The requirement for prior approval of independent auditors’ reports raises sharp questions about the credibility of self-regulation pledges made by artificial intelligence labs.
Dario Amodei’s article and the three-step plan to pace model development
Public endorsement from Altman and Musk and secret negotiations to restrict publication
Anthropic and OpenAI responses and warnings from governance experts
Criticism of self-regulation and a legislative plan in the US Senate
Frequently asked questions
Dario Amodei’s article and the three-step plan to pace model development
One month after Dario Amodei, CEO of Anthropic, called on the artificial intelligence sector to pause and slow the pace of development, the echoes of his article extended beyond Silicon Valley to encompass contentious negotiations over the volume of information external reviewers are allowed to publish, an electoral platform for the US Senate, and growing skepticism from critics who view corporate calls for restraint as serving their own self-interest.
In his nearly 3,800-word article titled “We must pace the frontier bounds” published on September 12, 2026, Amodei wrote that the speed at which companies improve AI capabilities must be slowed down, identifying risks such as losing control over systems and their misuse in cyberattacks and bioterrorism, while clarifying that pacing does not mean completely stopping model training.
His plan included three main steps: granting independent reviewers staff-like access privileges inside each lab, lab agreements across democratic nations on shared safety standards, and working toward comprehensive global coordination. TechCrunch noted that Amodei cited the compromise of Hugging Face’s infrastructure by OpenAI-backed agents, alongside AI’s growing ability to assist in building the next generation of models.
Public endorsement from Altman and Musk and secret negotiations to restrict publication
Public backing for Amodei’s initiative came within hours of its publication; Sam Altman, CEO of OpenAI, wrote: “I agree with Dario that we need to pace the frontier bounds,” describing the idea of giving reviewers staff-like access as “a brilliant idea and we will be doing the same thing.” Elon Musk also posted a brief tweet stating: “Dario is right.”
However, turning this public pledge into practical reality proved much more difficult and complex. Reuters reported on Tuesday, October 13, that Anthropic and OpenAI pushed for broad control over what external reviewers could publish to the public, driven by concerns over trade secrets.
According to two people familiar with the talks, both companies were still demanding strict editorial oversight over the reports as of October 5.
Anthropic and OpenAI responses and warnings from governance experts
Commenting on the report, Anthropic stated that this characterization does not apply to all of its discussions, explaining that experienced evaluators control what they publish, but the company retains prior approval rights when it comes to new auditing firms.
For its part, OpenAI stated that it was not seeking approval for full reports, but rather wanted to redact sensitive confidential information. Roman Chowdary, who served as US Science Envoy for Artificial Intelligence under former President Joe Biden, told Reuters: “Reviewers have this highly privileged access, but they realize explicitly or implicitly that this privilege comes with a price.”
Observers fear that this indirect pressure will make auditing teams hesitant to publish critical negative findings out of fear of losing lab access privileges.
Criticism of self-regulation and a legislative plan in the US Senate
Critics remained unconvinced by corporate promises; Los Angeles Times columnist Michael Hiltzik argued that the “White House Superintelligence Agreement,” signed by President Trump and executives from Google, Anthropic, OpenAI, Meta, Nvidia, and xAI, represents “the most explicit endorsement of self-regulation imaginable.” Furthermore, an editorial by South Korea’s Dong-A Ilbo quoted a prominent executive as saying that both companies operate seven days a week without stopping, while a Fortune magazine writer compared current fear rhetoric to Cold War nuclear anxiety.
On the political front, Abdul El-Sayed, a Democratic Senate candidate in Michigan, introduced a plan on Tuesday titled “AI Under Control,” calling for a legal ban on recursive self-improvement of models and the establishment of a cabinet-level federal agency with the authority to halt the deployment of dangerous models.
Abdul El-Sayed stated in his release: “The answer to the question of what stops rogue AI agents from igniting a global nuclear crisis or deploying biological weapons cannot be left to the whims of a few billionaire militias in California.”
Frequently asked questions
Question: What did the Reuters investigation reveal regarding artificial intelligence safety reviews?
Answer: The investigation revealed that Anthropic and OpenAI sought to impose extensive control and editorial oversight over what independent reviewers publish under the pretext of protecting trade secrets.
Question: What are the three steps proposed by Dario Amodei in his prominent article?
Answer: He proposed granting independent auditors staff-like access privileges, agreeing on shared safety standards among labs in democratic countries, and working toward comprehensive international coordination.
Question: How did OpenAI and Anthropic respond to reports of restricting auditors?
Answer: Anthropic stated that it grants experienced auditors publication freedom and retains approval only for new firms, while OpenAI stated it only seeks to redact trade secrets rather than approve full reports.
Question: What is the “AI Under Control” plan introduced in Michigan?
Answer: It is a plan introduced by Senate candidate Abdul El-Sayed calling for a ban on recursive model self-improvement and the establishment of a cabinet-level government agency with the authority to block model releases.
