Uncensored AI art model prompts ethics questions – TechCrunch

a new open source ai image generator Able to create realistic images from any text prompt has seen an astonishing speed in its first week. The steady proliferation of Stability AI, high fidelity but capable of running on off-the-shelf consumer hardware, is now being used by art generator services such as ArtBreeder, Pixelz.AI and more. But the unfiltered nature of the model means that not all uses are completely above board.

For the most part, use cases have been above board. For example, NovellAI is experimenting with static dissemination to produce art that can accompany AI-generated stories created by users on its platform. Midjourney has launched a beta that taps stable diffusion for greater photorealism.

But the stationary spread has also been used for less flavorful purposes. On the infamous discussion board 4chan, where the model quickly leaked, several threads are devoted to AI-generated art of nude celebrities and other forms of generated pornography.

Stability AI CEO Imad Mostak called it “unfortunate” that the model leaked on 4chan and stressed that the company was working with “leading ethics and technologies” around responsible releases on security and other mechanisms. One of these mechanisms is an adjustable AI tool, the Safety Classifier, included in the overall static diffusion software package that attempts to detect and block offensive or undesirable images.

However, the security classifier – while on by default – can be disabled.

Steady diffusion is very new territory. Other AI art-generating systems, such as OpenAI’s DALL-E 2, have implemented stricter filters for pornographic content. (The License Stable diffusion for open source restricts some applications, such as exploiting minorities, but does not tie the model to a technical level.) Also, many do not have the ability to create art of public figures, unlike static diffusion. Is. Those two abilities can be risky when combined, allowing bad actors to create obscene “deepfakes” that—worst case scenarios—can perpetuate abuse or be implicated in a crime they didn’t commit. Is.

A deepfake by Emma Watson created by Stable Diffusion and published on 4chan.

Women, unfortunately, are by far the most likely to be victims. A study done in 2019 Turns out, out of 90% to 95% of non-consensual deepfakes, about 90% are women. According to Ravit Dotan, AI ethicist at the University of California, Berkeley, this bodes well for the future of these AI systems.

“I worry about the other implications of synthetic images of illegal content – ​​that it will exacerbate the illegal behavior that is depicted,” Dotten told TechCrunch via email. “For example, the synthetic child would be [exploitation] Authentic child enhancement [exploitation], Will this increase the number of attacks by pedophiles?”

This view was shared by lead researcher Abhishek Gupta from the Montreal AI Ethics Institute. “We really need to think about the life cycle of an AI system that includes post-deployment use and monitoring, and think about how we can envision controls that will minimize the damage even in the worst case scenario. can,” he said. “This is especially true when a powerful ability [like Stable Diffusion] Goes wild that can cause real trauma to those against whom such a system can be used, for example, by creating offensive material in the likeness of the victim.

some of a Preview In the past year, when, on the advice of a nurse, a father took pictures of his young child’s swollen genital area and texted them to the nurse’s iPhone. The photos were automatically backed up to Google Photos and flagged as child sexual abuse material by the company’s AI filter, resulting in the person’s account being disabled and an investigation by the San Francisco Police Department.

If a legitimate photo can trip such a detection system, experts like Doton say, there’s no reason why deepfakes — and massive — generated by systems like Stable Diffusion.

“The AI ​​systems that people create, even if they have the best of intentions, can be used in harmful ways that they don’t anticipate and can’t prevent,” Dotton said. “I think developers and researchers often appreciate this point.”

Of course, the technology to create deepfakes has existed for some time, AI-powered or otherwise. A 2020 report good Deepfake detection company Senseity found that hundreds of explicit deepfake videos featuring female celebrities were being uploaded to the world’s largest pornography websites every month; The report estimated the total number of online deepfakes to be around 49,000, of which over 95% were porn. Actresses including Emma Watson, Natalie Portman, Billie Eilish and Taylor Swift have been the target of deepfakes since AI-powered face-swapping tools went mainstream several years ago, and some, including Kristen Bell, have spoken out against what they say. lets see. Sexual Exploitation,

But static diffusion represents a new generation of systems that can create incredibly – if not completely – convincing fake images with minimal work by the user. It’s also easy to set up, requiring few setup files and no graphics card costing several hundred dollars on the high end. Work is underway on even more efficient versions of the system that can run on the M1 MacBook.

steady spread

A Kylie Kardashian deepfake posted on 4chan.

Sebastian Burns, a Ph.D. Researchers from the AI ​​group at Queen Mary University of London believe the potential for scaling up automation and customized image generation are the big differences – and the main problems – with systems like stable diffusion. “Most lossy imagery can already be produced by conventional methods, but it is manual and requires a lot of effort,” he said. “A model that can produce near-photorealistic footage could give way to personal blackmail attacks on individuals.”

Burns fears that personal photographs scraped from social media could be used to generate static dissemination or any such models target pornographic imagery or images depicting illegal acts. There is an example. Indian investigative journalist Rana Ayyub, after reporting on the rape of an eight-year-old Kashmiri girl in 2018. became Targeted by Indian nationalist trolls, some of whom created deepfake porn on another person’s body with his face. The deepfakes were shared by the leader of the nationalist political party BJP, and as a result the harassment that Job received became so bad that the United Nations had to intervene.

“Stable spread provides enough adaptation to send automated threats against individuals either at risk of being faked or at risk of potentially harmful footage being published,” Byrne continued. “We already see that people are being extorted after accessing their webcams remotely. That step of infiltration may no longer be necessary.”

With steady proliferation in the wild and already being used to generate pornography – some non-consensual – it may be up to image hosts to take action. TechCrunch reached out to OnlyFans, one of the major adult content platforms, but did not hear back as of publication time. A spokesperson for Patreon, which also allows adult content, said the company has a policy against deepfakes and rejects images that “reproduce the likeness of celebrities and use adult references to non-adult content.” I keep it.”

If history is any indication, however, enforcement would likely be disproportionate – partly because some laws specifically protect against deepfaking as it pertains to pornography. And even though the threat of legal action pulls down some sites dedicated to offensive AI-generated content, there’s nothing to stop new ones from popping up.

In other words, says Gupta, it’s a brave new world.

“Creative and malicious users can abuse capabilities [of Stable Diffusion] To generate thematically offensive content on a large scale, using minimal resources to run estimates – which is cheaper than training an entire model – and then use them to drive traffic and attract attention to places like reddit and 4chan Publish it for you,” Gupta said. “A lot is at stake when such capabilities survive “in the wild”, where controls such as API rate limits, security controls on the type of output returned from the system are no longer in force.

Deepfakes for all: Uncensored AI art model prompts ethics questions

Leave a Comment