OpenAI publishes its Preparedness Framework

What it is

OpenAI published the Preparedness Framework as a beta document dated 18 December 2023^[1]. The company described it as a living document: not a rulebook fixed once and left alone, but a set of processes meant to be revised as understanding and as the models themselves change^[1]. Its stated purpose is to track, evaluate, forecast and protect against catastrophic risks posed by increasingly powerful models^[1]. Those verbs can be separated: tracking is knowing where the risks lie, evaluating is measuring them, forecasting is judging in advance what a future model might bring, and protecting is deciding what to do about it before and after release^[1].

Why OpenAI said it was needed

OpenAI stated in the document that the scientific study of catastrophic risks from AI had fallen far short of where it needed to be^[1]. That is a notable starting point for a safety framework: the evaluations that would tell a developer whether a model is dangerous were themselves immature. It helps explain why the document is written as a procedure for producing and updating evidence, rather than as a claim that the evidence already exists.

Four risk categories, four levels

The framework monitors four categories of risk: cybersecurity, chemical, nuclear and biological^[2]. What these have in common is that a model which materially helps in any one of them could cause harm that is wide in scale and hard to reverse — beyond what ordinary product-quality checks are built to catch.

Models are rated low, medium, high or critical^[2]. The rating is not a label attached after the fact; it is wired to a release rule, because only models rated medium or below may be deployed^[2]. A model judged high or critical therefore does not ship with a warning attached — it fails a condition for release.

Concretely, if a model is judged high on biological risk, the rule says it may not be deployed. How a rating is arrived at, and where the line between medium and high falls, therefore carries much of the weight in the document, since the release decision sits downstream of it.

Who evaluates, who decides

The Preparedness team was led by Aleksander Madry^[2]. It reports monthly to a new internal safety advisory group, which analyses the work and advises the chief executive and the board^[2]. The decision whether to release a new model is made by OpenAI's leadership on the basis of those reports, but the board can override that decision^[2]. OpenAI also said the work draws on feedback from people outside the company and is reviewed by qualified independent third parties^[2].

Taken together, these lines form a chain: a team does the evaluation, an advisory group analyses it, leadership decides, and the board holds an override.

What it amounts to

For a company that ships models continuously, writing this down turns a general promise of caution into a documented procedure — named risk categories, a rating scale, a release rule keyed to that scale, a team with a named lead, a monthly reporting line, and a written path for an override.

How far such a procedure constrains depends on questions the text leaves open: who assigns a rating, what happens when a rating is disputed, and how often the document is revised. On the last point, the framework's own framing is explicit — it is a beta, and it is meant to live.