President Joe Biden's administration is pressuring the tech industry and financial institutions to shut down a growing market for abusive sexual images created with artificial intelligence technology.
New generative AI tools have made it easier to turn someone's likeness into a sexually explicit AI deepfake and then share that realistic image in chat rooms and on social media. The victim may be a celebrity or someone else. or children — there's very little you can do to stop it.
The White House Call on Thursday In the absence of federal law, they are seeking voluntary cooperation from companies. By implementing a series of specific measures, officials hope to curb the private sector's ability to create, spread and monetize non-consensual AI-generated images, including explicit images of children.
“When generative AI emerged, everyone was speculating where the first real harm was going to happen, and I think we now know the answer,” said Arati Prabhakar, director of the White House Office of Science and Technology Policy and Biden's chief science adviser.
She told The Associated Press that there has been an “alarming increase” in non-consensual imagery facilitated by AI tools that primarily targets women and girls and threatens to upend their lives.
“If you're a teenage girl, if you're a gay kid, these are issues that people are experiencing right now,” she said. “We've seen an acceleration because of generative AI, which is evolving very quickly, and the quickest thing that can happen is for companies to stand up and take responsibility.”
The document, shared with The Associated Press ahead of Thursday's announcement, calls for action not only from AI developers but also from payment processors, financial institutions, cloud computing providers, search engines and the gatekeepers that control listings in mobile app stores – specifically Apple and Google.
The government said the private sector should work actively to “prevent the monetisation” of image-based sexual abuse, particularly restricting paying access to sites that promote explicit images of minors.
Prabhakar said many payment platforms and financial institutions have already stated that they will not support companies that promote abusive imagery.
“But sometimes it's not enforced. Sometimes there are no terms and conditions,” she said. “This is an example of where enforcement could be more rigorous.”
Cloud service providers and mobile app stores can also “regulate web services or mobile applications that are marketed for the purpose of creating or altering sexually explicit images without the consent of individuals,” the document said.
And whether the images posted online are AI-generated or real nudes, online platforms should make it easier for victims to remove them.
Taylor Swift is the best-known victim of pornographic deepfake images. When AI-generated images of the singer-songwriter being abused began circulating on social media in January, her most ardent fans fought back. Microsoft promised to step up safety measures after it was discovered that some of Swift's images were the product of the company's AI visual design tool.
In the United States and other countries, a growing number of schools We are also working on AI generation. The work depicted deepfake nudes of students, in some cases by classmates who had created AI-edited images and shared them with their classmates.
Last summer, the Biden administration Mediated Voluntary Commitment Big tech companies like Amazon, Google, Meta and Microsoft will have to implement a range of safeguards before releasing new AI systems to the public.
Biden then Ambitious Executive Order The bill, announced in October, aims to guide how companies develop AI so they can make profits without endangering public safety. While it focuses on a wide range of AI concerns, including national security, it also cited the emerging problem of AI-generated child abuse imagery and finding better ways to detect it.
But Biden also said his administration's AI protections need to be backed by legislation: A bipartisan group of senators is currently calling on Congress to spend at least $32 billion on AI development over the next three years. artificial intelligence and funding Measures to guide people safelyBut calls to enact such safeguards into law have largely been put on hold.
Jennifer Klein of the White House Gender Policy Council said encouraging companies to take voluntary steps “doesn't change the fundamental need for Congress to act here.”
Creating and possessing sexual images of children, even if they are fake, is already a crime under longstanding laws. Federal prosecutors earlier this month charged a Wisconsin man with using a popular AI image-generating tool called Stable Diffusion to create thousands of realistic images of minors engaged in sexual acts. The man's lawyer declined to comment after his arraignment hearing on Wednesday.
Yet there is little oversight over the technological tools and services that enable the creation of such images, some of which are hosted on ephemeral commercial websites that reveal little information about their operators or the underlying technology.
Stanford Internet Observatory I said in December Thousands of suspected child sexual abuse images have been discovered in LAION, a massive AI database that is an index of online images and captions and is used to train leading AI image creation tools such as Stable Diffusion.
London-based Stability AI, which owns the latest version of Stable Diffusion, said this week that it “did not approve the release” of the earlier model allegedly used by the Wisconsin man. Such open-source models are difficult to revert because their technical components are publicly available on the Internet.
Prabhakar said it's not just open-source AI technology that's causing harm.
“This is a much broader issue,” she says, “and unfortunately, this seems to be a category where a lot of people are using image generators, and we've seen an explosion in this space recently, but I don't think it breaks down neatly into open source and proprietary systems.”
——
AP reporter Josh Boak contributed to this report.
