Redwood Research Proposes Disclosure Rules to Protect CoT Monitorability
Redwood Research released a proposal addressing threats to AI monitorability from architectural evolution, endorsed and discussed by safety researcher Ryan Greenblatt. The core concern: if models shift from readable chain-of-thought (CoT) reasoning to thinking in unreadable activation states (so-called "neuralese"), or drop CoT entirely, humans lose a key window into model reasoning.
Confirmed
- Redwood Research's proposal warns that certain architectures could weaken or even remove CoT monitorability.
- The proposal calls on companies to transparently disclose three things: CoT-free reasoning capabilities, other evidence of monitorability, and policies for maintaining monitorability.
- Greenblatt said that based on limited public information, he views Google's Astra as a concerning relevant case.
Unconfirmed
- Earlier rumors claimed labs were racing toward "neuralese" research; per shared content this is not currently true, and it remains unverified.
Why it matters
- CoT monitorability is regarded as one of the key current tools for supervising powerful models. An architectural shift to opaque internal reasoning would significantly shrink the toolkit for safety evaluation and alignment verification; the proposal aims to preserve this supervisory baseline through transparency obligations.
2026-09-11 ~ 2026-09-11 · 5 related posts
Primary sources
- [source] Redwood proposes transparency rules as architectures risk removing chain-of-thought monitorability — RyanGreenblatt · 2026-09-11
- Ryan Greenblatt warns 'neuralese' architectures let AIs think in opaque activations, citing Astra — RyanGreenblatt · 2026-09-11
- Redwood Proposes CoT Monitorability Commitments Amid 'Neuralese' Speedrun Rumors — RyanGreenblatt · 2026-09-11
2 near-duplicate retellings: RyanGreenblatt · RyanGreenblatt