Former Anthropic Researcher Warns AI Companies Don't Fully Control Their Models

A former researcher at the AI company Anthropic has warned that artificial intelligence developers do not fully control the models they create, according to a report by Al Jazeera.

Jacob Coxon, identified as a former Anthropic researcher, said in comments cited by the outlet that AI companies "don't fully control" their models, a concern that echoes broader debate within the AI research community about the limits of interpretability and oversight for advanced AI systems.

The warning adds to a growing body of commentary from researchers and former industry insiders questioning whether safety and alignment techniques are keeping pace with the capabilities of the most advanced AI systems. Such concerns typically center on the difficulty of fully predicting or explaining the behavior of large, complex models, even for the organizations that built them.

Anthropic, along with other major AI developers, has publicly acknowledged the challenge of interpretability and has invested in research aimed at better understanding how its models arrive at particular outputs. The company has also previously stated that safety is a central part of its mission as it develops increasingly capable systems.

The report did not detail specific technical claims or evidence underlying Coxon's warning beyond the general statement about control, and Al Jazeera's report was presented as part of its video news coverage.