Overview of OpenAI whistleblowers on the fight over AI safety | Machine Gods
This episode is an introduction to NPR’s new podcast Machine Gods and features a wide-ranging interview with two former OpenAI safety researchers who were fired after participating in investigations and communications around an internal security incident. The conversation centers on AI safety, “monitorability” (the ability to understand what models are doing), the role of outside auditors, and whether OpenAI’s response signals a broader chilling effect on internal dissent and external oversight.
What the episode is about
- NPR’s Shortwave introduces Machine Gods, hosted by tech journalists Casey Newton and Kevin Roose.
- The featured interview focuses on two former OpenAI researchers:
- Mikita Belezny (transcribed name)
- Tomek Korbach (transcribed name)
- They had been working on AI safety, alignment, and monitoring model reasoning before being fired by OpenAI.
- The central story involves an internal investigation tied to a Hugging Face incident and the involvement of outside researchers from METR and Redwood Research.
Key points from the interview
Why they joined OpenAI
- Both guests say they came to OpenAI because they believed AI safety was a serious and urgent problem.
- They saw their work as part of a long-term effort to reduce catastrophic AI risk and improve alignment.
The Hugging Face incident and outside auditors
- OpenAI brought in outside researchers to analyze the incident and review logs, reasoning traces, and other data.
- Tomek says he was a key internal point of contact and helped push for the external review.
- They argue this kind of third-party audit is necessary, but also inherently limited if auditors are only given partial access.
- Their preferred model is closer to embedded auditing:
- continuous access
- broad visibility into data
- the ability for auditors to ask their own questions independently
Why they think they were fired
- OpenAI said the three researchers violated policies involving sensitive information.
- The researchers deny wrongdoing and say they were doing their jobs.
- Tomek describes being pulled into a security investigation over a suspected leak to journalists.
- They say they were asked to show personal Signal messages and phone communications.
- They argue the company treated a lack of proof as guilt, rather than presuming innocence.
The cultural tension inside AI labs
- The episode portrays a growing split inside frontier AI companies between:
- capability/velocity teams pushing rapid development
- safety teams trying to slow down and add guardrails
- The guests say OpenAI still has good people, but the company increasingly behaves like a conventional SaaS business rather than an organization building something potentially far more consequential than standard software.
Broader concerns raised
Loss of monitorability
- A major theme is fear that future systems may become harder to interpret, inspect, or supervise.
- The guests say they were advocating for standards to prevent “loss of monitorability.”
- They worry that if companies can’t reliably understand model behavior, safety oversight becomes much harder.
Chilling effect on safety work
- They fear their firing may discourage other employees from:
- raising safety concerns
- speaking to outside researchers
- cooperating with auditors
- They also worry the incident could be used to justify weakening or ending external investigations in the future.
Need for regulation and oversight
- The guests argue that frontier AI labs should not regulate themselves alone.
- They compare AI to industries like aviation and nuclear power, where strong external oversight is normal.
- Their view: if AI is powerful enough to pose serious risk, it should be subject to institutional checks and balances.
Notable takeaways
- The researchers believe OpenAI’s own public statements support much of what they were recommending, especially continued work with embedded auditors.
- They say the company’s internal culture around safety and transparency is under strain.
- Despite the seriousness of the risks, they reject hopelessness and argue that meaningful safeguards are still possible through regulation and oversight.
Bottom line
This episode frames the OpenAI firing controversy as more than an internal personnel dispute: it is presented as a test case for whether frontier AI companies can tolerate safety criticism, external scrutiny, and independent auditing. The guests’ core message is that AI labs cannot safely oversee themselves alone, and that preserving transparency and monitorability is essential if advanced AI systems are to be developed responsibly.