Deepfake images continue to cloud social media. Can they be stopped?

Ari Lightman admits he probably would fall for a deepfake.

“I cannot spot a deepfake,” he said.

That might be considered alarming, as Lightman is a Carnegie Mellon University professor of digital media and marketing who teaches classes measuring the impact of emerging technologies. Yet even he says he is not immune to the deception.

A deepfake is a photo, video or audio clip that has been generated by artificial intelligence yet looks or sounds realistic.

Creating deepfakes has become fairly easy, making these images and videos increasingly dangerous as they circulate on the internet and social media platforms.

Just last week, fake sexually explicit images of Taylor Swift were created and shared widely online.

Generative AI expert Bernard Marr said he is “deeply concerned and appalled” by the images of Swift.

“This misuse of technology is not only unethical but also raises significant moral and legal questions,” said Marr, an author and self-described futurist and influencer in business and technology. “It’s a stark reminder of the potential for AI to be used in harmful ways, particularly in violating individual privacy and dignity.”

Despite some safeguards being in place, there still are workarounds for those with “minimal technical knowledge” to create deepfakes, Marr said.

“This accessibility raises concerns about the potential for widespread misuse,” he said.

Lightman said he and others at Carnegie Mellon have followed generative AI tools for years. The Swift incident wasn’t surprising.

“It’s not a question of if (it will happen),” he said. “It’s a question of just when.”

Identifying AI images

The telltale signs of a deepfake image can be found within the person’s veins and coloration, Lightman said, among other attributes.

Attorney Roy Leonard, chair of the Legal Technology and E-Discovery Committee of the Allegheny County Bar Association, cited other possible deepfake signs, such as blurry backgrounds, unnatural shadowing or reflections on people’s glasses.

Because tools generating this content are more sophisticated, people have to become educated on how to recognize possible deepfakes, he said, admitting he’s not sure he could make the distinction.

Creating and enhancing AI platforms has become an arms race of sorts, Lightman said.

“You’re going to get people that have technology checkers and digital verification systems like digital watermarking to try and get better and better and better,” he said. “Hackers will find ways to continue to make AI images that elude the digital and physical fact checkers.”

Some AI images are purely fan-generated with no ill intended. Lightman compared the practice to writing fan fiction and “shipping” characters together.

“Shipping,” which comes from the word relationSHIP, is when fans pair two characters together and “ship” them as a pair — whether it’s platonic or romantic.

“A lot of these images get generated because people want to show their fandom, if you will,” he said. “It could be misinterpreted … and it might cause harm.”

The intent might not be malicious, but how the image is disseminated might change its effect. That is where misinformation comes into play, Lightman said.

Creating AI images can lead to financial harm or harm to reputation, he said.

He said someone could create a sexually explicit image of someone else’s likeness and threaten to disseminate it.

“I think people are looking to test sort of the technology boundaries, if you will, around spotting deepfakes,” he said. “It’s something we need to be very mindful of, be very careful of.”

Legal uncertainty

The images of Swift were “purposely designed to be malicious, for harm,” Lightman said.

There have been lawsuits against AI platforms for copyright infringement. One example accuses OpenAI and Microsoft of violating the New York Times’ copyright.

But lawsuits can be murky as more digital and privacy legislation is enacted.

“This is a whole new space for legal interpretation,” Lightman said.

Marr said he believes there’s an urgent need for more ethical guidelines for AI.

“We’re entering an era where distinguishing between real and synthetic media will become increasingly challenging,” he said.

USA Today identified 10 states that have passed laws banning exploitative deepfake pornography, or AI-generated images, audio files or videos with sexual content. As of now, Pennsylvania is not among them.

Laws always have lagged behind technology, Leonard said. Determining who owns an AI-generated or AI-enhanced image is an evolving area of the law, he said.

The 10 states that took on deepfakes mainly targeted pornographic images or political content, such as using deepfakes to influence an election.

“Creating a false pornographic image or video could have criminal implications, ranging from felony to misdemeanor,” Leonard said.

Sensity AI is a research company that has tracked online deepfake videos since December 2018. It found that, as of 2021, 90% to 95% of such videos are nonconsensual pornographic videos, and about 90% of those are nonconsensual pornographic videos of women, as reported in the MIT Technology Review.

“That, to me, is a red flag and should be a concern to everyone because of how AI-generated images and videos and content are being used to target … women in particular,” Leonard said.

As deepfakes become more mainstream, they are creating what’s called the “liar’s dividend,” Leonard said. The term was coined by law professors Bobby Chesney and Danielle Citron in the California Law Review and refers to how deepfakes can “make it easier for liars to avoid accountability for things that are, in fact, true.”

“On the one hand, we’re concerned with the use of deepfakes in terms of … the creation of false pornographic images, for example,” Leonard said. “But, on the other hand, there are real images that are out there that people are claiming are not real because you can’t believe what you see anymore.”

Deepfake-related legislation should be enacted in Pennsylvania — on the criminal side to deter creators and on the civil side to allow victims to be compensated, he said.

Sen. Lisa Boscola, D-Northampton, proposed legislation to modernize the state’s child pornography law to account for AI, which would add deepfakes to the list of materials prohibited.

The legislation also would establish civil liability for those who use AI and the companies that develop the programs, according to Boscola’s memorandum, written in November.

Future of AI deepfakes

Marr said he believes the landscape of AI and deepfakes likely will be complex.

“On one hand, these technologies hold immense potential for innovation and creative applications,” he said. “On the other, they pose serious ethical and societal challenges.”

Responsible development of AI technologies is essential, Marr said, and can be facilitated by ongoing research, regulation and public discourse.

“The future will likely involve a continuous balancing act between leveraging the benefits of AI and mitigating its risks,” he said.

In the legal sphere, Leonard said, lawyers have a heightened obligation to verify what content at a hearing or trial is authentic.

It’s becoming harder for deepfakes to be spotted, and he said he believes there will come a time when people will stop believing in real content.

“It’s frightening, isn’t it?” Leonard said. “I suspect it’s going to get worse.”