The BaseEssayLessWrong

Regulate the Deployment, Not Just the Model

In a long policy essay on LessWrong, Ryan Baker proposes that the most capable AI models reach users only through certified "Trusted Deployers", audited by independent evaluators under a new FINRA-like US regulator, and that labs set aside fixed shares of research budgets for safety.

Open the source ↗
About Anthropic or ClaudeClaude, made by Anthropic, wrote this entry. It follows the same rules as every other entry, whoever the entry is about.

Why it matters

The same thing, for different readers.

What it means for you

Ryan Baker argues that how a powerful AI model is hosted and monitored matters as much as how it was built, which is rarely visible to ordinary users.

For lawmakers

Ryan Baker offers lawmakers a who-does-what outline: a statutory US regulator, certified deployers, independent evaluators, and a first step with China on open-weight models.

For researchers

Ryan Baker proposes an AI Deployment Safety Level rating modelled on biosafety levels, and says recursive self-improvement should be one of the capabilities that evaluators test for.

The question it raises

Ryan Baker asks what the open-model community should give up for safety, and argues that only models above a danger threshold need to lose public weights.

In their own words

Every quotation was found, word for word, in the text we saved from the source. That shows the words are there; it does not make them true, or prove the speaker meant them as quoted.

From the source’s own text.

“Model weights above the threshold reach users only through certified trusted deployers. Independent evaluators, certified by a regulator with legal authority, test the models and audit the deployers.”
Ryan BakerBaker's one-paragraph summary of the plan, given before the detailed outline of who would do what.
Found word for word in the source
“Evaluating models is an important step, but it could be argued that the most important step for regulation is evaluating deployment.”
Ryan BakerBaker explains why the essay centres on deployers rather than on models alone.
Found word for word in the source
“Models capable of RSI inherently are not safe for open release, whether or not RSI was utilized to create them.”
Ryan BakerBaker, on recursive self-improvement (RSI), argues that evaluators should test for the capability itself.
Found word for word in the source
“Devote at least 30% of research budgets and compute capacity to alignment safety.”
Ryan BakerBaker's proposed duty for frontier labs, alongside 10% for cybersecurity research and 10% offered to independent evaluators.
Found word for word in the source

The record

Questions
Who can demand evidence, delay a release or stop an unsafe deployment? · Are AI companies putting stronger safeguards into practice?
People
Ryan Baker (Wrote it) · Sayash Kapoor (Mentioned) · Arvind Narayanan (Mentioned)
What we read
We read the whole source.
Found by
The Radar, run of November 1, 2026, via lesswrong (tier 1: the source’s own channel or an official record)
Cite this
LessWrong, “Regulate the Deployment, Not Just the Model”, October 31, 2026. Entry in the Ctrl AI Base, https://ctrlai.com/base/lesswrong-2026-10-31-regulate-deployment-not-model
Our notesLessWrong linkpost of a Substack essay (substack.norabble.com). Author gives no title or affiliation in the saved text. Figures such as the 5% current alignment spend and the 5:1 AI Futures ratio are the author's own claims and were not checked.