Below is a complimentary Demography Unplugged research note written by Hedgeye Demography analyst Neil Howe. Click here to learn more and subscribe.
|
The growth of text-to-image AI generators is spurring fresh concerns about destructive abuse in the hands of saboteurs, terrorists, and criminals. Programs like DALL-E allow users to create and publish, in seconds, a realistic image of (nearly) anything you can imagine. (The Washington Post) |
NH: Three weeks after Russia invaded Ukraine, a shocking video of President Volodymyr Zelensky spread on Ukrainian television and social media. Standing behind a podium, he told the Ukrainian army to lay down its weapons and surrender to Russian forces.
Actually: Zelensky never said any of these things. Russian hackers had made a deepfake. AI had manipulated Zelensky’s mouth to make it look like he was actually speaking the fake audio.
Fortunately for Ukraine, the video was quickly identified, debunked, and taken off most websites by Ukrainian authorities. But we have to wonder, what would have happened if the deepfake had been better sourced into official media sites--and if locals sympathetic to Russia had been better organized to keep publishing it?
When people hear the word "deepfake," they're usually referring to videos. This technology has sown confusion and distrust since it was invented eight years ago.
Until recently, the spread of deepfake videos has been limited by the rare expertise needed to create synthetic media. But now that's changing. Tech companies have released deepfake image generators that are easy for anybody to use.
And these firms are soon expected to release deepfake video generators, making it just as easy for people to create their own imaginary videos.
OpenAI, originally founded by Elon Musk and Sam Altman, is on the cutting edge of this entire industry of AI-generated content. It's all based on natural language input (text prompt) and some sort of text, image, or video output.
Last April, OpenAI released DALLE-2, an advanced artificial neural network based on something called GTP3 ("generative pretrained transformer," version three). The text-to-image AI generator allows users to submit a text description of a scene, and the model will produce several images based on the prompt. It can conjure up anything from a photo of an astronaut riding a horse to an image of protesters outside the Capitol Building. The model generates over 2M images daily from 1.5M users.
How does it work? The researchers give DALL-E a database of images with text descriptions.
After comparing billions of text descriptions with images, it is able to create its own images through text and pixel associations. While the images are strikingly realistic, they fall short of photographic accuracy--because their purpose is simply to convey plausibility. Faces often look slightly distorted and dreamlike.
In response to OpenAI’s success, both Meta and Google have announced they are developing text-to-video AI generators. And smaller startups are also entering the field.
These companies argue that AI-generated content will revolutionize marketing and design. Many of these ventures are motivated by the excitement of showcasing the creativity of the human mind.
Yet it’s easy to imagine how this technology could be used for destructive purposes. Within seconds, anyone can now create fake news: an image of a politician flipping off a grandma or a video of a president telling his army to surrender.
How will we ever know what is real? There are already examples of politicians claiming legitimate videos, like that of George Floyd’s murder, are deepfakes.
So far, the most common use for these technologies has not been the generation of fake news, but porn. A 2019 study found that out of approximately 15K deepfake videos online, non-consensual porn accounted for 96% of them, with creators imposing women’s faces on porn actors’ bodies. Typically famous women are targeted, but everyday people have also been victims. It has even been used to make child pornography.
To prevent misuse, OpenAI and its competitors have added restrictions to what kind of images can be created. For example, users aren’t allowed to generate pictures that are sexual, could influence an election, or feature a public figure. And any shared image must indicate it is created by AI. However, there are already Reddit threads and Discord servers on how to circumvent these rules.
It seems obvious that, given the enormous potential for harm, leaving these companies to police their own users is not enough. More regulations feel inevitable.
But thus far Western democracies have been slow to act. In the United States, any regulation of free expression faces a high bar. The U.S. Supreme Court has held that the government can’t prohibit speech simply because it’s a lie.
Thus, the government can’t ban deepfakes outright. To be sure, a company or individual could be sued for damages if the creations were libelous--in the case of, say, deepfake porn. But libel laws don't usually apply to public figures. Nor would they stop someone from sharing a fake image of protesters at the Supreme Court.
Europe, by contrast, is starting to take action. This summer, the EU adopted new regulations for large tech companies to clean up bots and deepfakes. If signatories, like Meta, Google, Twitter, and TikTok, do not comply, they will be fined up to 6% of their global revenue.
And in the U.K., an independent government commission recently recommended that sharing deepfake porn should be made illegal.
China is cracking down the hardest. It’s a criminal offense to share a deepfake without disclosing its synthetic origin. And in January, the CCP presented an additional set of proposals: Algorithms cannot promote synthetic content, even if its origins have been disclosed. Companies that disobey can be fined up to $16K on a first offense. Future violations could result in criminal prosecution.
Given U.S. free-speech laws, it’s safe to say we won’t be going the Europe or China route anytime soon. But our media ecosystem will be completely defenseless if fake images and videos remain unregulated.
In a 2021 AP poll, nearly all Americans (95%) said that the spread of misinformation is already a problem. Enabling the average person to create photorealistic material is going to make the problem even worse.
In 2019, Texas became the first state to pass a law banning deepfakes--specifically, publishing political deepfakes within 30 days of an election. California followed with a similar ban.
But such laws are unlikely to stand up to a First Amendment challenge. Various congressional bills targeting deepfakes have failed in recent years due to lawmakers' objections that they threatened First Amendment rights.
The most likely form of deepfake regulation that doesn’t run afoul of the First Amendment may be a “clear disclosure” provision, which would require AI-generated media to contain a disclaimer stating that the image or video has been manipulated.
A coalition of tech companies led by Adobe has also developed a new content-verification standard that shows when and where an image was made and how it has been changed. These measures won't stop abuse outright, but they will make it harder for creators to mislead.
| To view and search all NewsWires, reports, videos, and podcasts, visit Demography World. For help making full use of our archives, see this short tutorial. |
* * *
ABOUT NEIL HOWE
Neil Howe is a renowned authority on generations and social change in America. An acclaimed bestselling author and speaker, he is the nation's leading thinker on today's generations—who they are, what motivates them, and how they will shape America's future.
A historian, economist, and demographer, Howe is also a recognized authority on global aging, long-term fiscal policy, and migration. He is a senior associate to the Center for Strategic and International Studies (CSIS) in Washington, D.C., where he helps direct the CSIS Global Aging Initiative.
Howe has written over a dozen books on generations, demographic change, and fiscal policy, many of them with William Strauss. Howe and Strauss' first book, Generations is a history of America told as a sequence of generational biographies. Vice President Al Gore called it "the most stimulating book on American history that I have ever read" and sent a copy to every member of Congress. Newt Gingrich called it "an intellectual tour de force." Of their book, The Fourth Turning, The Boston Globe wrote, "If Howe and Strauss are right, they will take their place among the great American prophets."
Howe and Strauss originally coined the term "Millennial Generation" in 1991, and wrote the pioneering book on this generation, Millennials Rising. His work has been featured frequently in the media, including USA Today, CNN, the New York Times, and CBS' 60 Minutes.
Previously, with Peter G. Peterson, Howe co-authored On Borrowed Time, a pioneering call for budgetary reform and The Graying of the Great Powers with Richard Jackson.
Howe received his B.A. at U.C. Berkeley and later earned graduate degrees in economics and history from Yale University.
