Google Has Some Thoughts On AI

Google is out this week with a set of proposals for AI governance that includes a number of recommendations that, wait for it…would be good for Google. Titled “A Pragmatic Approach to AI Governance in America,” the 21-page outline touches on six areas of concern involving so-called “frontier” AI models: the future of work; protection of minors; information integrity; copyright and creativity; energy infrastructure; and privacy. The pragmatism is in the eye of the beholder.

The proposal takes its biggest swing at the question of who should regulate frontier AI models. Google’s answer: an independent self-regulatory body under the supervision of an unspecified federal agency. It calls for creating a federally supervised frontier AI regulatory organization (FARO) that would “standardize and verify frontier AI labs’ safety and security practices prior to launching frontier models on the U.S. market, including by ensuring that companies are publishing and complying with frontier AI frameworks and engaging with annual governance process audits.”

The FARO board, according to the proposal, would include “a combination of independent directors and industry representatives, balancing considerations of safety and security against the need for bold innovation and speed to market in a global race.”

Google likens the structure to the Financial Industry Regulatory Authority (FINRA), a self-regulatory body comprised of financial broker/dealer members overseen by the Securities and Exchange Commission (SEC), or state bar associations that are overseen by state supreme courts.

Unmentioned by Google is that FINRA, at least, was established by act of Congress and given explicit authority to enforce SEC regulations over its members. The FARO, as described in the proposal, would be more of a standards-setting organization than an enforcement agency.

Working in concert with federal government experts in different safety domains, the FARO should maintain a repository of standards and benchmarks for building, testing, and deploying the most advanced AI systems.

It could also establish other safety rules based on industry best practices, guidelines from industry bodies like the Frontier Model Forum, and standards from the American National Standards Institute and the International Standards Organization. Ideally, governments around the world would implement international reciprocity agreements with the U.S. government, such that models verified by the FARO would be automatically verified in reciprocal jurisdictions.

Google’s other big idea is to explore “new types of value-exchange” with publishers—so long as it doesn’t involve mandatory licensing of publishers’ content for training AI models.

Using publicly available web data for training models is a transformative, non-expressive use — like an art student taking inspiration from walking through a gallery — that should remain protected under fair use in the U.S. and text-and-data-mining exceptions abroad. While this use is legally protected, responsible model developers should give website owners choice and control over whether their content is used in model development via simple, machine-readable robots.txt tags, like the Google-Extended control. Such controls enable markets to work and allocate value where appropriate.

“Where appropriate” is doing a lot of work in that paragraph. It also elides the fact that robots.txt is a voluntary protocol with no technical enforcement mechanism to back up “website owners choice and control” over the use of their content. It’s up to the bots whether to honor robots.txt tags.

As for any further consideration of copyright, “the focus should again be on outputs — in this case, whether a specific image or piece of text actually copies an existing work, regardless of how it was created,” the proposal says.

It’s hard to know who the intended audience is for the proposal. The copyright portion simply restates a position Google and other tech companies have long held and advocated for, both publicly and in court.

The FARO proposal seems intended to imply a concession on Google’s part to the need for regulation of AI. But substantively, it seems more like a bid to preempt state AI regulations without actually creating formal federal regulations beyond industry-developed standards and “best practices.”

To be fair, there is something to be said for involving outside expertise in crafting any federal standards regarding highly technical and fast-moving developments such as AI. That’s especially so now, after Donald Trump and Elon Musk have effectively denuded the federal government of expertise of any kind in favor of political hackery or simply no governance at all.

Beyond that, however, it seems there’s less here than meets the eye.

Share