By Connie · Last reviewed: April 2026 — pricing & tools verified · This article contains affiliate links. We may earn a commission at no extra cost to you if you sign up through our links.
OpenAI Launched a Safety Fellowship Hours After the New Yorker Reported It Deleted 'Safely' From Its Mission
April 7, 2026 · 8 min read
On April 6, 2026, OpenAI announced a Safety Fellowship for external researchers — the same day The New Yorker reported OpenAI had removed "safely" from its IRS mission statement and dissolved its internal superalignment team. The fellowship runs Sept–Feb 2027, pays a stipend + compute, and is open to researchers from CS, social science, cybersecurity, and more. Applications close May 3, 2026.
The Timing Everyone Noticed
On April 6, 2026, OpenAI posted to its social accounts: "Introducing the OpenAI Safety Fellowship." Hours earlier, The New Yorker had published an investigation — based on IRS filings, internal documents, and sources — with a different kind of OpenAI safety story.
According to the investigation: OpenAI dissolved its superalignment team (the internal group tasked with solving alignment for superintelligent AI). It also dissolved its AGI-readiness team. And in its IRS 501(c)(3) filings, OpenAI had quietly removed the word "safely" from its stated mission.
The original mission, as OpenAI filed it: "to build artificial general intelligence safely and in a way that benefits all of humanity." The 2026 filing reads: "to build artificial general intelligence in a way that benefits all of humanity."
One word. Gone. The AI safety community noticed.
What the Safety Fellowship Actually Is
The OpenAI Safety Fellowship is a funded external research program. Selected fellows receive:
- A monthly stipend for the program duration
- Computing resources and API credits
- Mentorship from OpenAI safety researchers
- Optional access to a shared workspace in Berkeley, California (remote participation allowed)
The program runs from September 14, 2026 to February 5, 2027. Fellows are expected to produce a significant research output — a paper, benchmark, or dataset — by the end.
Who Can Apply
Eligibility is deliberately broad. OpenAI is accepting applications from:
- Computer scientists and ML researchers
- Social scientists studying AI's societal effects
- Cybersecurity and privacy researchers
- Human-computer interaction researchers
- Practitioners with strong track records in adjacent fields
OpenAI explicitly says it values research ability and technical judgment over specific academic credentials. Applications close May 3, 2026. Accepted candidates will be notified by July 25, 2026.
Priority Research Areas
| Research Area | Focus |
|---|---|
| Safety evaluation and ethics | Benchmarks, red-teaming, moral reasoning in AI systems |
| Robustness and scalable mitigation | Making safety techniques that scale with capability |
| Privacy-preserving safety methods | Safety without requiring sensitive training data |
| Agentic AI oversight | Controlling AI agents that take multi-step autonomous actions |
| High-severity misuse domains | Bioweapons, CBRN risks, large-scale manipulation |
The Debate: Outsourcing vs. Rebuilding
The AI safety research community's reaction split along a familiar line.
On one side: the fellowship is genuinely useful. External, independent researchers studying frontier AI safety is better than nothing. The stipend and compute access make it possible for academics and independent researchers who couldn't otherwise afford to work on these problems.
On the other side: a fellowship for external researchers is not a replacement for internal safety infrastructure. The superalignment team existed to work directly on aligning future, more capable models — with access to OpenAI's own systems, weights, and training pipelines. External fellows have none of that access.
The distinction matters because the hardest alignment problems are model-specific. You cannot solve them from the outside.
What This Means for the Broader AI Safety Landscape
OpenAI is not alone in restructuring safety teams. Google DeepMind, Anthropic, and Meta have each shuffled safety-adjacent teams as product velocity has accelerated. But OpenAI's restructuring drew particular attention because of the IRS filing change — a legal document, not a PR statement.
Anthropic, by contrast, has published research on AI's labor market effects and maintains a large internal safety team. Its mission — "the responsible development and maintenance of advanced AI for the long-term benefit of humanity" — remains unchanged in its filings.
The question now is whether the fellowship will produce research that meaningfully influences OpenAI's product decisions, or whether it serves primarily as external credibility for a company that is racing toward AGI faster than its internal safety structure can keep up.
Key Dates
| Date | Milestone |
|---|---|
| April 6, 2026 | Fellowship announced; New Yorker investigation published same day |
| May 3, 2026 | Application deadline |
| July 25, 2026 | Accepted candidates notified |
| September 14, 2026 | Program begins |
| February 5, 2027 | Program ends; research output due |
FAQ
Sources
Get the best AI tools tips — weekly
Honest reviews, tutorials, and Happycapy tips. No spam.