Machine Gods/Pod/26/10 Fired OpenAI Safety Researchers Speak Out + We Look Inside Pandora’s Box
Properties
- date-created
- 2026/10/24
- Podcast: Machine Gods/Pod
- Hosts: Person/Kevin Roose, Person/Casey Newton
- Summary: Two fired OpenAI safety researchers on a leak investigation and why AI audits need independent investigators, plus a look at how OpenAI runs.
Readwise highlights
- Tomek describes the Hugging Face incident as the most intense period of his career, when AI safety preparation suddenly became real.
- He realized earlier work had been a drill and decisions now carried immediate consequences.
- Short, contractor-style audits struggle because investigators cannot know every relevant question beforehand.
- Tomek Korbak recommends continuously embedded auditors with unrestricted access, like employees rather than temporary consultants.
- Mikita Balesni worked with Jasmine Wang on cross-industry standards for monitoring whether AI systems remain understandable.
- He shared internal memos with OpenAI leaders and external experts while saying he disclosed no confidential intellectual property.
- OpenAI investigated Mikita Balesni and Tomek Korbak over a suspected leak about less-monitorable AI architectures.
- They denied leaking information and surrendered Signal messages from their personal phones to investigators.
- OpenAI investigated whether the researchers leaked information about less-monitorable architectures to journalists.
- Investigators asked them to show personal Signal messages, yet they were fired without a chance to explain the context.
- The researchers considered Signal and disappearing messages standard security practices for sensitive work.
- Tomek Korbak said OpenAI effectively presumed guilt when employees could not prove their innocence.
- Require independent AI auditors to have broad access, confidential employee channels, and authority beyond voluntary contractor arrangements.
- Tomek Korbak points to regulated industries where auditors can inspect everything and speak privately with staff.
- OpenAI employees reportedly fear speaking to journalists and third parties after researchers were asked to show personal phones.
- Rumors suggest OpenAI may cut off Meter and discredit the independent safety organization.
- Frontier labs may lack the incentives to prioritize safety because private-company leadership must maximize revenue for stakeholders.
- That financial pressure can conflict directly with minimizing risks from increasingly capable systems.
- Mikita Balesni sees OpenAI operating more like a SaaS company than an organization building a potentially superhuman system.
- Fast product innovation and the rigor required for safe AGI development create a deep cultural tension.
- The researchers reject hopelessness and argue that enforceable checks and balances can substantially reduce AI risks.
- They compare AI governance with aviation and nuclear regulation, where oversight coexists with continued innovation.
- OpenAI's former researchers urged the company to maintain embedded audits, preserve monitorability, and restore its open internal culture.
- They argued employees must be able to raise safety concerns and collaborate with outside experts.