OpenAI launches a fellowship program for independent AI safety researchers

OpenAI has announced the Safety Fellowship, a pilot program supporting independent researchers who work on AI safety and alignment. The program aims to develop the next generation of specialists who will tackle some of the most important technical challenges of the era.

It targets independent scientists and academic researchers working on how to make modern language models more predictable, interpretable and safe. Fellows will receive funding, access to OpenAI's computing resources and the opportunity to work directly with the company's internal Safety team.

The grant amounts and eligibility terms have not yet been disclosed. The pilot status implies that details will be refined as the first results come in.

AI alignment remains one of the central topics in debates about the technology's long-term risks. As language models grow more powerful, the question of how to ensure they behave in line with human intentions and values becomes ever more pressing. OpenAI, as one of the leading developers of such systems, is in an awkward position: it creates potential risks and looks for ways to reduce them at the same time. The Safety Fellowship is an attempt to address part of this problem by backing outside researchers who are not directly tied to the company's commercial agenda.

The Safety Fellowship is not the first initiative of its kind in the industry. Anthropic, founded by former OpenAI employees, positioned itself from the start as a lab focused on safety. Google DeepMind and academic groups at MIT, Stanford and Berkeley are actively researching interpretability, robustness and RLHF. Still, none of the existing programs comes directly from the lab that built ChatGPT and sets the pace for the entire industry. Against this backdrop, OpenAI's new program looks like an attempt to move beyond in-house R&D and build a broader independent community.

The word "independent" in the program's description deserves a closer look. Historically, funding from large tech labs has raised conflict-of-interest questions: it is hard to assess a technology's risks objectively when your livelihood depends directly on its commercial success. That is why separate centers emerged in academia, such as MIRI, ARC and CHAI, funded independently of the industry giants. How OpenAI will ensure the fellows' real intellectual independence is the key question. The answer will determine whether the program becomes a genuine contribution to science or a PR tool for positioning.

The pilot nature of the initiative suggests the company is still feeling out the right format: who should take part, which research to fund, how to publish results. The first cohorts of fellows will most likely set the trajectory, as happened with similar programs from Open Philanthropy or Machine Learning Collective.

The mere appearance of such a program from the world's most visible lab is a signal to the whole field. It acknowledges that AI safety problems are too complex for one company to solve alone. For researchers working on alignment with minimal funding, it is a concrete opportunity, regardless of what ultimately motivated OpenAI to launch the program.