(Credit: Chesnot / Contributor / Getty Images Entertainment via Getty Images)
Anthropic supports an amended version of California AI bill SB 1047, but says it still lacks clarity in some areas and leaves room for improvement.
"The new SB 1047 is substantially improved, to the point where we believe its benefits likely outweigh its costs," CEO Dario Amodei writes in a letter to California Governor Gavin Newsom. "However, we are not certain of this, and there are still some aspects of the bill which seem concerning or ambiguous to us."
SB 1047 has already cleared a 32-1 vote in the state Senate and an initial vote in the Assembly. It now goes to a final vote by the full Assembly, which must occur by Aug. 31. If it passes, it would advance to Newsom to sign or veto by Sept. 30, according to Reuters.
The bill faces opposition from those who say it could stifle innovation, such as Google and Meta. Last week, some members of Congress, including Nancy Pelosi, wrote a letter to Newsom calling the legislation "well-intentioned but ill-informed."
Anthropic had similar concerns about hampering innovation after reading the first draft, but those fears have been "greatly reduced" in the new version. With the new amendments, the bill "likely presents a feasible compliance burden for companies like ours," Amodei says, citing the "importance of averting catastrophic misuse," which he believes could begin to take shape in the next one to three years.
SB 1047 could help protect the public by requiring AI companies to adopt and fully disclose safety and security protocols (SSPs).
"Nothing prevents companies from making misleading statements about their SSPs or about the results of the tests they have conducted as part of their SSPs," Amodei says. "It is a major improvement, with very little downside, that SB 1047 requires companies to adopt some SSP (whose details are up to them) and to be honest with the public."
The bill could also incentivize companies to mitigate abuse of their systems, and push forward research efforts to greater understand those risks.
As for the cons, Amodei says the proposal disproportionately affects large AI companies and largely exempts smaller players. It's also still too restrictive in some areas. For example, it gives the Attorney General license to enforce the bill before any harm has occurred, which could lead to overreach.