GabyA deepfake is a forged video or audio recording created to imitate a real person. Researchers warn that people can detect fewer than a quarter of speech samples created by artificial intelligence. That helps explain why there’s growing concern that criminals and fraudsters could use this technology to steal money from people.
People do not recognize deepfakes very well.
Early synthetic-speech systems often needed thousands of voice samples to generate believable audio. Today’s algorithms can reproduce a person’s voice from a three-second snippet of conversation.
A team of London researchers used an algorithm to create 50 samples of deepfake audio and presented them to 529 participants. The participants correctly identified the fakes in only about 73% of cases. Results improved only slightly after people received training on how to spot aspects of deepfakes.
The study’s author, Kimberley May, says this confirms earlier findings that humans struggle to reliably detect fake speech, even after training to recognize artificial content. She also notes the experiment used relatively old algorithms.
This raises the question of whether people will become even less able to detect deepfakes made with more advanced technologies. For example, Apple recently announced software for the iPhone and iPad that can create a copy of a user’s voice from just 15 minutes of recording.
Between benefits and risks
As the Daily Mail documented, criminals have already used deepfakes. In 2019, the CEO of a British energy company was persuaded to transfer a large sum to a fake supplier after receiving a deepfake audio recording of his boss’s voice.
Despite the risks, the technology can also be beneficial—for example, for people with limited speech or for those who may lose their voice due to illness.
The senior author of the study, Professor Lewis Griffin, says the complexity of AI and the wide availability of these tools puts society on the brink of both significant benefits and serious risks. He recommends that governments and organizations develop strategies to combat misuse.
How to recognize a deepfake?
Deepfake technology, a form of AI, began to emerge around 2014. These systems learn about a target person from photographs and videos, capturing different angles and imitating behavior and speech. Still, there are telltale signs that can separate the real from the fake.
Unnatural eye movement
It’s hard to replicate blinking and natural eye movement. Unnatural or missing eye motions are strong warning signs of a deepfake.
Unnatural facial expressions
If a person’s facial expression looks off or doesn’t match the context, it may indicate that one image has been overlaid onto another.
Misplaced facial features
If a person’s face points one way while their nose or eyes point another, be skeptical about the video’s authenticity.
Lack of emotion
If a person’s expression doesn’t match what they’re saying—flat affect or mismatched emotion—that can be a sign of “facial morphing” or image stitching.
Awkward posture and unnatural movements
Because deepfakes typically focus on faces, body shape and posture are often wrong: the head and body may not align, or movements may look abrupt and disjointed between frames.
Odd coloring, hair, and teeth
Abnormal skin tones, strange lighting, and incorrect shadows can betray a fake. Deepfake algorithms often struggle to render individual hairs or distinct outlines of teeth, so missing detail in hair or teeth can be a giveaway.
Blur or misalignment
Blurry edges or misaligned elements—such as where the face meets the neck—are common artifacts in deepfakes.
Intermittent noise or strange audio
Poor lip-sync, robotic-sounding voices, odd pronunciation, digital background noise, or an inconsistent soundscape can all indicate a deepfake.
Oddness revealed in slow motion
Slow down playback or watch on a larger screen to spot errors. Enlarging the mouth area can reveal whether the lips actually match the speech.
Reverse image search
Use reverse image search to find similar images or videos online; that can help show whether media has been altered. Some advanced forensic tools aren’t publicly available, but access to them can be useful for verifying suspicious content.
