OpenAI Launches Safety Fellowship Amid Damning Investigation and Dissolved Teams

The most striking feature of the OpenAI Safety Fellowship announcement is not its content but its calendar. On April 6, 2026, OpenAI published details of its new fellowship program supporting independent AI safety research. Within hours, Ronan Farrow and Andrew Marantz's 18-month investigation into OpenAI landed in The New Yorker, drawing on over 200 pages of internal documents and interviews with more than 100 people. The investigation raised serious questions about Sam Altman's trustworthiness and OpenAI's actual commitment to safety.

Farrow himself drew the connection publicly on X, writing that the fellowship announcement "arrives hours after our investigation described how OpenAI dissolved its superalignment and AGI-readiness teams and dropped safety from the list of its most significant activities on its IRS filings." His post received approximately 284,000 views and 3,600 likes, establishing the dominant public narrative around the fellowship as one of strategic timing rather than genuine institutional commitment. The crypto news outlet BSCN amplified the framing with its headline: "OPENAI ANNOUNCES SAFETY FELLOWSHIP HOURS AFTER ARTICLE INVESTIGATING CEO'S INTEGRITY." Whether the timing was coincidental or deliberate, the fellowship's launch is now inseparable from the investigation's revelations.

The fellowship announcement must be understood against a documented institutional pattern. Since mid-2024, OpenAI has dissolved three dedicated internal safety teams in rapid succession. The Superalignment team, created with a public pledge of 20% of OpenAI's compute resources, was disbanded in May 2024 after reportedly receiving only 1-2% of compute on the company's oldest hardware. Its co-leaders — Ilya Sutskever and Jan Leike — both departed, with Leike publicly stating that safety had taken a backseat to product development.

Five months later, in October 2024, the AGI Readiness team was dissolved when its head advisor Miles Brundage resigned. Then in February 2026, just two months before the fellowship announcement, OpenAI shut down its Mission Alignment team after only 16 months of operation. Each dissolution was accompanied by the departure of senior safety-focused personnel. This sequence raises a fundamental question: if OpenAI could not sustain internal safety teams with direct access to its models and infrastructure, what does an external fellowship with a five-month duration and limited compute access realistically accomplish? Critics argue the pattern suggests that safety work is valued primarily as a public relations asset rather than as a core organizational function.

The fellowship's structure reveals an interesting strategic choice. Rather than rebuilding internal safety capacity — which OpenAI has repeatedly built and dismantled — the company is now outsourcing safety research to independent scholars through a time-limited program. Fellows receive mentorship, compute resources, and a monthly stipend over approximately five months, and are expected to produce research papers or datasets. The program explicitly targets "the next generation of talent" and prioritizes proposals on safety evaluation, ethics, robustness, scalable mitigation, privacy-preserving methods, agentic oversight, and high-severity misuse.

This externalization model has both legitimate benefits and structural limitations. On the positive side, independent researchers may produce work that is less constrained by corporate incentives and more credible to regulators and the public. The research areas listed align with genuine frontier safety challenges. On the other hand, external fellows lack the deep model access that internal teams had, their five-month engagement is far shorter than the multi-year timelines safety research typically requires, and the program's outputs — papers and datasets — may have limited impact on OpenAI's actual deployment decisions. The EU AI Act's requirements for safety assessments of high-risk AI systems may also be a motivating factor, as fellowship-generated research could help demonstrate regulatory compliance without requiring permanent internal investment.

The fellowship does not exist in a competitive vacuum. Anthropic, founded by former OpenAI safety lead Dario Amodei, is simultaneously running its own Anthropic Fellows Program with applications open for May and July 2026 cohorts. This parallel creation of external safety fellowship programs by the two leading frontier AI companies signals that AI safety talent has become a strategic resource worth competing for.

The competition carries an ironic undertone. Amodei left OpenAI in part because of concerns about its safety culture — concerns he documented in over 200 pages of internal notes — and has characterized the fundamental problem at OpenAI as being "Sam himself." Now both companies are vying for the same pool of independent safety researchers. For prospective fellows, this competition is arguably beneficial: it creates more funded positions, more mentorship opportunities, and more institutional support for safety work. But it also raises the question of whether these programs will produce research that meaningfully influences how either company develops and deploys its models, or whether they primarily serve as talent scouting and reputation-building exercises in an increasingly scrutinized industry.

The New Yorker investigation provides crucial context that the fellowship announcement alone cannot. According to reporting, Altman internally expressed that "vibes don't match a lot of the traditional AI-safety stuff," suggesting a philosophical skepticism toward the very field the fellowship aims to support. Dario Amodei's 200-plus pages of internal notes reportedly paint a picture of an organization where safety concerns were systematically deprioritized.

Gary Marcus framed the stakes bluntly, asking whether anyone would want Altman "deciding, unilaterally, whether or not it is ok to release" a model capable of enabling bioweapons or cyberattacks. The investigation also revealed that OpenAI dropped safety from its list of most significant activities on its IRS filings — a bureaucratic detail that may carry more weight than any public announcement. These revelations suggest a gap between OpenAI's public safety narrative and its internal priorities. The fellowship, regardless of its operational merits, now exists within this credibility gap. Its ultimate significance will depend less on the quality of research it produces and more on whether OpenAI demonstrates sustained institutional commitment to safety beyond time-limited external programs.