This story originally appeared in Digital Future Daily, POLITICO’s newsletter about how technology is redefining global power.
Subscribe here.
The impact of deepfakes on society — and elections particularly — has been
an anxiety for years. Easy-to-use generative AI tools have recently moved it from an issue in niche areas to a top security risk across the board. Before the Biden robocall, AI deepfakes were used in attempts to disrupt elections in
Slovakia and
Taiwan.
Congress has taken note. Sen. Mike Rounds (R-S.D.) told POLITICO that as the Senate hashes out its priorities on AI legislation, there is growing recognition that tackling the use of AI in campaign ads and communications should top the list. And after explicit deepfakes of Taylor Swift spread on X last week, lawmakers
renewed calls for urgent legislation on the issue.
A reminder: no federal laws currently prohibit the sharing or creation of deepfakes, though several bills have been proposed in Congress and some states have passed laws to crack down on manipulated media. The Federal Election Commission, too, has been
considering rule changes to regulate the use of A.I. deepfakes in campaign materials.
“Deepfakes is the first test that generative AI has thrown at us because it fundamentally eliminates all trust,” Vijay Balasubramaniyan, CEO of the phone fraud detection company Pindrop, told Steven Overly on a
POLITICO Tech podcast episode that delved into the Biden robocall incident. “If we can’t get together and figure out how to solve that problem, yeah, the killer robots will definitely get us.”
No surprise that’s easier said than done. One especially tricky part will be figuring out how to tackle the full range of manipulated media — from older techniques like splicing in fake audio to the new generative AI-fueled advancements, and all the hybrids in between. The robocall, for one, was not a very advanced audio deepfake, according to Matthew Wright, who chairs Rochester Institute of Technology’s cybersecurity department.
“There are tools available now that that can do a better job, and consequently be more dangerous,” he told DFD.
Looking at the proposed federal bills and enacted state laws, it turns out there’s not a whole lot they collectively agree on, starting with even what should be regulated.
California and Washington’s laws target false depictions only of political candidates, while Texas and Minnesota go further to include those created with the intention of harming a political candidate or influencing election outcomes.
Consensus on what constitutes a deepfake is also lacking. Some bills distinctly cover images and video, while others extend to audio.
“This episode does highlight how important it is to have audio be included in these efforts,” said Mekela Panditharatne, counsel for the Brennan Center’s Democracy Program. “It could be kind of separated and done piecemeal. But I do think it makes sense to consider those different forms of gen-AI production together.”
Piecemeal seems to be the way regulation on deepfakes is moving. Wright drew parallels with the landscape for privacy legislation, where a patchwork of laws offer varying levels of protection.
A key question is who should be held accountable: phone service providers, platforms, developers or distributors of the deepfakes? How you answer that ends up defining the focus of proposed solutions.
At the federal level, bills have assigned responsibility to two main groups, said Panditharatne. The first includes the actors that fall under campaign finance disclosure requirements: campaigns, super PACs and donors. Often, the resulting bills address the timing of deepfakes — like
one act that bans false endorsements and knowingly misleading voting information 60 days before a federal election — or transparency, as in the case of Rep. Yvette Clarke’s (D-N.Y.) bill which requires that political ads reveal their use of AI-generated material through mandatory labeling, watermarking or audio disclosures.
The second category targets deepfake disseminators, so long as they meet certain knowledge or intent requirements in some cases. Rep. Joe Morelle’s (D-N.Y.) Preventing Deepfakes of Intimate Images Act would make it illegal to share deepfake pornography without consent.
“There is relatively little attention both at the federal and state level in holding other actors to account for deepfakes,” Panditharatne added, giving social media companies and AI developers as examples.
As with past content moderation issues, social media giants enjoy some protection from legal liability under federal law (thanks to the famous Section 230 of the Communications Decency Act), which complicates such efforts. The bipartisan
Senate NO FAKES Act is one attempt; it proposes holding liable anyone who makes or publicly shares an unauthorized digital replica — including companies — and allowing for penalties that start at $5,000 per violation.
Still, it’s unclear to Wright whether any regulations under consideration, or industry solutions in development, could have prevented the Biden robocall. Wright said he has built a deepfake detection tool of his own, but also offers one solution for which the technology does not currently exist on phones. “Every microphone is going to have to have even live audio being constantly re-certified. That might have to be what’s required.”
The design of the scheme exploited an area on which detection focuses less: a direct line with no real-time feedback from social media and limited playback capabilities.
Enforcing the regulations being floated will require some sort of detection mechanism (
many
have
been
invented). But for now, some bad actors with just a voter registration list, phone, and 30-second clip of a political figure can inevitably fly under the radar. The FTC has
sponsored a challenge with a $25,000 top prize for the most effective approach to safeguard against the misuse of AI-enabled voice cloning, covering everything from imposter fraud to using someone’s voice without consent in music creation. Its suggestions include real-time detection and monitoring to alert users to voice cloning or block calls.