A responsible scaling policy (RSP) is a public commitment by an AI developer that links the training and deployment of increasingly capable models to predefined capability thresholds and to safeguards that must be in place before each threshold is crossed. In AI safety, it is a governance instrument rather than a technical method: it states in advance what the developer will measure, which results would count as dangerous, and what the developer will do, including pausing scaling, if the required protections are not ready.
The idea answers a practical problem of artificial intelligence development. Model capabilities often appear unpredictably with scale, so a policy written only after a dangerous capability is observed would arrive too late. An RSP therefore works like an if-then commitment: if evaluations show a model approaching a specified risk level, then specified security and deployment measures become mandatory.
How a responsible scaling policy works
A responsible scaling policy works by running dangerous capability evaluations at defined points in the model lifecycle and comparing the results against capability thresholds that trigger stronger obligations. The policy names the risk domains it covers, typically chemical, biological, radiological and nuclear (CBRN) uplift, offensive cyber operations, and autonomous AI research and development.
Evaluations are scheduled before training large runs, at intermediate checkpoints and before deployment. Because evaluations can underestimate a model, most policies require a safety buffer: testing continues at regular compute or capability increments, and the developer acts when a model comes close to a threshold, not only when it clearly crosses it.
The following table shows the core components that recur across published policies and what each one is meant to achieve.
| Component | What it specifies | Purpose |
|---|---|---|
| Capability thresholds | Levels of dangerous capability, such as meaningful uplift to weapons development | Define when obligations change |
| Evaluation protocol | Which tests run, when, and with what elicitation effort | Detect thresholds before they are crossed |
| Required safeguards | Security and deployment measures tied to each level | Match protection to risk |
| Pause commitment | Halt training or deployment if safeguards are missing | Prevent racing ahead of protection |
| Governance and review | Responsible officer, board oversight, external input | Make decisions accountable |
Safety levels and capability thresholds
Safety levels group models by the severity of risk their capabilities could enable and attach a minimum set of safeguards to each group. Anthropic introduced the first published RSP in September 2023 with AI Safety Levels (ASL), a scheme loosely modeled on biosafety levels used for handling dangerous pathogens.
Each level raises two kinds of requirements at once: deployment standards, which limit how the model can be misused by people who access it, and security standards, which protect the model weights from theft. The table summarizes the general pattern; exact definitions differ between developers and have been revised over time.
| Level | Capability profile | Typical safeguards |
|---|---|---|
| ASL-1 | Systems with no meaningful catastrophic risk | Standard software practices |
| ASL-2 | Early signs of dangerous knowledge, no reliable uplift | Usage policies, basic weight security, red teaming |
| ASL-3 | Meaningful uplift to non-expert misuse in CBRN domains | Hardened weight security, targeted classifiers, misuse monitoring |
| ASL-4 and above | Uplift to state programs or substantial autonomous AI research | Defined progressively as capabilities approach |
A defining feature is that higher levels are specified incrementally. Developers commit to defining the next level before models reach the current one, because detailed safeguards for capabilities that do not yet exist are hard to design with confidence.
How it compares with related safety frameworks
A responsible scaling policy is one member of a family of frontier safety frameworks that share the same if-then logic but differ in terminology and scope. After Anthropic, OpenAI published its Preparedness Framework and Google DeepMind published its Frontier Safety Framework, and many developers signed the Frontier AI Safety Commitments at the 2024 Seoul summit, which asked signatories to publish such frameworks.
The table contrasts an RSP with neighboring concepts in the AI safety topic so that their roles do not blur.
| Instrument | Main question it answers | Relation to an RSP |
|---|---|---|
| Responsible scaling policy | When must stronger safeguards apply? | The overarching commitment |
| Dangerous capability evaluation | Does the model have a risky capability? | Supplies the evidence for thresholds |
| AI red teaming | How can the system be broken or misused? | One input to evaluations and safeguards |
| Model card | What are the model’s properties and limits? | Often reports RSP evaluation results |
| Responsible AI program | How does an organization manage AI harms broadly? | Wider scope, usually not threshold based |
The key distinction is that an RSP focuses on severe, large-scale risks from frontier models, while responsible AI programs also cover bias, privacy and everyday product harms. The two are complementary rather than interchangeable.
Implementing a responsible scaling policy in practice
Implementation turns written commitments into operational procedures with owners, schedules and evidence. A credible program designates a responsible scaling officer, records each capability determination in writing, and lets staff report noncompliance without retaliation.
- Elicitation effort: evaluators use fine-tuning, scaffolding and tool access so that tests approximate what a motivated adversary could achieve.
- Capability report: a documented argument that a model is or is not below a threshold, with the evidence and its uncertainty.
- Safeguards report: a documented argument that required protections are implemented and effective.
- External review: input from independent experts or government AI safety institutes on evaluations and decisions.
- Policy updates: versioned revisions published with explanations of what changed and why.
The decision table below illustrates how evaluation outcomes translate into actions under a typical policy.
| Evaluation outcome | Required response | Reason |
|---|---|---|
| Clearly below threshold | Proceed with current safeguards | Risk is within accepted level |
| Close to threshold | Apply next-level safeguards as a precaution | Evaluations can underestimate capability |
| Threshold crossed, safeguards ready | Deploy with upgraded protections | Risk is mitigated to acceptable level |
| Threshold crossed, safeguards missing | Pause training or deployment | Core commitment of the policy |
Limitations and criticisms
The main limitation of a responsible scaling policy is that it is largely self-imposed and self-assessed. The developer sets its own thresholds, runs most of its own tests and decides whether its safeguards are adequate, so credibility depends on transparency and outside scrutiny.
- Evaluation gaps: tests can miss capabilities that appear only with better prompting, tools or later fine-tuning, a problem related to sandbagging.
- Vague thresholds: qualitative definitions such as meaningful uplift leave room for interpretation.
- Competitive pressure: a pause is costly when competitors continue, which tests the strength of the commitment.
- Revisable commitments: developers can amend their own policies, so critics ask for public reasoning behind every change.
- Narrow scope: catastrophic-risk focus does not address misalignment problems such as deceptive alignment unless explicitly included.
Supporters argue that RSPs create verifiable public commitments and a template for regulation. Critics respond that voluntary frameworks are no substitute for binding rules. Both views agree that external evaluation and clear reporting make a policy considerably more trustworthy.
The short version
A responsible scaling policy is a promise an AI company makes before building stronger systems. It says: we will test each new model for dangerous abilities, and if a model gets too capable, we will add stronger protections first. If those protections are not ready, the company commits to stop and wait.
Think of a laboratory that handles germs. A simple lab can work with harmless samples, but dangerous viruses require sealed rooms and strict access rules. An RSP applies the same idea to AI: the more powerful the model, the stricter the rules around it.
Where it makes a difference
Consider a frontier AI lab finishing a new model. Before release, its evaluation team runs a battery of biology capability tests, including question sets reviewed by virologists and uplift trials in which non-experts attempt complex laboratory planning with and without the model.
The results are ambiguous: the model does not clearly cross the threshold for meaningful uplift, but it scores close to it and outperforms the previous model by a wide margin. Under the policy, closeness is enough. The responsible scaling officer recommends activating the higher safety level as a precaution, which adds input and output classifiers for weapons-related requests, tighter internal access to the weights and bug bounties for jailbreaks.
Launch slips by several weeks while the safeguards are tested. The outcome is that the model ships with protections matched to its assessed risk, and the lab publishes a summary of its capability and safeguards reports so that outside experts can check the reasoning.