Former Anthropic researcher Jacob Coxon has become a prominent voice in the AI safety debate after publicly criticizing both Anthropic and OpenAI and warning about the trajectory of advanced AI development. Major media outlets have highlighted his resignation from Anthropic and his claims that leading AI labs are moving too quickly toward increasingly capable systems.
As the discussion around his comments grows, it is worth considering the experience supporting his observations as carefully as the observations themselves.
According to Zeki Data’s talent intelligence, Jacob Coxon has a notably strong mathematical background. Zeki Data records show he won bronze and silver medals at the International Mathematical Olympiad in 2016 and 2017, an achievement that places him among an exceptionally small group of elite young mathematicians.
Zeki Data also shows that before joining frontier AI labs, Coxon spent a brief period (12 months) as a machine learning engineer at InBalance Health, an AI company focused on hormonal health and headquartered in Beirut. His academic publication record prior to OpenAI aligns more closely with applied health and medical research. The one paper he published during his time at Cambridge University used bayesian analysis to look at survival rates from tuberculosis meningitis.
There is little question that Coxon has operated close to the frontier of AI development. Public reporting shows he worked at both OpenAI and Anthropic and was involved in model pre-training research during a period of rapid advancement in foundation models.
The more interesting question is whether proximity to frontier AI development should automatically be interpreted as deep expertise in AI safety.
Zeki Data’s analysis suggests caution here. While Coxon has technical credentials and experience at leading AI labs, his publicly visible research record in AI safety is relatively limited. Zeki identifies a single mechanistic interpretability paper associated with his time at OpenAI, co-authored with a small group of researchers. The work was released publicly but was not published in a major journal or conference proceeding. Public records also show his name among hundreds of contributors acknowledged in the GPT‑4o system card, a distinction shared by many OpenAI employees.
None of this invalidates his concerns. It does, however, highlight an important distinction that is frequently lost in viral discussions: being close to a technology and being one of the field’s foremost experts on a specific topic are not necessarily the same thing.
“On a subject as existential as Ai’s threat to humanity, we need the real experts inside the labs to step forward and be heard on the record not unreliable witnesses including the CEOs,” observed Tom Hurd, Co-Founder and CEO of Zeki Data.
Coxon’s comments may resonate precisely because they come from someone who helped build advanced AI systems. Yet it is equally important to understand the nature of the expertise behind those comments. His perspective appears best understood as that of a talented mathematician and AI practitioner with direct experience inside frontier labs, rather than as a researcher whose career has been primarily devoted to foundational AI safety research.
Image Credit: The Financial Times