OpenAI Researchers Challenge Their Dismissals in Public Letter

Three safety researchers recently dismissed by OpenAI have publicly challenged the narrative surrounding their terminations. The open letter warns of the potential chilling effect on employees willing to voice their concerns about the risks associated with AI.

Tomek Korbak, Jasmine Wang, and Mikita Balesni addressed their letter to OpenAI’s Safety and Security Committee, Safety Advisory Group, and Mission Advisory Council. They expressed worry that communications regarding their firing could deter colleagues from speaking openly. They also emphasized the importance of researchers challenging decisions and collaborating with external experts, stating, “AI is a distinct technology, and OpenAI is a unique company.”

OpenAI refuted claims linking the researchers’ dismissals to their vocal safety concerns. The company asserted that the trio violated policies concerning sensitive information, following an exhaustive investigation. Newsweek has reached out to OpenAI and the dismissed researchers for further comments and awaits responses.

The Open Letter

The letter titled “OpenAI cannot make AI safe on its own” was made public by Korbak, Wang, and Balesni. It cautions about the negative impact their dismissals could have on others within OpenAI. The former employees voiced concerns that communications around their terminations have generated fear among their peers.

Historically, OpenAI encouraged open discourse on safety concerns, they claimed, but staff now feel unsure about their standing.

Korbak, Wang, and Balesni argue that current safety concerns necessitate working environments free from fear and vague policies.

Disputes Over Dismissals

The researchers challenged claims regarding the reasons behind their dismissals, particularly the allegation they leaked information on AI architectures. “We were not responsible for the leak,” they wrote.

They denied having improperly communicated with external organizations, asserting any interactions were within their job requirements.

Korbak noted he was verbally informed his termination related to communications with an independent organization, METR. “No detailed explanation or supporting documentation,” Korbak commented on X. His engagements with METR were part of his job duties.

The open letter discusses Korbak’s concerns about decreasing abilities to monitor AI reasoning, believing his dismissal related to these raised issues.

Mikita Balesni’s Insights

Balesni publicly disputed OpenAI’s reasoning, suggesting their prioritization of safety clashed with OpenAI’s corporate interests. He stated the accusations of leaking company intellectual property were baseless. His work ensured AI monitorability, collaborating closely with external parties.

“Throughout, Mikita regularly checked with authorities and sanitized shared materials,” the trio noted.

Jasmine Wang’s Account

Wang challenged OpenAI’s explanation about her accessing an executive’s email, which was sanctioned for recruiting purposes. “I requested IT remove access once unnecessary,” Wang stated on X. Her repeated requests went unheeded.

Wang maintained that transparency was maintained when inadvertently accessing sensitive emails, urging OpenAI to provide a detailed list of allegations to address.

Recommendations for OpenAI

The former researchers recommended that OpenAI and similar AI companies continue independent safety assessments. They warned against reducing interactions with external organizations capable of scrutinizing powerful systems.

They advocated for preserving AI monitorability, emphasizing industry-wide commitment. The researchers urged maintaining an open culture, encouraging communication with external experts.

OpenAI’s Response

On October 9, OpenAI defended its decision, affirming that the researchers violated sensitive information policies. The company detailed breaches of trust but declined to specify conduct due to privacy.

OpenAI disputed claims about dismissals being linked to safety advocacy, underscoring the importance of spirited debates and open discussions. The company expressed commitment to collaboration with third-party safety assessors.

OpenAI agreed with the significance of industry cooperation in monitoring advanced systems, asserting that trust is crucial for fostering an open culture. The company acknowledged the researchers’ past contributions while steadfastly denying any connection between safety advocacy and their dismissal.

Broader Implications

This disagreement highlights challenges faced by companies developing frontier AI. Balancing confidentiality and effective partnerships with external evaluators remains complex. Fear of repercussions may deter researchers from engaging with outside evaluators.

The dismissed researchers caution that they form a critical frontline defense against risks inherent in increasingly advanced AI systems.