Office Hours: April 15, 2026

[andrewmchang]

Our next office hours are happening Wednesday, April 15, 2026 at 1630–1700 UTC!

Learn how T&S peers are using AI for safety workflows, ask for support in implementing open safety models, and connect directly with RMC model partners. We'll also discuss how you can contribute to ongoing projects that the RMC is working on!

• Discord event
• Google Meet
• #roost-model-community

Proposed Agenda

Introduction / Updates

• Recap of what the ROOST Model Community is and the purpose of office hours
• Updated ReadME with eligibility criteria!
• Requesting feedback on the guide for open safety models in T&S
• Choosing our next project sprint!

Demo / Presentation

• Child Safety Policies & Datasets from OpenAI and Everyone.AI

Q&A / Show and Tell

• What questions do you have about the RMC? What resources would be most beneficial?
• What progress have you made in testing, using, or evaluating different open safety models?
• What use cases are you most interested in these models for? What would you like to see more of?

Admin:

• Contributing to upcoming RMC projects

What's Next:

• Introduce yourself in GitHub Discussions
• Start a new discussion! Whether it's an implementation question, research findings, or model specific feedback

[cassidyjames]

Notes

RMC Updates

• Updated RMC README
  • Currently links out to Eligibility Criteria sheet; looking to pull this directly into the repo
• First draft of guide for open safety models in T&S merged in!
  • Feedback/iteration from community welcome, please give it a look

OpenAI Teen Safety Policy Pack

github.com/openai/teen-safety-policy-pack

Example policies are not comprehensive; they’re a few basic examples of safety risk that are fairly generic and fundamental/broadly applicable for products addressing teens. Target six specific risks:

• Harmful body ideals
• Graphic violent content
• Graphic sexual content
• Dangerous content
• Age-restricted goods and services
• Dangerous roleplay

Goal is to allow developers to take these examples and use as-is, but also to give examples of what a good policy is and how it can be written; examples written by in-house subject matter experts with deep expertise in teen safety combined with LLM/prompting background.

Policies are structured as Markdown files, designed to be used directly as an LLM prompt with gpt-oss-safeguard to classify user-generated content. Repo also includes synthetically-generated validation datasets for each policy as a CSV.

Use cases:

• Live moderation of content (e.g. social network) to flag conversations for safety risks
• Historical logs from your platform to identify/label content across your platform

Next steps: have people use policies, contributing their own policies to the RMC repo: adapted, stripped down or expanded upon, translated to different languages, etc.

Q&A

• Q: Will we see new gpt-oss-safeguard model developments openly released soon?
  • A: there’s a focus on multi-modal which is harder to release under open licenses; we’ll likely see new developments available via API
• Q: What’s the background/motivation behind how these specific policy areas were determined?
  • A: Caveat, these policies are by no means exhaustive; each platform/developer will need to determine and adapt the correct policies for themselves
  • Specifically intending to help developers with teen safety; policies are focused on areas that are subtly different for teens vs. adults
  • Another criteria for the included policies: how broad is the concern?
• Q: How were these datasets generated/compiled; how did you determine when the quality was good enough to be used?
  • A: wanted to provide a small representative evaluative dataset for each policy
  • Generated synthetically by OpenAI policy team, balancing the included user prompts across the risk levels as defined by the policies
  • Datasets from one company are inherently limited; for example, it may invite more scrutiny from the company’s legal teams and other departments. Ideally more third-party/aggregate datasets can be compiled from a variety of sources.
• Q: How sensitive is the model to changes or variations in policy?
  • A: model is trained to be BYOP, so it’s pretty sensitive; example policies are pretty detailed, including examples and multiple defined labels
  • Policies were originally crafted by policy team at OpenAI, tested against production data to find edge cases, then iterated on by the policy team
  • Since gpt-oss-safeguard has been trained as a safety model, there is an inherent inertia/default judgement; in some areas it’s more or less steerable
  • OpenAI is mostly looking at chatbot-type traffic, which is different from other types of user-generated traffic; datasets shared to the RMC would be really helpful as they develop future models!
  • Would love feedback from RMC community on how their policy creation processes work, and if there are ways these datasets, models, and tooling can improve those processes.
• Q: How much conversation history is recommended to be passed into the classifier?
• Q: Best way to provide feedback on these policies/gpt-oss-safeguard model?
  • A: Share in ROOST Discord or RMC GitHub repo and ROOSTers can bubble feedback up
  • Can also email OpenAI folks directly with feedback, though that’s less helpful for the open community :)