# How do you spot a deepfake or a cloned voice in 2026?

Date: 2026-10-28 · Author: Robert Scales · https://guidetoartificialintelligence.com/en/learn/deepfakes-voice-cloning-c2pa-2026/

**Answer:** In 2026 you often can't spot a good deepfake or cloned voice by eye or ear, so the reliable defense is to verify through a separate channel: call the person back on a number you already have, use a family code word, and check where a file came from. Content credentials and watermarks help when they are present, but they can be stripped, so their absence proves nothing.

A **deepfake** is an image, video or audio clip made or altered with AI to show someone saying or doing something they never did. A **cloned voice** is a synthetic copy of a real person's voice, made from a short recording. Both are now cheap, fast and convincing enough that the old advice, look for odd hands or listen for a robotic tone, catches only the sloppy ones.

This is an update to the site's [2024 guide to spotting deepfakes](/en/learn/deepfakes-spot-them/). The short version: shift from detecting to verifying.

## Why your eyes and ears are no longer enough

The FBI put it plainly in a 2025 warning about criminals who sent AI-generated voice messages pretending to come from senior U.S. officials: AI content is often difficult to identify, and a cloned voice can sound nearly identical to someone you know. The bureau still lists visual clues worth a glance, such as distorted hands or feet, irregular faces, odd glasses or jewelry, wrong shadows, unnatural movement, or a lag in a voice call. But a clean result on those checks tells you very little.

The scams follow a familiar pattern. A message arrives that seems to come from a boss, an official or a family member. It builds trust, then asks for something: money, a gift card, a two-factor code, a click on a link, or a move to another messaging app. The AI makes the voice or face believable; the request is where the harm happens.

## What does work: verify through another channel

Instead of judging the media, check the person and the source:

1. **Call back.** Hang up and call the person on a number you already had, not one given in the message. The FBI recommends exactly this.
2. **Use a code word.** Agree on a secret word or phrase with family members that only you know, and ask for it in any urgent call.
3. **Slow down.** Urgency, secrecy and unusual payment methods are the real warning signs, whatever the voice sounds like.
4. **Trace the file.** For a photo or video, look for the original source: who first posted it, and whether reliable outlets show the same scene from other angles.
5. **Protect accounts.** Turn on multi-factor authentication and never read a code aloud to a caller.

## A worked example: the urgent call from a grandchild

A grandmother gets a call. The voice is her grandson's, panicked: he has been in an accident and needs money wired today, and please don't tell his parents.

| Step | What she does | Why it works |
|---|---|---|
| 1 | Says she will call right back, and hangs up | Breaks the pressure of the moment |
| 2 | Calls her grandson's own cell number | A cloned voice cannot answer his real phone |
| 3 | If he doesn't answer, calls his parents | The request for secrecy is itself a red flag |
| 4 | Asks for the family code word if the caller calls again | A clone trained on public videos won't know it |
| 5 | Reports the call to police or the anti-fraud center | Helps others targeted by the same group |

None of these steps requires deciding whether the voice was real.

> **Try it:** Pick a family code word this week and share it in person with the people most likely to be impersonated, such as children, grandchildren and parents. Practise the call-back rule once with a pretend call.

## What are content credentials (C2PA) and watermarks, and do they work?

**Content credentials** are an open standard from C2PA (Coalition for Content Provenance and Authenticity), an industry group whose founders include Adobe and Microsoft. A camera, editing app or AI generator that supports the standard attaches a signed record to a file saying where it came from and what was done to it, such as "generated with an AI model" or "cropped in an editor." The record is protected by cryptography: if someone tampers with it, the signature breaks and verification tools flag it.

A **watermark** is different: a signal hidden in the pixels, sound or text itself, often invisible, that a detector can later read. Google's SynthID is a well-known example. Watermarks can survive when a file is copied or screenshotted, which is exactly when the attached credentials tend to be lost.

Both help, within limits. The C2PA's own explainer stresses that credentials don't judge whether content is true; they show only who signed the record and that it hasn't been altered. It also says plainly that the record can be removed, which is why the standard now pairs credentials with watermarks or fingerprint lookups so a stripped file can be matched back to its record. A 2024 report from the U.S. National Institute of Standards and Technology (NIST) reaches the same conclusion from the other side: metadata is often stripped as content moves across platforms, and watermarks can often be removed too.

So use them one way only. A valid credential saying a photo came from a news agency's camera, or that an image was made with AI, is useful evidence. A file with no credential or watermark proves nothing either way, because most content still carries none and bad actors can remove them. In Europe, the AI Act's transparency rules, which apply from August 2026, require providers of generative AI to give their output a machine-readable marker, with systems already on the market given until December 2, 2026 ([details here](/en/learn/applies-under-eu-ai-act-august-2026/)). That should make labels more common. Until they are everywhere, the call-back habit remains your best protection.

## Sources

- C2PA and Content Credentials Explainer (version 2.3): https://spec.c2pa.org/specifications/specifications/2.3/explainer/Explainer.html
- Reducing Risks Posed by Synthetic Content: An Overview of Technical Approaches to Digital Content Transparency (NIST AI 100-4): https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.100-4.pdf
- Senior US Officials Impersonated in Malicious Messaging Campaign (I-051525-PSA): https://www.ic3.gov/PSA/2025/PSA250515