Miju Labs

← All buyers · OpenAI

Preston Bowman

Program Manager, Human Data Team · OpenAI

low confidenceTier B · Map and watchFrontier labstext

Bowman is a programme manager on OpenAI's Human Data team whose visible work is expert-graded evaluation. He is the fourth-listed author on HealthBench, the benchmark built from conversation-specific rubrics written by 262 physicians, and a core Research contributor on GPT-4.5. He is the clearest public example of an OpenAI human-data PM who runs a large, credentialled expert panel end to end.

Background

His public record is short and almost entirely OpenAI-era. A third-party org-chart profile describes him as a programme manager with a background in data science and neuroscience research and lists prior time at Duke University; Duke's Neurobiology department lists a Preston Bowman as an undergraduate researcher, which is consistent but not the same as confirmed.

What is solid is the work. HealthBench (May 2025) is a 262-physician programme producing 48,562 rubric criteria for grading model responses in realistic health conversations — not multiple choice, but expert-authored criteria per conversation. Running that is a human-data programme management job, and Bowman is credited as an author rather than in acknowledgements. He is also a core Research contributor on GPT-4.5 (Feb 2025).

What they run now

Expert-graded evaluation programmes: recruiting credentialled professionals, getting them to author grading rubrics, and turning those rubrics into a benchmark a model can be scored against. HealthBench is the template — domain experts write what a good answer looks like, at conversation level, along behavioural dimensions such as accuracy, instruction-following and communication. That template generalises directly to any domain where quality is a matter of professional judgement rather than a correct answer.

What they buy

Expert panels and the rubric artefacts they produce: recruitment of credentialled professionals, rubric authoring, per-conversation grading criteria, and the grading passes themselves. HealthBench shows the scale he can run (hundreds of experts, tens of thousands of criteria). Whether the physician panel was sourced in-house or through a vendor is not public; the vendor question is inference.

Career

  1. presentProgram Manager, Human Data Team, OpenAI
  2. —Research (neuroscience), Duke University — Listed as prior employer on a third-party org chart; Duke Neurobiology lists a Preston Bowman as undergraduate researcher — identity match not confirmed

Interviews & talks

Why it matters here

Moderate but real. He does not work on images, so he is not a direct buyer of creative data. He matters as a proof and a pattern: the rubric-per-item, expert-authored evaluation design Julian would need to sell for design and visual work is precisely what he already built for medicine. A pitch that says 'HealthBench for design craft — practising art directors authoring per-artefact criteria' is legible to him in one sentence, and he is internally adjacent to the people who do own image data.

How to reach

LinkedIn only; no X handle found. Open by naming HealthBench and the specific design choice in it — conversation-specific rubrics rather than a fixed checklist — and ask whether anything equivalent exists for visual and design output, where there is no ground truth. Offer a worked example: a small set of design briefs with per-brief criteria authored by practising designers, and inter-grader agreement numbers. Speculation: he is more likely to be a referrer than a buyer here, so ask for the right internal owner rather than a deal.

What we could not establish
  • Whether the Duke link is the same person.
  • No education, no start date, no X handle found.
  • Whether he still sits on the Human Data team in 2026 — the third-party org chart is marked unverified and the hard credits stop in 2025.
  • No public statements or talks found at all.

Why they are on the map

LinkedIn profile: "Program Manager - Human Data Team" at OpenAI, San Francisco, California, United States, in role since 2024-10. source — evidence dated 2026-09

Runs vendor projects; decides if a pilot converts.

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. HealthBench paper (author list)
  2. Introducing GPT-4.5 credits
  3. The Org profile: Program Manager, Human Data Team at OpenAI
  4. Duke Neurobiology personnel listing