Imagine it’s late October 2020, and that there’s fierce competition for the remaining undecided voters in the presidential election. In a matter of hours, a deepfake video depicting a candidate engaged in unsavory behavior goes viral, and thanks to microtargeting, reaches those who are most susceptible to changing their vote. Deepfakes—the use of AI to generate deceptive audio or visual media depicting real people saying or doing things they did not—are a serious threat to democracy and lawmakers are aggressively responding. Unfortunately, their current efforts will largely be ineffective.
Brandie M. Nonnecke (@BNonnecke), PhD, is founding director of the CITRIS Policy Lab at UC Berkeley and a fellow at the Aspen Institute’s Tech Policy Hub and the World Economic Forum. She studies human rights at the intersection of law, policy, and emerging technologies, with her current work focusing on issues of fairness and accountability in AI.
Last month, Governor Gavin Newsom signed California’s AB 730, known as the “Anti-Deepfake Bill,” into law. The intention to quell the spread of malicious deepfakes before the 2020 election is laudable. But four major flaws will significantly impede the law’s success: timing, misplaced responsibility, burden of proof, and inadequate remedies.
The law applies only to deepfake content distributed with “actual malice” within 60 days of an election—a forced time constraint that does not reflect the enduring nature of material posted online. “What happens if content is created or posted 61 days before an election and remains online for months, years?” asks Hany Farid, a professor and digital forensics expert at UC Berkeley who works on deepfake detection.
To ensure that the law does not infringe upon free speech rights, it incorporates exemptions for satire and parody. However, AB 730 is ambiguous on how to efficiently and effectively determine these criteria—ambiguity that nefarious actors are likely to game. By claiming satire and parody when the material is contested, a deepfake could be tied up in a lengthy review process for removal. Like the manipulated video of House Speaker Nancy Pelosi to make her appear intoxicated, a drawn-out review process to determine the video’s intent enables it to further gain virality and spur a contagion of negative effects.
The law exempts platforms from the responsibility to monitor and stem the spread of deepfakes. This is due to Section 230 of the Communications Decency Act, which provides platforms with a liability safeguard against being sued for harmful user-generated content, especially if they are acting in good faith to remove the content. Court interpretations since the law’s passing in 1996 have broadened platforms’ immunity, even if they deliberately encourage the posting of harmful user-generated content.
Instead, the law places responsibility on producers of deepfakes to self-identify manipulated content and on users to flag suspicious content. This is like having a Wall Street broker trading on inside information to self-identify intentions and suspicious transactions, or asking a con artist to make victims sign a terms of service agreement before they get swindled. These tactics will be unenforceable and ineffectual. Nefarious actors will not voluntarily disclose their creations as deepfakes. They’ll use botnets—connected communities of bots that interact with each other to quickly spread content through a social network—to evade detection. The damage to public perception will be done well before the content is flagged and reviewed for takedown.
According to the law, any registered voter may seek a temporary restraining order and an injunction prohibiting the spread of material in violation. It is not hard to imagine this being manipulated by special interest groups to tie up contentious content in a lengthy review process for removal, as well as inviting public skepticism regarding its veracity. By introducing doubt, the damage to public perception will be done without needing removal. This is especially problematic for content that truthfully depicts a candidate engaging in unsavory or illegal behavior, but is being claimed by supporters to be a malicious deepfake. When there is no definitive truth, everything is a lie.
Content spread through platforms has tangible effects on our democracy and public safety. To mitigate the spread and impact of malicious deepfakes, platforms must be required to play a more proactive role. Last month, senators Mark Warner, the Democrat of Virginia, and Marco Rubio, the Republican of Florida, sent identical letters to leading social media platform companies urging them to establish industry standards to deal with deepfakes. If the California legislature really wants to address the spread of malicious deepfakes, they must put pressure on platforms.
Burden of Proof
Again, the law only pertains to deepfakes posted with “actual malice” or “the knowledge that the image of a person has been superimposed on a picture or photograph to create a false representation,” and that there is an “intent to injure the candidate’s reputation or to deceive a voter into voting for or against the candidate.” Proving actual malice will not be straightforward. Clear and convincing evidence, which is often difficult to obtain, will be necessary to determine intent. Because of the high burden of proof to determine actual malice, a lengthy review process will likely ensue and allow the deepfake to continue to spread.