responsible scaling guideline (RSP)

in English: responsible scaling policy

advancedspreading

A public, voluntary commitment by an AI development company regarding what safeguards it will introduce for increasingly dangerous capabilities of its models before further developing or releasing them.

In plain terms

Imagine virus laboratories: the more dangerous the pathogen they work with, the stricter the airlock, protective gear, and access control. The Responsible Scaling Policy applies the same logic to AI: the more capable a model is, the stronger the safeguards that come with it.

The company pre-defines which capabilities it considers dangerous, for example, if the model were to substantially assist in creating biological weapons or independently carry out a cyberattack. It also specifies what it will do if the model reaches this level: it will strengthen security, withhold the release, or suspend development.

It differs from general ethical statements in that it consists of 'if-then' commitments: if the measurement shows this, then this step follows. This makes it possible to hold the company accountable externally for what it promised and whether it kept it.

It is important that this is not legislation but self-regulation. The company writes it, the company modifies it, and the company largely enforces it.

How it works

The company sets capability thresholds, i.e., levels above which the model could cause severe, large-scale harm.
It assigns a protection level to each threshold: how the model's weights (the trained model) are protected from theft, and how misuse is prevented during use.
New models are tested both before release and during development; for example, experts deliberately try to extract dangerous assistance from them (red teaming).
If the model crosses the threshold but the corresponding safeguards are not yet ready, the original logic dictates that the company will not release it or will slow down development. Newer versions phrase this more flexibly at several companies, making it conditional.
The results are approved by internal owners and management, partly reported publicly, and the policy itself is revised from time to time.

An everyday example

A new chat model appears, and you read in the news that the developer has "activated a higher safety level" for it. This means that according to the company's own policy, the model has reached a capability threshold, so stricter filters and protections have been enabled. You may notice only that the model answers certain chemical or biological questions more cautiously or refuses them.

Who is affected

everyoneThese commitments determine how much safety testing an AI model undergoes before it reaches millions of people.

entrepreneur, executiveWhen choosing an AI provider, this shows how seriously your supplier takes risks and what commitments it makes.

developerSome of the built-in restrictions and refusals in the model follow from these guidelines, so they also influence the behavior of the application or agent you build. → Code

lawyer, citizenVoluntary commitments will gradually become legal expectations, so it is worth understanding what the regulation grew out of. → Law

A common misconception

MythThe responsible scaling policy is a mandatory rule that a regulator enforces on companies.

In realityThis is essentially a voluntary corporate commitment, written by the company itself and open to change at any time. Legal obligations come only from laws on similar subjects, such as the EU AI Act or California's SB 53, which requires the largest developers to publish their safety framework and also to follow it.

Limits and risks

The company acts simultaneously as rule-maker and enforcer, while independent external oversight remains limited for now.
Commitments can change: several companies have relaxed or rephrased earlier promises. OpenAI confirmed in 2025 that it may adjust its requirements due to competitors' moves, and Anthropic replaced its previous strict shutdown commitment with a more flexible one in 2026. Therefore, always check the latest version.
Measuring dangerous capabilities is immature; a test can easily under- or overestimate what a model can actually do.
The guideline focuses on catastrophic risks and does not provide answers for everyday harms (false information, bias, data protection).

Similar, but not the same

EU MI-rendelet (AI Act) – The law enforced by authorities with fines is mandatory, while the responsible scaling directive is a company's own voluntary commitment.
usage policy – The license specifies what you can use the model for, while the responsible scaling policy states what the company commits to during development and release.
responsible AI principles – Those are general value statements, but the guideline is tied to measurable thresholds and prescribes concrete actions.

Related concepts

Broader

AI safety, AI governance, risk management, self-regulation

Narrower

AI safety level, capability threshold, model evaluation, red teaming

On the edge

frontier model, EU MI-rendelet, general-purpose AI model, systemic risk, fitting

A brief history

The first such guideline was published by Anthropic in September 2023, modeled on the security levels of biological laboratories. The concept was described a few days later by the research organization now known as METR (then ARC Evals) as a general framework. In May 2024, at the Seoul AI Summit, sixteen major AI companies committed to publishing a similar security framework.

Where things stand

As of late 2026, most of the largest AI developers have such a document, under different names. The EU's code of practice for general-purpose AI models, published in 2025 (which companies join voluntarily), and some US state laws, such as California's SB 53, already expect a similar framework. However, how the content and compliance are verified varies greatly from company to company.

Status: 2026 November

Also known as

RSP · responsible scaling policy · responsible scaling policy · Responsible scaling policy · frontier safety framework · frontier model security framework · Preparedness Framework · if-then commitments

Sources

This entry was machine-edited and has undergone machine expert review; the “In the news” section updates automatically from Agent Age stories.