As the world grapples with deepfakes, AI companies agree to a set of principles

Is this some sort of solution to the onslaught of artificial intelligence (AI)-generated deepfakes and child sexual abuse material? are using AI tools to create “authentic” manipulated content. The nonprofit Thorn and All Tech is Human have successfully brought AI companies to the table to create new AI standards, especially for child safety.

These principles include reconsidering the training data used for AI models, as well as new detections such as watermarking AI generations and preventing generation tools from creating AI-generated child sexual abuse material (AIG-CSAM). Contains the need for solution development. . To date, 11 technology companies have registered, including Meta, Google, Anthropic, Microsoft, OpenAI, Stability.AI, and Mistral AI.

“We are facing a rare moment, an opportunity, to stay on the right path with generative AI and ensure that children are protected as the technology is built,” Thorne said. In a statement on “AI Safety by Design.'' Principles of AI. The intention to expand the scope remains. “In the coming weeks, we will be adding companies to the list of major industry players committed to Safety by Design’s generative AI principles,” said David Polgar, Founder and President of All Tech Is Human.

Thorne's window of opportunity may be closing quickly. A recent example: In March, when we reviewed Meta's ad library for Facebook and Instagram, we found that an app called Perky AI was used after the two platforms showed ads containing blurred, fake nude images of young celebrities. pointed out.

Sen. Richard Blumenthal (D-Conn.), the committee's chairman, said at a hearing of the U.S. Senate Privacy, Technology, and Law Subcommittee that “probably, hypothetically, it's a science fiction thing that will happen at some point in the future.'' It's not,” he summed up. , early this month. The hearing discussed Perky's advertisement.

Recently, deepfake images were found to be appearing as part of search results for certain search phrases on Microsoft Bing and Google Search. Both companies subsequently confirmed that the material had been identified and removed.

Key principles that AI companies have signed on to include taking a serious look at the datasets used to train AI models. Central to this is the early detection of her CSAM and child sexual exploitation material (CSEM) from these models. Meta, Microsoft, and others target risks to children “alongside the adult sexual content in our video, image, and audio production training datasets.”

“This effort represents an important step forward in preventing the misuse of AI technology to create or disseminate child sexual abuse material and other forms of sexual harm against children.” Microsoft Chief Digital Safety Officer Director Courtney Gregoire said.

In December, researchers at Stanford University's Internet Observatory announced they had discovered more than 1,000 images of child exploitation in a popular open-source image database called LAION-5B. This is used to train generative AI tools such as Stable Diffusion 1.5, a popular and incredibly realistic text-to-image generator. Stability AI did not create or manage that database, but it was immediately removed from the process.

AI companies recognize the importance of identifying generated content from actual content, and in the case of objectionable content generated by malicious parties with harmful intent, identifying the source. We understand that this is important. Meta, Microsoft, Google, and others claim to be working on solutions that embed signals in content as part of the process of generating images, audio clips, or videos.

In its pledge, Stability AI states that its purpose is to “prohibit the use of deceptive generative AI for the purpose of sexually harming children and explicitly ban AIG-CSAM from our platform.” It states that there is.

Another solution is to watermark AI-generated content to identify its author and source. Progress has been made on this front. For example, Meta and OpenAI confirmed earlier this year that every generation on their platform will include a watermark or label. “This measure will help distinguish between human and synthetic content, which is important to protect user privacy and combat the spread of deepfakes,” said White & Brief Advocates & Solicitors founder and CEO Managing Director Nilesh Tribhuvann told HT at the time.

There's also the Content Provenance and Authenticity Coalition for Content Provenance and Authenticity (C2PA), led by Adobe and including Google, Microsoft, Intel, Leica, Nikon, Sony, and Amazon Web Services, which requires all content produced to be “content authenticated.” We encourage people to seek information.