A new study has warned that AI tools may soon be able to manipulate people's online decision-making, raising concerns about the potential for deepfakes and other forms of digital deception. Researchers at the University of California, Berkeley, have discovered that certain types of machine learning algorithms can be used to create highly realistic fake content, such as videos and audio recordings, that are nearly indistinguishable from real things.

The study, which was published in the journal Nature Machine Intelligence, found that these algorithms can be used to create "adversarial" content that is specifically designed to deceive people into making certain decisions. For example, a fake video of a politician saying something they never actually said could be created using one of these algorithms, and then shared online in order to influence public opinion.

The researchers behind the study say that their work has significant implications for the way we think about AI and its potential uses. "We're at a crossroads here," said one of the researchers. "On the one hand, AI has the potential to revolutionize our lives and solve some of the world's toughest problems. But on the other hand, it also poses a very real risk of being used for nefarious purposes."

The study's findings were based on a review of existing research on adversarial machine learning, which is a type of AI that is specifically designed to manipulate or deceive people. The researchers found that several types of algorithms could be used to create fake content that is highly realistic and difficult to distinguish from real things.

One of the most concerning implications of this research is the potential for deepfakes to be used in politics and other areas where decisions are made by people who may not have access to all the facts. For example, a politician who is opposing a candidate could use a deepfake video or audio recording to make it seem like their opponent said something they never actually did.

The researchers behind the study say that they are not advocating for the use of AI to manipulate people's decisions. Rather, they want to raise awareness about the potential risks and to encourage developers to take steps to prevent these types of attacks from being used in the future.

To mitigate these risks, the researchers suggest a number of different strategies. One approach would be to use more robust forms of machine learning that are designed specifically to detect and resist adversarial attacks. Another approach would be to require developers to disclose when they have created or are using AI-powered content that may be used for deception.

Ultimately, the study's findings highlight the need for a more nuanced and informed conversation about the potential uses and risks of AI. As AI continues to evolve and become more powerful, it is likely to play an increasingly important role in our lives - both for good and for ill.

• Nature Machine Intelligence
• University of California
• raising concerns
• people online decision-making
• study