exclusive
With artificial intelligence, voices can be cloned deceptively real with little effort. Criminals use this to commit crimes with audio deepfakes, like SWR-show research. The security authorities can hardly keep up.
It was an ordinary Friday when Corinna Busch from Neuss received a call with a blocked number. On the other end a troubled voice: “Mom, it’s me. I caused an accident and killed a woman. I’ll pass you on to the policeman.” This is how Corinna Busch tells it.
The voice undoubtedly sounded like her son’s. The alleged police officer told her on the phone that her son was in custody and that he would release him on bail. After that he hung up. Shocked, her husband Wolf Busch called her son. In the end it turned out: The whole thing was a fake. Her son wasn’t with the police, he was at home.
Because the voice sounded so deceptively real, the Buschs are convinced that someone used artificial intelligence (AI) to clone their son’s voice. However, the investigators were just as unable to determine whether the criminals actually used AI as they were to find the perpetrators.
There are now several online services that make “voice cloning” easily possible using AI. research of SWR-Investigative format full screen now show that criminals also abuse this technology – for example for identity theft or fraudulent shock calls. On request, the Federal Ministry of the Interior stated that the security authorities were “faced with diverse and complex challenges every day” due to technological developments.
Cases of fraud with AI not recorded
An inquiry to the 16 state criminal investigation offices (LKA) shows that the Busch family is not an isolated case. In North Rhine-Westphalia alone, according to police crime statistics, more than 8,000 crimes in connection with shock calls and grandchildren’s tricks were committed in 2022. Most of these cases could not be solved. “I make no secret of the fact that it is difficult to conduct these investigations,” says Daniela Dässel, spokeswoman for the LKA North Rhine-Westphalia, in an interview full screen.
Nationwide, the number of shock calls is not collected, the Federal Criminal Police Office said on request. No LKA records separately how many of these crimes were committed with the help of AI. The advances in AI technology are causing concern for investigators: The LKA Hessen points out that audio fakes were a challenge just a year or two ago. In the meantime, however, there are some websites that are free of charge. It is possible to achieve “good results” with short, high-quality recordings.
Using a recent survey, the tech company McAfee collected figures on the extent of the problem. The company surveyed around 7,000 people from seven countries online about audio fakes using AI. Just over 1,000 of those questioned were from Germany. 22 percent stated that they or acquaintances had already experienced an audio scam with AI. Worldwide it is 25 percent.
It’s that easy to clone voices
According to the survey, AI software can clone voices with an accuracy of 95 percent. One of these software comes from the US start-up ElevenLabs. The company only released a version with German voices at the end of April. That hadn’t happened before. “ElevenLabs” states that the number of users went from zero to more than one million in five months.
Also full screen tested voice cloning. With the help of “ElevenLabs” and an open source program, the editors were able to clone the voice of podcaster Ariana Baborie in a deceptively real way. She was shocked when confronted with the result: “It was very difficult for me to realize that it’s not really me,” she said when she heard her own clone.
For many people, programs like ElevenLabs can become dangerous if misused. The SPD politician Sawsan Chebli is particularly affected by hatred and hostility on the Internet. She goes along with the interview full screen assume that language programs can add a whole new dimension to this problem.
She sees it as a possible weapon in the hands of those who are trying to defame her and try to draw her into the Islamists’ circle: “I actually think that would be brutally dangerous because so many who have a certain image of the enemy would believe it . And if you feed the distorted image with my voice – I think that would be brutal.”
Troll calls to the police
A case from the USA shows that artificially generated voices can even scare people to death. Internet trolls have cloned the voices of Milwaukee, Wisconsin writer Patrick Tomlinson and his partner Niki Robinson. With the fake voices, trolls apparently often call the police and report crimes to the couple’s address. As a result, heavily armed police officers keep storming her house, as surveillance videos show full screen present.
This is called “swatting” and is derived from the word “swat”, i.e. the name for a US special police unit. “In the past few years there have been 42 swatting calls. Most of them in the last ten months,” says Tomlinson in an interview with full screen. According to the couple, the swatting calls began after Patrick Tomlinson tweeted critically about a comedian. When asked what the perpetrators want to achieve with the swatting calls, he replies: “They want us dead.”
Fake not recognized
How do you recognize a voice clone? The tech startup ElevenLabs advertises an “authentication tool” that is supposed to recognize whether an audio track was created with their software. full screen created an audio file with ElevenLabs, slightly post-processed it with an open source program and had it verified by the authentication tool. It didn’t recognize the fake.
ElevenLabs had no comment as of press time. The company prohibits the improper use of its software on its website. It states: Eleven Labs does everything possible to minimize the risk of harmful abuse with protective measures and to take action against it.
EU wants transparency with deepfakes
Politicians have become aware of the dangers of AI and are working on the so-called “Artificial Intelligence Act” at European level. According to a spokeswoman for the EU Commission, this should come into force by May 2024. The draft envisages categorizing AI applications according to risk and regulating them depending on the level of risk.
But there is criticism. “The EU draft provides for transparency obligations for some deepfakes, but it makes exceptions,” criticizes Maria Pawelec. She researches artificial intelligence and disinformation at the University of Tübingen. In her opinion, deepfakes should be marked without exception.
According to the audio forensic expert Patrick Aichroth from the Fraunhofer Institute, the current technology favors the attackers: “Basically, the whole thing is always a game of cat and mouse. As soon as procedures for the detection of forgeries are developed, new forgeries are found that circumvent this detection .”
Judith Simon from the German Ethics Council is also pessimistic about the future. She advises the federal government on how to deal with artificial intelligence. In an interview with full screen she says: “I think we will have to manage a faster learning curve to deal with such technologies.”
The Busch family feels helpless after the shock call with the son’s deceptively real voice, although they had contacted the police. The only thing the police advised them was that they tell as many people as possible so that more people don’t fall for such shock calls.