- The number of AI-generated deepfakes is exploding as tools become more sophisticated and widely available.
- Most deepfakes are pornographic and target everyone from children to celebrities.
- Here are some attempts to stop the explosion of deepfakes.
The risks of artificial intelligence may seem overwhelming. For all the benefits that technology offers, there are also abuses.
One of the big problems posed by AI is deepfakes. Deepfakes are AI-generated videos, images, or audio that are designed to imitate a victim’s unrealistic behavior.
Some deepfakes are superimposed to resemble real video footage, while others are completely computer-generated.
A 2019 study by Deep Trace called “The State of Deepfakes” found that 96% of deepfake videos are pornographic.
Deepfake researcher Henry Ajdar, co-author of the study, told Business Insider that while the numbers may have changed, the problem remains serious.
As AI tools like Dall-E, Stable Diffusion, and Midjourney become more widely available, it is becoming easier for people with little technical knowledge to create deepfakes.
“The overall amount of deepfake content that is pornographic has exploded, although the overall percentage is now going down,” Ajdel said. “We’re talking about millions of victims around the world.”
Although the content is fake, the sense of humiliation, trauma, and intimidation felt by the victims is very real.
A British teenager committed suicide in 2021 after a deepfake porn image of her was created and shared on a Snapchat group by other students at her school, according to a report in BBC News.
Last month, deepfake porn of Taylor Swift began circulating online, prompting Elon Musk to ban searches for the pop superstar on X.
While the situation may seem bleak, there are tools and methods available that can protect against AI identity manipulation.
Deepfake detection
Digital watermarks, which clearly mark content as AI-generated, are being championed by the Biden administration as one solution.
These labels are intended to raise public awareness and make it easier for platforms to scrape and remove harmful fake content.
Google and Meta have both announced plans to start labeling material created or modified by AI with “digital credentials” to make the origin of the content more clear.
OpenAI, the creator of ChatGPT and the image generation tool DALL-E, also supports visual watermarks and hidden metadata that reveal an image’s history, in line with the Coalition on Content Provenance and Authenticity (C2PA) standards. We plan to include both.
There are also dedicated platforms designed to verify the origin of online material. sensitivityThe company behind the 2019 study on deepfakes has developed a detection service that alerts users via email when they are watching content that contains telltale AI-generated fingerprints.
However, even if the image is clearly fake, the subject may still feel victimized since most images still do not have a watermark.
“poison”
Defense tools that protect images from tampering are still in the early stages of development, but are considered a more powerful solution.
These tools offer users the option of processing images with imperceptible signals that, when run on AI-powered systems, create an unusable, blurry, and confusing mess.
For example, Nightshade, a tool created by researchers at the University of Chicago, adds pixels to images that are corrupted when fed into an AI, but still appear as intended to humans.
One of the researchers, Ben Zhao, told NPR, “Nightshade literally adds little poison pills inside the artwork in a way that tries to confuse the training model with what’s actually in the image. You can think about it,” he said.
This technology is designed to protect artists’ intellectual property, but can be used with any photo.
“This is a really good first line of defense for people to feel like, ‘Okay, it’s okay to upload a picture of my friend’s birthday that weekend,'” Ajder said. .
Regulation can make a difference
At least 10 states already have a patchwork of legal protections in place for deepfake victims, according to the Associated Press.
But recent high-profile cases have increased pressure on lawmakers to deter and punish malicious uses of AI and deepfakes.
The Federal Communications Commission has banned AI-generated robocalls after a hoax call using an AI-generated voice resembling Joe Biden’s was made during the New Hampshire primary.
And in January, a bipartisan group of senators introduced a federal bill known as the DEFIANCE Act that would allow victims to sue those who create and distribute sexual deepfakes of themselves. , it is a civil issue rather than a criminal issue.
A bill introduced last May by Representative Joe Morrell to criminalize the sharing of deepfakes has not progressed.
But many of the new laws are at odds with free speech advocates. The reason, according to Henry Ajdel, is that some people see the creation of personal deepfakes as similar to the fantasies someone might have in their head about unrequited love. If no one knows about pornographic content, will anyone actually be harmed?
Crime deterrence
This has implications for UK law, with the Online Safety Act making it illegal to distribute deepfake porn, but making deepfake porn illegal.
“My counterargument is that if you’re creating this content, you’re actually creating something that can be shared in a way that’s not a fantasy that you can’t really share,” Ajder says. .
Although difficult to prosecute, he believes criminalizing deepfake porn remains an important deterrent. Some are using AI tools to create content for personal consumption, Ajdel said. But he added that it was important to ensure that people knew this was a criminal activity to deter people who were simply curious.
Governments can also put pressure on search engine providers, AI tool developers and social media platforms where content is distributed.
In India, a deepfake porn scandal involving a Bollywood actress prompted the government to rush legislation and pressure big tech companies to prevent AI-generated content from spreading online.
“We shouldn’t joke that it’s possible to get rid of the problem completely,” Agider admits. “I think the best thing we can do is create as much friction as possible and try to be very intentional in creating this content.”
