New AI Job Titles · Safety
Model Policy
Model Policy is an OpenAI job title for the people who decide what a frontier model will and will not do. The seat sits in Safety Systems, where the Model Policy team aligns model behaviour with human values and norms, and it produces behavioural specifications that reach training, evaluation and monitoring rather than press releases. OpenAI posts it in San Francisco as Model Policy, with domain variants including Model Policy, Frontier Cyber Risk and Model Policy Manager, Chemical & Biological Risk. It is not a government affairs role.
What does a Model Policy role actually do?
It writes the rules the model enforces on itself, in a form engineering can act on. The posted scope runs from designing and maintaining model policies for a risk domain, to translating threat models into behavioural specifications and evaluation criteria, to drawing the practical boundary between a legitimate request and one that would materially help someone cause harm. The output is a taxonomy and a spec, and the spec is then operationalised in training data, evaluation benchmarks and safety monitoring.
That last part is the mechanism worth naming. A refusal policy that lives in a document changes nothing; a refusal policy that reaches the training and evaluation pipeline changes what the model does. The role exists because somebody has to own the translation between those two states, and neither a lawyer nor a research engineer was doing it.
Why does OpenAI call this policy when other labs call it safety?
Because at OpenAI the deliverable genuinely is a policy: a written rule set with domains, thresholds and edge cases, authored by a named owner. Adjacent work at other frontier companies more commonly sits under a Trust & Safety or AI Safety umbrella, where the framing is a research and enforcement discipline rather than an authorship one. Neither naming is wrong. The difference tells you where the work reports and what it is measured on.
For anyone reading job boards, the practical consequence is that this title does not travel. Searching for Model Policy finds OpenAI and misses the equivalent seats elsewhere, in the same way that searching for Forward Deployed Engineer misses the Applied AI Engineer postings that describe the same job.
How is Model Policy different from AI policy?
| Model Policy | AI policy / government affairs | |
|---|---|---|
| Faces | Inward, at the product | Outward, at regulators and legislators |
| Decides | What the model does and refuses | What the company argues for |
| Deliverable | Taxonomies, behavioural specs, evaluation criteria | Submissions, positions, engagement |
| Ships into | Training, evaluation, monitoring | Consultations and public record |
| Adjacent skill | Threat modelling in a risk domain | Regulatory and legislative process |
The two share a word and almost nothing else. A candidate who reads the title as regulatory affairs applies to the wrong job, and a CTO who slots this work under a public affairs reporting line gets specifications nobody ships.
Why are there cyber and bio variants of the same title?
Because the hard part is domain judgment, and domain judgment does not generalise. The cyber variant exists to sit on the line between defensive security research and assistance that materially enables intrusion, across malware, exploit discovery and credential abuse. The chemical and biological variant sits on the equivalent line between supporting drug discovery and public health work and lowering the cost of misuse. Both are dual-use problems where the same request can be either thing depending on who is asking and why.
Splitting one title into named domain seats is a decision with a cost, so it tells you something. It says the company concluded that a generalist policy author could not hold both lines competently, and that getting either one wrong is expensive enough to fund a specialist. The same logic drives the AI red team researcher seat, which attacks the resulting behaviour to find out where the specification is wrong.
Should you create this seat inside your own company?
Only if you train or fine-tune models whose refusal behaviour you actually control. If you deploy someone else’s model through an API, the vendor already owns the refusal policy and your equivalent job is much smaller: choosing acceptable-use boundaries, configuring what your application permits, and monitoring what gets through. That work usually belongs with your AI governance function or the Chief AI Officer, not in a new headcount line.
The transferable idea is narrower than the title. Somebody in your organisation should own, in writing, what your AI systems are permitted to do, and that document should be testable rather than aspirational. Most companies have the first half and not the second.
Title and role scope verified against OpenAI careers postings: Model Policy (San Francisco) · Model Policy, Frontier Cyber Risk · Model Policy Manager, Chemical & Biological Risk (August 2026).
Model Policy: common questions
Is Model Policy a policy job or a technical job?
Technical, in the sense that the output is a specification rather than a position paper. The role designs and maintains the policies that govern model behaviour, translates threat models into behavioural specifications and evaluation criteria, and works with research and engineering to get those specifications into training, evaluation and monitoring pipelines. The title says policy because the deliverable is a rule set. The job is closer to specification authorship than to lobbying.
How is this different from OpenAI’s public policy or government affairs work?
Direction of travel. Government affairs faces outward: legislators, regulators, standards bodies. Model Policy faces inward at the product and decides what the model will and will not do for a user. Both use the word policy and they are not the same function. A candidate who reads the title as regulatory affairs will apply to the wrong job.
Why are there separate cyber and chemical-biological variants?
Because dual-use judgment does not generalise across domains. Deciding where legitimate security research ends and material assistance to an attacker begins requires someone who knows offensive security. The equivalent line in biology requires someone who knows biosecurity. OpenAI has posted Model Policy, Frontier Cyber Risk and Model Policy Manager, Chemical & Biological Risk as separate requisitions, which is what it looks like when a company decides one generalist cannot hold both lines.
Do other AI labs have this role under a different name?
The work exists across the frontier labs. The label does not travel with it. Adjacent work at other companies more commonly sits under Trust & Safety or AI Safety umbrellas, so searching for Model Policy will find OpenAI and miss the others. Read the responsibilities rather than the title, as with most of the roles in this cluster.
What does a Model Policy role pay?
The title is not publicly benchmarked. The sample is small, recent, and concentrated at one company, which is the same limitation that applies to the other titles on this site’s new-roles cluster. Treat any specific figure quoted against this title as unverified until the sample grows.
