Amid months of security debacles and plummeting approval of its AI technology, OpenAI announced on Monday that four independent assessors will investigate its internal models and test the effectiveness of its safety features.
asserted
assessors → plummet → features
It may sound like a step in the right direction, but there’s a glaring problem: The evaluators have limited access to the startup’s technology, and—in many cases—OpenAI gets to approve each member of the research teams and can choose which staff they speak to.
uncertain
they → sound → staff
Many AI researchers have pointed out since yesterday’s announcement that the assessors don’t appear to have the “employee-like access” that CEO Sam Altman committed to in September.
asserted
Altman → point → September
The four assessors—METR, Apollo Research, Transluce, and Goodfire—are research labs that evaluate AI models for safety risks.
asserted
that → evaluate → risks
Two METR staff members, along with one researcher from another nonprofit AI safety research organization, investigated OpenAI in the wake of the July Hugging Face security incident.
asserted
members → investigate → incident
The researchers initially had only two days to work on-site and were later invited two more times—which amounted to a total of six days.
asserted
which → have → days
Following the publication of the team’s findings, one of the researchers admitted a lack of “good approaches for understanding/overseeing the activity and aims of AI” when multiple agents work together, as in the Hugging Face breach.
asserted
agents → follow → breach
“Outputs from analysis [AI] agents were often missing key details, wrong, overconfident, or really hard to understand,” the researcher, Ryan Greenblatt, wrote on X in August.
asserted
researcher → miss → August
The researchers’ report also acknowledged that the volume of material under review meant they “had to heavily delegate our analysis to AI agents,” which have “significantly worse judgment and reliability than human researchers.”
asserted
which → acknowledge → researchers
It’s not exactly a complete study.
asserted
It → ’ → ?
OpenAI said Monday it would allow METR to conduct its assessments with undefined “agreed access, resources, and safety arrangements.”
asserted
METR → say → access
In a footnote, the startup acknowledged that the scope of the investigation would be “less expansive than what METR outlines in their alignment assessment proposal,” which in October attempted to outline a comprehensive method to understand the worst plausible outcomes of a developer’s AI agents.
asserted
which → acknowledge → agents
Although Apollo Research, another assessor, is set to “work within OpenAI for four months,” its announcement says that OpenAI can withhold information because of “technical infeasibility, security risks, or IP protection” and will select the staff members that assessors can speak with.
asserted
assessors → set → that
Similarly, Transluce can only access models, safety documentation, and researchers pre-approved by OpenAI.
asserted
Transluce → access → OpenAI
The details of Goodfire’s assessment are the most vague, with the AI research lab stating that it will have “deep access to OpenAI’s internal models and systems”—whatever “deep access” means—to use its “interpretability platform to monitor training runs, debug undesired behavior, and improve alignment.”
asserted
access → state → alignment
Beyond the obscure language, OpenAI and the research labs’ framing of the assessments as “collaborations” that can work toward “stronger shared standards—emphasis on “shared”—is perhaps the most telling
asserted
that → work → standards
A truly independent investigation of OpenAI would not include OpenAI having final say on much of what is being investigated, or being able to dictate different terms to each assessor, with deeply different levels of time and access.
asserted
what → include → time
As METR itself said in October regarding how to evaluate unreleased AI models, independent experiments would “ideally have codebase and infrastructure access comparable to what an employee studying similar questions would typically have,” including training code and the company’s simulation tools.
asserted
employee → say → code
While the announcement suggests that OpenAI no longer feels it can just selectively disclose its own models’ “unexpected or concerning behavior,” it’s a far cry from the firm’s stated goal of “public transparency across the industry.”
Disclosure:
uncertain
it → suggest → industry
The Center for Investigative Reporting, the parent company of Mother Jones, has sued OpenAI for copyright violations.
asserted
Center → sue → violations
OpenAI denies the allegations.
asserted
OpenAI → deny → allegations