Miju Labs

← All buyers · OpenAI

Mia Glaese

VP of Research and Safety, OpenAI · OpenAI

medium confidenceTier A · Talk to firstFrontier labstext · multimodal

Amelia 'Mia' Glaese runs research and safety at OpenAI after the July 2026 reorganisation folded the safety organisation back under research and created her role; safety teams now report to her. Before OpenAI she led DeepMind's Sparrow, the paper that introduced decomposing 'good behaviour' into natural-language rules and asking human raters about each rule separately. That methodology is the intellectual ancestor of every rubric-based human-judgement contract a lab signs, which makes her unusually easy to talk to about taste rubrics.

Background

Glaese joined OpenAI from Google DeepMind in 2024 and was, by late 2024, the company's head of alignment research. She was subsequently VP of Research and Head of Alignment, and on 11 July 2026 was promoted to the newly created VP of Research and Safety when Chief Research Officer Mark Chen merged safety into research, saying safety work would have an earlier and more direct role in shaping model, product and launch decisions. Johannes Heidecke, head of safety systems for two years, left at the same time; Saachi Jain was named interim head of safety systems.

Her DeepMind work is the more useful lens for a data vendor. Sparrow (2022) argued that instead of asking raters for a global 'which is better' judgement you should break the requirement into explicit rules and collect targeted judgements per rule, then train rule-conditional reward models. It also required the agent to surface supporting evidence when collecting preference judgements. That is precisely the shape of a well-specified creative-taste rubric: decompose 'good design' into craft, originality, coherence, and ask separately.

What they run now

All of OpenAI's safety teams plus alignment research, integrated into the frontier-model research org. In practice: model behaviour and policy adherence, refusal and over-refusal calibration, preparedness and launch gating. The work is cross-modal — image and video generation carry the heaviest safety review inside OpenAI — so she sits above both the Model Behavior group and the image/video safety leads.

What they buy

Specifies rather than purchases. What flows from her remit: rule-conditional human preference data, red-team and adversarial rating campaigns, expert graders for policy-sensitive content, and rubric design itself. For creative modalities the safety-adjacent demand is real and large — likeness, style attribution, brand and trademark, cultural-sensitivity judgements on generated images and video — and those are exactly the judgements you need trained human raters rather than crowdworkers for. Her appearance at OpenAI's AI Trainers community event shows she engages with the rating workforce directly. Inference on specific vendors; OpenAI's contracted human-data suppliers are not public in this pass.

Career

  1. 2024Research (Sparrow / alignment), Google DeepMind — Lead author on Sparrow, 2022
  2. 2024Head of alignment research, OpenAI — Reported publicly Oct 2024
  3. 2026-07VP of Research, Head of Alignment, OpenAI
  4. 2026-07 – presentVP of Research and Safety, OpenAI — Safety teams report to her

On the record

Rubric-decomposed human judgement · 2022-09

Break requirements for good behaviour into natural-language rules and ask raters about each rule separately, which yields more targeted judgements and rule-conditional reward models. source

Limits of rule-following · 2022-09

A model that learns to follow the rules can still exhibit distributional biases. source

Interviews & talks

Why it matters here

High, but as a methodology buyer rather than an aesthetics buyer. She will not commission 'is this image beautiful'. She will commission 'here are eleven rules about likeness, style and cultural depiction; grade each separately with rationale'. If Julian can frame creative taste as a decomposed rubric with inter-rater agreement statistics — Sparrow's own framing, handed back to its author — she is one of the few people senior enough to make a cross-modal programme happen. The risk is that her attention is consumed by safety-systems firefighting after three leadership departures.

How to reach

Frame as measurement, not as creative services. Open by citing Sparrow's rule-decomposition and showing the same structure applied to visual judgement, with per-rule inter-rater reliability numbers from a real designer panel. Speculation: she is more likely to be reached through a research-shaped artefact (an arXiv preprint or a well-argued blog post with data) than through a commercial approach, and the OpenAI Forum's AI-trainer community is an unusual side door.

What we could not establish
  • Education not found.
  • Whether image/video safety evaluation sits under her or under the applied image org.
  • Her current public activity level — no 2026 talks or posts of hers were located, only news coverage.
  • X handle confirmed from an X profile listing, not from a bio naming the employer.

Why they are on the map

Human Data lead on GPT-4o system card (2024); co-author of GDPval (expert-graded eval); named VP Research & Safety, safety teams report to her after Heidecke exit (Jul 2026) source — evidence dated 2026-07

Decides whether the problem is funded.

How OpenAI buys data

Has a dedicated Human Data team that 'turns human feedback into reliable signals for training and evaluation' via bespoke campaigns with 'external vendors and AI trainers' and has open PM/TPM/EM/TLM roles on it (2025-26 job posts). Ended its Scale AI relationship after Meta's stake (Jun 2025); reported to be a major revenue source for Mercor and to use Surge and Handshake, while building in-house data capacity. GDPval (Oct 2025) and HealthBench (May 2025, 262 physicians) show it buys large expert-graded eval sets. Image models ship with a named 'Data & Evaluation' team (Dec 2025 credits).

A central Human Data team (PMs, TPMs, engineers) works with all research teams and vendors. Glaese, a former Human Data lead, is now VP Research & Safety (Jul 2026). Image generation sits under Multimodal (Dhariwal), with its own data leads and an 18-person Data & Evaluation group, plus Model Behavior. Sora was shut down in Mar 2026 and Peebles left in Apr 2026. Image work was folded into GPT Image 2 / 2.5.

Known vendors: mercor, surge-ai, handshake-ai, scale-ai (ended Jun 2025)

Sources

  1. OpenAI has folded safety into research again
  2. Mia Glaese profile
  3. Improving alignment of dialogue agents via targeted human judgements (Sparrow)
  4. Amelia Glaese, LinkedIn
  5. OpenAI Loses Sixth Safety Leader in Two Years