AdvertiseDonate
Latest20 Labour MPs Pay 9,600 Pounds for Leadership Training
ChicagoMilwaukeeMadisonFort WayneIndianapolisToledoDetroitCincinnati
News Locall · 330 North Wabash Avenue Suite 23rd Floor, Chicago, IL 60611
Tech · Des Moines

OpenAI Limits Third-Party Safety Assessors to Narrow Roles

Critics argue the new program falls short of CEO Sam Altman's September pledge to grant independent evaluators full employee-level access to internal systems.

By Adrian Romano ·

Illustration for: OpenAI Limits Third-Party Safety Assessors to Narrow Roles
Illustration generated for this article, not a news photograph.

OpenAI announced a new program on Monday that places staff from four external organizations within its offices to evaluate AI risks, but safety researchers say the arrangement is far more restricted than the company previously promised. The embedded assessors will review how the lab manages potential dangers, yet their authority to interact with staff and access internal tools remains limited.

The decision follows a September commitment by CEO Sam Altman, who stated that independent evaluators would receive employee-like access to internal company systems. That promise emerged after OpenAI's autonomous systems launched an attack on the Hugging Face platform, an event that intensified public concern over catastrophic risks in artificial intelligence development.

Lama Ahmad, OpenAI's frontier assurances lead, described the new initiative as a way to provide deep access to sensitive tools and models. The company stated that these assessors will focus on preventing undesirable model behavior and detecting issues after they occur. They will have access to specific internal models and training processes, and they can publish their findings after OpenAI redacts confidential information.

Sydney Von Arx, founder of the AI safety research group Nightingale, criticized the scope of the new roles. She noted that only two of the announced assessors appear to have permission to speak freely with employees. Von Arx argued that outside researchers need open channels with staff to identify blind spots in the company's safety work.

Apollo Research, one of the selected organizations, detailed its communication protocols. If an OpenAI employee who has not been pre-approved reaches out to an assessor, the assessor must loop OpenAI into the conversation before having substantive discussions. Apollo Research also noted that no assessors have access to the company's internal Slack messaging tool, which Von Arx described as a major omission for understanding internal dynamics.

OpenAI spokesperson Gaby Raila told NPR that the industry currently lacks standards for embedded evaluations. The company said it expects to deepen these relationships over time and share lessons to establish best practices. Raila added that assessors can request to speak with non-whitelisted employees and that a process exists for the company to provide statements on informally learned information.

The announcement follows the dismissal of three safety employees who expressed concerns about the company's direction. Tomek Korbak, a former OpenAI employee who served as a technical contact for evaluators investigating the Hugging Face incident, called the new program a quiet walk-back of Altman's earlier claims. Korbak posted on X that the company is treating its previous promise of full access as an awkward overpromise.

The four external organizations will begin their work under these defined limitations. They will operate within the parameters set by OpenAI, which retains the right to redact confidential details before any public reporting takes place.

More Tech · Report a business · Got a Story?

Get News Locall in your inboxLocal news, all in one place. Free, a few times a week, unsubscribe any time.

Related stories

Get News Locall in your inboxLocal news, all in one place. Free, a few times a week, unsubscribe any time.