushes you to skip verification, the more it is the attack working.
Run the refusal test: propose the callback and watch the response; resistance to verification is itself confirmation.
Swap the question to the trust anchor: not "is this real?" but "can I confirm through a channel they do not control?"
Carry the same channel-break into the phishing and account-recovery pages — it is the cluster's shared spine.

The honest takeaway is that a deepfake is only as dangerous as your willingness to act on it without confirming, and confirmation is the one thing an attacker cannot synthesise for you. The defence is not a sharper eye for the glitch — that eye is losing to the generator — but a procedural step you always control: leave the channel, re-establish on a known one, and verify the claim rather than the media. For the same defense in text, see our phishing guide; for the same defense in the recovery process, see our account-recovery guide; for the account perimeter and the break-glass that a synthetic call might try to talk you out of, see our password-manager guide and our survivable-passwords guide; and for the reporting and incident mechanics after a call has already done its damage, see the official guidance in the sources below rather than a guide page.

A plain landline telephone handset resting in its cradle with the coiled cord visible, on a clean light surface — The handset resting in its cradle, cord coiled and known, is the callback made physical: the line you call back on rather than the one that rang. The inbound call is the attacker's channel; this one is yours, and the verification that defeats the perfect voice happens here, not there.

Common Questions

Frequently Asked Questions

Can I still spot a deepfake by looking for artifacts?

Less reliably every year, and this is the honest premise of the page. The artifact-hunting method — squinting for the glitch, the wrong shadow, the uncanny blink — is eroding because generation quality is outpacing detection, and a targeted synthetic call or clip may have no flaw a human can find. The defense that survives is channel verification, not content inspection.

What is the callback rule?

Hang up, then call back on a number you already have — from your contacts, a known card, or the official site — never a number the caller gave you. The callback defeats both the deepfake and the account-recovery attack, because it re-establishes a channel the attacker does not control, and the real person confirms or denies in a sentence. It is the same channel-break the phishing guide relies on, applied to voice.

The caller knew my name and details — does that prove it is real?

No; the personalization is the attack, not the proof. A targeted synthetic call uses information gathered elsewhere — breaches, social media, a prior message — to sound credible, exactly as a spear-phish does. What they know is not the tell; what they are asking you to do right now, under time pressure, is. The same inversion the phishing guide names applies to the voice.

What if it is a video call rather than a phone call?

The same channel-verification principle applies, and the medium does not change the rule. Confirm through a second, independent channel — a text to their known number, a call to a colleague, an in-person check — before acting on anything the video asks of you. Real-time video raises the stakes and the pressure, which is precisely why the verification step matters more, not less.

The voice sounded distressed and urgent. Surely I should have acted?

The instinct is humane and it is exactly what the attack exploits. Real-time pressure is the deepfake's enabler, engineered to collapse the verification step, so the more the moment rushes you to skip confirming, the more it is the attack working. A real emergency survives a callback in seconds; a synthetic one does not. The callback is not a failure of compassion — it is the only compassion that reaches the real person.

Is there any artifact I can still rely on?

Not as a defence, and the page will not pretend otherwise. Any specific glitch you learn today may be gone in the next model generation, which is the eroding fast in one sentence. The tells that do not erode are behavioural — the unexpected channel, the ask, the pressure, and above all the resistance to being verified — because they live in the attack's structure, not the synthesiser's output, and no improvement in generation quality removes them.

What do I do after a call like that, if I already acted?

Act fast on the known channels, and treat it as an incident rather than a shame. Contact your bank directly on the number on your card, change any credentials you gave, tell the real person the call impersonated, and report it to the official body in your country. The recovery and hardening mechanics are in the account-recovery and password-manager pages; the reporting routes are jurisdiction-specific and live in the sources below, not on a guide page.

Ask our AI. Received a call or clip that sounded or looked exactly like someone you know? Tap the chat icon and describe the context — how they reached you, what they knew, what they asked, and how fast they pushed — and it can help you run the callback and the refusal test without ever asking you to act inside the channel. The defence is procedural, not perceptual; the chat can help you plan the verification, not replace your judgement.

More security & verification guides like this

We treat a deepfake as only as dangerous as your willingness to act on it without confirming, and we swap the eroding perceptual question for the surviving procedural one: leave the channel, re-establish on a known one, and verify the claim rather than the media.

Explore All Guides