See how real AI-generated images have been converted

Seeing has not been believing for a long time. The photos have been falsified and manipulated Almost as long as photography has existed.

Now, it doesn’t even require reality to make the photos appear authentic, just artificial intelligence responding to a prompt. Even experts sometimes struggle to know if one is real or not. Can?

The rapid advent of artificial intelligence has raised alarm bells that the technology used to fool people is advancing much faster than the technology that can identify tricks. Tech companies, researchers, photo agencies, and news organizations are scrambling to catch up, trying to set standards for content ownership and provenance.

The advances are already fueling misinformation and being used to stoke political divisions. Authoritarian governments have created ostensibly realistic news stations to further their political goals. Last month, some people fell for images showing Pope Francis wearing a puffy Balenciaga jacket and an earthquake ravaging the Pacific Northwest, even though neither of those events had occurred. The images were created with Midjourney, a popular image generator.

On Tuesday, as former President Donald J. Trump turned himself in to the Manhattan district attorney’s office to face criminal charges, AI-generated images showing actor Bill Murray as president in the White House surfaced on Reddit. Another image showing Trump marching in front of a large crowd with American flags in the background was quickly re-shared on Twitter without the disclosure that accompanied the original post, pointing out that it was not actually a photograph.

Experts fear that technology could accelerate the erosion of trust in the media, in government and in society. If any image can be made and manipulated, how can we believe everything we see?

“The tools will get better, they will get cheaper, and there will come a day when you won’t believe anything you see on the internet,” said Wasim Khaled, chief executive of Blackbird.AI, a company that helps clients fight misinformation.

Artificial intelligence allows virtually anyone to create complex works of art, such as those found now exhibit at the Gagosian Art Gallery in New York, or realistic images that blur the line between what is real and what is fiction. Enter a text description and the technology can produce a related image; no special skills are required.

There are often hints that the viral images were created by a computer rather than captured in real life: the lavishly dressed pope had glasses that seemed to melt on his cheek and blurry fingers, for example. AI art tools also often produce nonsense text. Here are some examples:

However, rapid advances in technology are eliminating many of those flaws. The latest version of Midjourney, released last month, is capable of rendering realistic hands, a feat that notably eluded early imaging tools.

Days before Trump turned himself in to face criminal charges in New York City, footage of his “arrest” circulated on social media. They were created by Eliot Higgins, a British journalist and founder of Bellingcat, an open source research organization. He used Midjourney to imagine the arrest, trial, and imprisonment of the former president in an orange jumpsuit and her escape from him down a sewer. He posted the images on Twitter, clearly marking them as creations. Since then, they have been widely shared.

The images were not meant to mislead anyone. Instead, Higgins wanted to draw attention to the power of the tool, even in his childhood.

The Midjourney images, he said, were able to pass the facial recognition programs Bellingcat uses to verify identities, usually of Russians who have committed crimes or other abuses. It is not hard to imagine governments or other nefarious actors fabricating images to harass or discredit their enemies.

At the same time, Higgins said, the tool also had trouble creating convincing images with people who aren’t as photographed as Trump, like Britain’s new prime minister Rishi Sunak or comedian Harry Hill, “who probably isn’t as well known outside from United Kingdom”.

Midjourney was not amused by any means. She suspended Mr. Higgins’ account without explanation after the images spread. The company did not respond to requests for comment.

The limits of generative images make them relatively easy for news organizations or others attuned to risk to spot, at least for now.

Still, stock photography companies, government regulators and a music industry trade group have moved to protect their content from unauthorized use, but technology’s powerful ability to mimic and adapt is complicating those efforts.

Some AI image generators have even rendered images, a giddy homage to “Twin Peaks”; Will Smith eating handfuls of pasta, with distorted versions of the watermarks used by companies like Getty Images or Shutterstock.

In February, Getty accused Stability AI of illegally copying more than 12 million Getty photos, along with captions and metadata, to train the software behind its Stable Diffusion tool. In his lawsuit, Getty argued that Stable Diffusion diluted the value of the Getty watermark by incorporating it into images that ranged “from the bizarre to the grotesque.”

Getty said the “blatant theft and freeriding” was carried out “on a staggering scale.” Stability AI did not respond to a request for comment.

Getty’s lawsuit reflects concerns raised by many individual artists: that artificial intelligence companies are becoming a competitive threat by copying content they don’t have permission to use.

Trademark violations have also become a concern: Artificially generated images have replicated NBC’s peacock logo, albeit with unintelligible lettering, and show the familiar curvy Coca-Cola logo with additional O’s in the name.

In February, the US Copyright Office weighed in on artificially generated imagery when evaluating the case for “Zarya of the Dawn,” an 18-page comic written by Kristina Kashtanova with art generated by Midjourney. The government administrator decided to offer copyright protection to the text of the comic, but not to its art.

“Due to the significant distance between what a user can command Midjourney to create and the visual material Midjourney actually produces, Midjourney users lack sufficient control over the generated images to be treated as the ‘mastermind’ behind them. “, the office explained. in your decision.

The threat to photographers is quickly outpacing the development of legal protections, said Mickey H. Osterreicher, general counsel for the National Press Photographers Association. Newsrooms will find it increasingly difficult to authenticate content. Social media users are ignoring labels that clearly identify images as artificially generated and choosing to believe they are real photos, he said.

Generative AI could also make it easier to produce fake videos. This week, a video surfaced online that appeared to show Nina Schick, an author and expert on generative artificial intelligence, explaining how the technology was creating “a world where shadows blend with reality.” Ms. Schick’s face then lit up as the camera panned back, showing a body double in her place.

The video explained that the deepfake had been created, with Ms Schick’s consent, by Dutch company Revel.ai and Truepic, a California company exploring broader digital content verification.

The companies described their video, which features a stamp identifying it as computer generated, as the “first digitally transparent deepfake.” Data is cryptographically sealed in the file; Image tampering breaks the digital signature and prevents credentials from appearing when using trusted software.

The companies hope that the badge, which will come with a fee for commercial customers, will be adopted by other content creators to help create a trusted standard with AI images.

“The scale of this problem will accelerate so quickly that it will drive consumer education very quickly,” said Jeff McGregor, CEO of Truepic.

Truepic is part of the Coalition for Content Authenticity and Provenance, a project created through an alliance with companies like Adobe, Intel, and Microsoft to better trace the origins of digital media. Chipmaker Nvidia said last month that it was working with Getty to help train “responsible” AI models using Getty-licensed content, with royalties paid to artists.

On the same day, Adobe introduced its own image generation product, Firefly, which will be trained using only images that are licensed or from its own stock or no longer protected by copyright. Dana Rao, trust director of the company, he said on his website that the tool would automatically add content credentials, “such as a nutrition facts label for images,” that identified how an image had been created. Adobe said it also planned to compensate taxpayers.

Last month, model Chrissy Teigen wrote on Twitter that she had been fooled by the pope’s puffy jacket, adding that “there’s no way I’m going to survive the future of technology.”

last week, a series of new AI images showed the Pope, again in his usual robe, enjoying a tall glass of beer. His hands seemed mostly normal, except for the wedding band on the pontiff’s ring finger.

Additional production by Jeanne Noonan DelMundo, Aaron Krolik and Michael Andre.