How to — tell if the person on the video call is real

You will finish this with a rule you can run on any video call in under a minute: which requests must be verified, what to ask on the spot, where to confirm them — so a face on a screen never carries a payment or a password by itself.

That rule is now necessary because your eyes have lost. Tavus, a video-conversation startup, ran a one-minute call study with its Griffin model and reports that 48% of participants believed they had been talking to a real person; its previous systems topped out at around 2%. Text got there first — a UC San Diego study found GPT-4.5 judged human in 73% of five-minute chats, more often than the actual human in the room. Video catching up is the story of the past year, and it is why "I'd know a fake if I saw one" stopped being a plan. If you want the background on what you're actually up against, AI 101 — What is a deepfake? walks through how these are made and why detection keeps losing ground.

1. Decide in advance what must be verified

The attacks that work don't fool you for an hour; they only need to fool you at the moment you decide. In January 2024, an employee at a Hong Kong firm sent $25 million after a video call on which her chief financial officer and colleagues turned out to be deepfakes — Deloitte's write-up of the case notes that she was never on a call with any of them. She verified nothing because nothing in her routine said to.

So write the trigger list before you need it: any instruction to move money, change bank details, approve a payment, or hand over a credential gets verified out-of-band, whoever appears to be asking — your CEO, your bank, your mother. Once the list exists, the call itself stops being the place where you make the judgment.

2. Ask for something you didn't announce

An unplanned request is the cheapest test you have. Ask them to turn toward the window, hold up a number of fingers you name, or answer something only that person could answer — last week's decision, the name on the whiteboard, the thing you argued about on Tuesday.

Be honest about what it proves. Replayed and face-swapped footage dies on an unplanned prompt; real-time systems improvise, which is exactly what Tavus's model does. Treat a smooth answer as "not obviously fake" rather than "confirmed real," and treat a refusal or an excuse to skip it as a serious signal. A real person spends two seconds turning around; a stalled feed spends them apologizing.

3. Confirm on a channel you already trust

This is the step no product can replace, and it is why the Hong Kong case was avoidable: end the call or pause the request, then reach the person on contact details you already had — the number in your phone, the email thread that predates this conversation, the colleague sitting two desks away. Never the number or link that arrived with the request. If your finance lead asks for a transfer, the confirmation happens somewhere they did not choose for you.

4. Read the platform's proof signals, and know their limits

Video products are starting to carry verification of their own. Zoom has opened a beta for World ID's Deep Face check, aimed at financial approvals, healthcare calls and executive decisions: a host can require it in the waiting room or request it mid-meeting, and a verified participant gets a "Verified Human" badge on their tile, with the comparison done on-device. Three honest caveats — it is an enterprise beta, not a default; a badge proves a verified human is present, not that the request in front of you is legitimate; and the absence of a badge proves nothing yet, because most callers have never enrolled. For files and recordings rather than live people, the equivalent signal is content provenance and marking, covered in What is AI watermarking?.

5. When you can't verify, slow the room down

If the person can't be confirmed on a second channel, the answer is a delay, not a decision. Attacker urgency is the tell: a real colleague, a real bank and a real relative tolerate five minutes while you check; a scam depends on the window staying open. Say you're confirming and will call back, then initiate the callback yourself. Escalating costs you a coffee; guessing costs someone a payroll.

Don't do this

Don't hunt for artifacts — the stutter in the teeth, the eyes that blink wrong, the background that swims. That is the one part of the stack improving fastest (2% to 48% in a single model generation), and it makes you the worst kind of confident: the participants in Tavus's study who grew suspicious usually did so in the first 20 seconds, which is less detection than luck. And don't paste a screenshot of a call into a random "is this AI?" detector to settle it — accuracy across those tools varies wildly, and you have just uploaded a private frame to a site you'd never heard of this morning.

How you'll know it worked

Run the rule once on something low-stakes: one unplanned question, one confirmation on a channel you already had. Then wait for the real test — an urgent request arriving on a call. You should feel the pause become automatic before you feel the decision, and the proof is that the request waits while you verify it, then goes through unchanged. A rule you have exercised once is a rule you'll actually use when the amount is large.

Have you had a video call where you genuinely weren't sure the person was human? What tipped you off? Tell us in the comments.