Copyright Office Declines to Bless Hacking to Study AI Bias (1)

The US Copyright Office is rejecting an exemption to a statutory hacking ban allowing researchers to circumvent digital security measures to probe artificial intelligence models for bias and discrimination.

Register of Copyrights Shira Perlmutter recommended denying the exemption because the issues raised by its supporters are linked to third-party control of online platforms, not federal copyright law, so “an exemption would not ameliorate their concerns,” according a final rule set to publish Oct. 28 in the Federal Register. The Library of Congress accepted the recommendation and didn’t include the exemption in the new rule.

Perlmutter acknowledged the importance of AI trustworthiness research as a policy matter, but noted that Congress and other agencies may be best positioned to act on this emerging issue, according to the rule.

The 1998 Digital Millennium Copyright Act prohibits circumventing technological measures to protect copyrights—including passwords, encryption, and digital locks. Under the DMCA, the Copyright Office implements exemptions to that general ban and reviews existing and proposed exemptions every three years.

The aim of the Section 1201 proceeding—so-named for the DMCA provision that established the process—is to assess whether technological protection measures controlling access to copyrighted works adversely affectindividual users’ ability to make lawful uses of copyrighted works. In October 2023, the Copyright Office advanced seven proposed or expanded exemptions, including one for researchers to access “copyrighted generative AI models, solely for the purpose of researching biases” within them.

Three policy groups submitted comments in support of the AI research exemption, arguing that permitting independent hackers to legally circumvent digital security measures to probe AI models for bias would boost transparency and trust in the polarizing technology. The Justice Department’sintellectual property crime unit also endorsed the measure, calling independent research “essential to ensuring the integrity and safety of AI systems.”

Technology, streaming services, and news industry groups opposed the exemption, citing concerns about piracy of copyrighted works and disincentivizing innovation.

Reasoning, Potential Effects

Proponents of the exemption failed to prove that the DMCA’s prohibition on circumvention of TPMs has or will adversely affect their ability to conduct such research, according to Perlmutter’s letter.

The Copyright Office found the adverse effects identified by proponents weren’t actually due to TPMs, but rules set forth in corporate terms of service or safety guidelines, Suzanne Wilson, the Copyright Office’s general counsel, said at a press briefing on Friday. Effects of those measures are “simply not within the scope of 1201,” she said, but may be a place where “other agencies or Congress may want to act.”

The office’s reasoning is fair, said Saul Ewing LLP partner Darius Gambino. “The Library of Congress took a look at this and all the different opinions and said, ‘This is more of a issue for the Congress to deal with,’” he added.

The National Telecommunications and Information Administration, part of the Department of Commerce, said in a letter the exemption should be modeled after the existing shield for security research. There’s sufficient evidence that researchers are hesitant to continue their work without an exemption, according to the agency.

Gambino acknowledged the comparison to security research, but said the software industry was well developed when that exemption took effect. The AI research proposal, on the other hand, deals with “an entirely different class of software that is still fairly new, that’s still being worked out, that the other branches of government haven’t spoken to yet,” he said.

It’s encouraging to see the NTIA support the exemption and acknowledge the chilling effect on research, said Amit Elazari, CEO and co-founder of regulatory engagement platform OpenPolicy, which submitted comments in support of the exemption. But the register’s rejection reasoning seems like a technicality and “legalese argument,” she said.

“We are not at a point where we have three years to wait for the next process here,” Elazari said. There is an urgent need to cultivate an ecosystem of AI red teaming—structured testing to find flaws and vulnerabilities.

She said “it’s a missed opportunity,” but indicated proponents may ask Congress to consider the register’s recommendations for AI research safe harbor.

Exemptions in the new rule include text and data mining of audiovisual and literary works, computer programs for repair of commercial industrial equipment, and computer programs for vehicle operational data. They’ll remain in force until October 2027.