How can I spot AI fakes in images, voices and phone calls?

What deepfakes are and why there are more of them

A deepfake is an image, video or audio recording created or altered with artificial intelligence (AI) to look or sound real. The name combines deep learning, a method by which AI learns, with the word fake.

Germany's Federal Office for Information Security (BSI) distinguishes three techniques for faces: swapping faces in videos, controlling the facial expressions and lip movements of a real person, and inventing entirely new faces of people who do not exist. For voices there are two routes. Either the AI reads any text aloud in the imitated voice, or it converts one speaker's voice into someone else's.

This needs less material than many people think. According to the BSI, a few minutes of good video are enough for a face swap. For voices, the requirement can be cut to a few minutes, and researchers are working on methods that need only seconds. German consumer advice centres (Verbraucherzentralen) observe that such fakes can now be made with free apps and without technical knowledge.

Clues in images and videos

The BSI lists typical flaws, called artefacts, that can appear in manipulated faces:

Visible seams: a border can show around the edge of the face, skin colour or texture may change, and sometimes double eyebrows shine through.

Blurred details: teeth and eyes often look soft because the AI struggles to reproduce sharp outlines.

Limited expressions and inconsistent lighting: some facial expressions or lighting situations come out wrong. When the head turns sharply to the side, the face often becomes blurry or distorted.

The consumer advice centres add lip movements that do not quite match the words, and a voice or intonation that suddenly sounds different from usual.

Their key warning, though, is that deepfakes are getting so good that you should not rely on such technical flaws. A clean image is no proof of authenticity. Looking at the content is often more useful. Criminals use the faces of celebrities to promote supposed investments, crypto offers, diet products or miracle cures. Warning signs include exaggerated promises, no coverage in reputable media or on the person's official channels, and websites without a legal notice or company details.

Clues in voices and phone calls

For artificially generated voices, the BSI describes these possible signs: a metallic sound, mispronounced words, a monotone delivery, intonation or an accent that does not match the real person, odd noises, and a noticeable delay before an answer comes.

A case reported by Bavarian public broadcaster BR shows what this looks like in practice. An elderly woman received a call from a withheld number. Supposedly her daughter was on the line, crying after an accident, and the mother was certain she recognised her voice. A fake police officer then took over and demanded money. The family grew suspicious, called the real daughter and reached her while she was out shopping. Another relative also noticed that the name of a nearby town was mispronounced. The family suspected the voices had been taken from voice messages and social media.

The pattern is the classic shock call, just with a faked voice. According to German police, the perpetrators rely on fear and time pressure and sometimes keep victims on the line until the money is handed over, so they cannot talk to anyone. That pressure is the clearest warning sign, clearer than any audio glitch.

Why detection apps and labels are not enough

There are programs meant to detect deepfakes automatically. The BSI points out, however, that they often only work reliably under certain conditions and can fail against new faking techniques. In a large competition in 2020, the best detection program reached only about 65 percent accuracy, where pure guessing would have given 50 percent. The consumer advice centres write that such tools do not yet offer reliable protection.

In the EU, Article 50 of the AI Act has applied since 2 August 2026. It requires providers to mark AI-generated content in a machine-readable way, and those who use and publish deepfakes generally have to disclose that they are artificial. The European Commission has published guidelines on this. Criminals, however, do not follow such rules. A missing AI label therefore does not prove that something is genuine.

Provenance information such as Content Credentials takes a different approach. It is backed by an alliance called C2PA, which says it includes more than 500 companies, among them Adobe, Google and Microsoft. Such records can show how an image was created and whether it was edited. Our assessment: this helps when a record exists. Many images have none, though, and again the absence of a record is not proof of a fake.

What actually protects you

Because fakes keep improving, it pays to rely less on visual flaws and more on firm habits. For suspicious calls, German police recommend:

Hang up immediately. That is not rude; it gives you time to think.

Call back yourself, using a number you already know, not one given to you during the call.

Agree on a family password that callers must give in an emergency. Police in Upper Franconia also suggest asking questions only the real person could answer.

Never hand money or valuables to strangers, and do not share details about your finances.

If a call seems suspicious, inform the police straight away (110 in Germany), and always report it if you have been scammed.

For images and videos online, the consumer advice centres advise: do not click on the ad or sign up, check whether reputable media are covering the video, take a screenshot of suspicious content and report it via the platform's reporting function.

As prevention, be careful about sharing your own voice and face online. Police point out that publicly shared voice recordings and videos can serve as raw material for fakes. In Germany, shortening or removing your entry in the phone directory also makes it harder for scammers to find you.