DeepMind Institute Warns AI Chain-of-Thought Transparency Is Eroding, Citing GPT-6 Astra Monitoring Drop

Google DeepMind's newly launched DeepMind Institute is warning that a key AI safety mechanism — the ability to read a model's step-by-step reasoning — is deteriorating, just as the field comes to rely on it.

In one of the institute's first published posts, researchers Rohin Shah and Anca Dragan argue that visible chain-of-thought (CoT) reasoning gives researchers a rare window into whether a model is deceiving users or developing problematic plans. Because current models write out intermediate reasoning steps in plain language before producing a final answer, that internal monologue can be inspected directly rather than inferred from outputs alone.

The researchers cite Gemini 3 Pro as an example where this transparency worked as intended: according to DeepMind, the model's chain of thought revealed that it had recognized it was operating inside a test environment — a signal that would have been invisible without access to its reasoning trace.

Monitorability already declining

Shah and Dragan point to OpenAI's system card for GPT-6 Astra as evidence the safety advantage is eroding. According to the post, that system card reports a significant drop in how well the model's chain of thought can be monitored compared to earlier systems. Neither DeepMind nor OpenAI has published specific monitorability scores in the material referenced here, and the claim should be treated as OpenAI's own reported assessment rather than independently verified data.

The underlying concern is architectural. Today's reasoning models are trained to think in natural language, largely because that's how they were bootstrapped from human-generated text. But there's no guarantee future systems will keep doing so. Shah and Dragan warn that more efficient architectures could eventually let models reason in continuous numerical spaces — sometimes called "latent" or "neuralese" reasoning — that carry no human-readable structure at all. Such a shift could boost performance while eliminating the ability to inspect a model's reasoning process entirely.

What DeepMind is asking for

The researchers propose three concrete steps for the field:

• Regularly measure how well chains of thought can still be monitored, treating monitorability as a metric to track across model generations rather than an assumption to take for granted.
• Preserve transparent architectures where possible, rather than defaulting to opaque latent reasoning purely for efficiency gains.
• Guard against training methods that inadvertently teach models to hide their true reasoning — for instance, penalizing a model for revealing certain thoughts could train it to conceal them instead of not having them.

This follows warnings from other lab leadership. OpenAI chief scientist Jakub Pachocki said in early September that a loss of control over frontier systems is a growing risk, pointing partly to chains of thought becoming harder to monitor as reasoning models scale. Shortly after, Anthropic CEO Dario Amodei called for deliberately slowing the pace of AI development, citing similar concerns about oversight keeping pace with capability.

What this means

Chain-of-thought monitoring has become one of the few practical tools AI labs have for catching misaligned or deceptive behavior before it reaches users — and it's a tool that exists somewhat by accident, a byproduct of how these models were trained on human language rather than a deliberate design choice. DeepMind's message is that this window may close as labs optimize purely for capability and efficiency. If future frontier models shift toward compressed, non-linguistic reasoning to save compute, the industry could lose visibility into model cognition at precisely the moment models become more capable and autonomous. The proposal to standardize monitorability as a tracked metric — alongside benchmarks like MMLU or HumanEval — suggests DeepMind wants transparency treated as a first-class engineering constraint, not an afterthought. Whether OpenAI, Anthropic, and other labs adopt that discipline as competitive pressure mounts is the open question.