An NPR member station
Play Live Radio
Next Up:
0:00 0:00
Available On Air Stations

AI-generated deepfakes are moving fast. Policymakers can't keep up

An image from a Republican National Committee ad against President Biden features imagery generated by artificial intelligence. The spread of AI-generated images, video and audio presents a challenge for policymakers.
Republican National Committee
An image from a Republican National Committee ad against President Biden features imagery generated by artificial intelligence. The spread of AI-generated images, video and audio presents a challenge for policymakers.

Updated April 27, 2023 at 6:11 PM ET

This week, the Republican National Committee used artificial intelligence to create a 30-second ad imagining what President Joe Biden's second term might look like.

It depicts a string of fictional crises, from a Chinese invasion of Taiwan to the shutdown of the city of San Francisco, illustrated with fake images and news reports. A small disclaimer in the upper left says the video was "Built with AI imagery."

The ad was just the latest instance of AI blurring the line between real and make believe. In the past few weeks, fake images of former President Donald Trump scuffling with police went viral. So did an AI-generated picture of Pope Francis wearing a stylish puffy coat and a fake song using cloned voices of pop stars Drake and The Weeknd.

Artificial intelligence is quickly getting better at mimicking reality, raising big questions over how to regulate it. And as tech companies unleash the ability for anyone to create fake images, synthetic audio and video, and text that sounds convincingly human, even experts admit they're stumped.

"I look at these generations multiple times a day and I have a very hard time telling them apart. It's going to be a tough road ahead," said Irene Solaiman, a safety and policy expert at the AI company Hugging Face.

Solaiman focuses on making AI work better for everyone. That includes thinking a lot about how these technologies can be misused to generate political propaganda, manipulate elections, and create fake histories or videos of things that never happened.

Some of those risks are already here. For several years, AI has been used to digitally insert unwitting women's faces into porn videos. These deepfakes sometimes target celebrities and other times are used to take revenge on private citizens.

It underscores that the risks from AI are not just what the technology can do — they're also about how we as a society respond to these tools.

"One of my biggest frustrations that I'm shouting from the mountaintops in my field is that a lot of the problems that we're seeing with AI are not engineering problems," Solaiman said.

Technical solutions struggling to keep up

There's no silver bullet for distinguishing AI-generated content from that made by humans.

Technical solutions do exist, like software that can detect AI output, and AI tools that watermark the images or text they produce.

Another approach goes by the clunky name content provenance. The goal is to make it clear where digital media — both real and synthetic — comes from.

The goal is to let people easily "identify what type of content this is," said Jeff McGregor, CEO of Truepic, a company working on digital content verification. "Was it created by human? Was it created by a computer? When was it created? Where was it created?"

But all of these technical responses have shortcomings. There's not yet a universal standard for identifying real or fake content. Detectors don't catch everything, and must constantly be updated as AI technology advances. Open source AI models may not include watermarks.

Laws, regulations, media literacy

That's why those working on AI policy and safety say a mix of responses are needed.

Laws and regulation will have to play a role, at least in some of the highest-risk areas, said Matthew Ferraro, an attorney at WilmerHale and an expert in legal issues around AI.

"It's going to be, probably, nonconsensual deepfake pornography or deepfakes of election candidates or state election workers in very specific contexts," he said.

Ten states already ban some kinds of deepfakes, mainly pornography. Texas and California have laws barring deepfakes targeting candidates for office.

Copyright law is also an option in some cases. That's what Drake and The Weeknd's label, Universal Music Group, has invoked to get the song impersonating their voices pulled from streaming platforms.

When it comes to regulation, the Biden administration and Congress have signaled their intentions to do something. But as with other matters of tech policy, the European Union is leading the way with the forthcoming AI Act, a set of rules meant to put guardrails on how AI can be used.

Tech companies, however, are already making their AI tools available to billions of people, and incorporating them into apps and software many of us use every day.

That means for better or worse, sorting fact from AI fiction requires people to be savvier media consumers, though it doesn't mean reinventing the wheel. Propaganda, medical misinformation and false claims about elections are problems that predate AI.

"We should be looking at the various ways of mitigating these risks that we already have and thinking about how to adapt them to AI," said Princeton University computer science professor Arvind Narayanan.

That includes efforts like fact-checking, and asking yourself whether what you're seeing can be corroborated, which Solaiman calls "people literacy."

"Just be skeptical, fact-check anything that could have a large impact on your life or democratic processes," she said.

Copyright 2023 NPR. To see more, visit

Shannon Bond is a business correspondent at NPR, covering technology and how Silicon Valley's biggest companies are transforming how we live, work and communicate.