OpenAI's New Reasoning Method Alarms Safety Experts — and Signals a Hiring Shift

OpenAI has built a new reasoning method into its Astra model that is drawing pushback from AI safety researchers, who say it could make it harder to monitor how the system actually reaches its conclusions.

The technique, sometimes called "recurrent depth" or "opaque recurrence," has the model process queries in loops rather than working through a step-by-step chain of thought, TechCrunch reported September 2. That loop-based approach produces fewer legible traces of the model's reasoning than the sequential, readable chain-of-thought output that has become a standard tool for catching misalignment or unsafe behavior before it reaches users. A widely cited 2025 position paper signed by researchers across several labs argued that chain-of-thought monitorability is "a new and fragile opportunity for AI safety," and urged developers to weigh how their design choices affect it — the exact tradeoff now in dispute.

OpenAI's use of the technique in Astra is reportedly limited so far. But according to reporting cited by TechCrunch from The Information, Anthropic and Google DeepMind are already discussing similar approaches, suggesting the shift could spread across major labs rather than stay confined to one.

The concern from safety researchers centers on oversight. Buck Shlegeris, CEO of Redwood Research, warned that "if OpenAI pushes this technique further, they'll have the option to massively increase the recurrence and totally destroys CoT monitorability," referring to chain-of-thought monitoring. Ryan Greenblatt, Redwood's chief scientist, raised a related alarm about scaling the technique further, warning the model could end up reasoning "entirely or almost entirely in latent space" — meaning outside any format humans can read or audit. AI safety advocate Zvi Mowshowitz and OpenAI chief scientist Jakub Pachocki were also named in the reporting.

What this means for job seekers

This is exactly the kind of development that keeps interpretability and AI safety roles in demand even as broader tech hiring stays choppy. When a leading lab's own architecture choices become a point of public alarm among researchers at organizations like Redwood Research, it reinforces that labs need people whose job is specifically to audit, monitor, and stress-test how models reason — not just people building the models faster.

For job seekers eyeing this space, a few things are worth knowing. These roles increasingly sit inside dedicated safety or alignment teams at labs like OpenAI, Anthropic, and Google DeepMind, rather than being folded into general ML engineering. Backgrounds that show up often include machine learning research, formal verification, cognitive science, and policy analysis — not just software engineering credentials. Nonprofits and research groups such as Redwood Research also function as a talent pipeline, giving candidates a way to build a public track record on interpretability work before or alongside applying to labs directly.

The debate over techniques like opaque recurrence is also a reminder that this field moves on the news cycle as much as the hiring cycle: understanding what chain-of-thought monitoring is, why it matters, and why researchers are worried about losing it is becoming table-stakes knowledge for anyone applying into AI safety, policy, or trust-and-safety roles at a lab — not a niche specialty.

If you're building toward this kind of role, following how labs and independent researchers publicly debate these tradeoffs is itself useful preparation. It shows employers you understand the stakes of the work, not just the tools.

Sources

• OpenAI's new reasoning technique alarms AI safety experts — TechCrunch
• Chain of Thought Monitorability: A New and Fragile Opportunity for AI Safety — arXiv