Deepfakes – how to recognise AI fakes

Whether it’s a faked video, a bogus audio recording or a manipulated image, deepfakes are, at the same time, both fascinating and frightening. How do you recognise deepfakes and what should you do if your content has been misused? We investigate.

Have you seen the videos of Barack Obama ranting and the Mona Lisa talking? Have you seen the Pope in a thick down coat and Donald Trump fleeing from the police? Deepfakes, i.e. videos, images or audio recordings that have been faked using artificial intelligence (AI), are appearing with increasing frequency and seem astonishingly real. With the right online tools and apps, it is easier than ever to manipulate faces and voices and circulate them via social media.

How deepfakes work

Deepfakes (from the words “deep learning” and “fake”) use artificial intelligence to substitute faces and voices in videos, images and audio recordings. This can be quite entertaining. For example, when voice conversion lets you suddenly sing like Marilyn Monroe. Or when you use face swaps to give yourself a six-pack like Schwarzenegger. Using both techniques together opens up endless possibilities – and presents many dangers, since deepfakes are used in fake news and scams.

Recognising deepfakes

Fortunately, it is still relatively easy to expose deepfakes. But the fakes are getting better and better. It is therefore important that we train our perception. Pay attention to the context and discrepancies:

1. Dubious source: Who posted the video/photo? Can you find the photo or video on reputable news portals?
2. Unnatural details: Are the facial expressions and pronunciation correct? How does the person blink? Can you see any edges or illogical shadows on the face?
3. Differences in quality: Is the resolution of the face higher than in the rest of the crop? Are there any pixelated or blurred transitions?
4. Poor lip synchronisation: Are the lip movements in sync with the spoken words? Does the person have a monotone, metallic voice?
5. Overall image: Does the person shown behave and talk realistically? Would this person really say something like that?

Some Platforms and tools could also be helpful. However, these are never completely reliable. When in doubt: if something seems wrong to you, it probably is.

How to counteract deepfakes

What should you do if fraudsters have used deepfakes to clone your voice or face ID, or if your content has been used for a deepfake without your consent?

1. Collect evidence: Take screenshots or save copies of the fraudulent content. These are useful if you want to take legal action later.
2. Report the incident: Notify the platform or website on which the deepfake was published. Most platforms have guidelines for dealing with this and will remove the content.
3. Get legal help: Creating and distributing harmful deepfakes is not (yet) illegal in Switzerland. However, depending on the content, deepfakes may infringe copyright or privacy rights and can damage your reputation. It is best to seek advice through your legal protection insurance.