Mia Glaese

Mia Glaese is an OpenAI research leader working on human data, alignment and coding agents. In a February 23, 2026 interview, she described herself as a VP of research responsible for the Codex, Human Data and Alignment teams. Her Human Data team collaborated with Frontier Evals on SWE-bench Verified, a coding benchmark she later argued had become less useful for measuring improvements in frontier agents.[1]

SWE-bench Verified curated tasks from Princeton’s original SWE-bench, which asks an agent to resolve a repository issue and grades its patch through tests. Glaese credited its creation to collaboration between Human Data and the Frontier team, represented in the interview by Olivia Watkins. She defended having three experts independently review problems: assessing a problem and its proposed patch requires understanding the surrounding codebase. She also acknowledged that more reviews might have helped.[1]

Glaese considered Verified valuable when it was released, but argued that later score gains could increasingly reward guessing particular implementation details rather than better software engineering. Watkins described tests that demanded an unspecified function or argument name, or additional features absent from the task description. Glaese emphasized the consequence: failing such a test did not necessarily mean an implementation was poor, because the tests accepted only a narrow portion of the viable solutions.[1]

On February 23, 2026, OpenAI recommended SWE-bench Pro, describing Verified as increasingly contaminated. Glaese supported moving toward harder, more representative problems, while warning that Pro would eventually need replacement too. She encouraged researchers to create and share evaluations and supported adopting better benchmarks developed elsewhere.[4][1]

Glaese’s evaluation priorities extended to tasks lasting hours or days, design choices that fit a team’s preferences, readable code and future maintainability. “Is the code nice?” she asked in the interview. She described these qualities as meaningful to people working with coding agents, even when they are harder to measure than whether a small GitHub issue has been resolved.[1]

She identified a practical tradeoff: human judgment can assess whether code is good, but it is slow and complex. Automated tests provide fast grading and straightforward comparisons across the industry. Passing them, however, does not establish whether an open-source maintainer would merge the proposed patch.[1]

In July 10, 2026 reporting, WIRED’s Maxwell Zeff described a staff memo announcing Glaese’s expanded role as VP of research and safety. OpenAI’s safety teams would report to her, with Saachi Jain becoming interim head of Safety Systems and reporting to Glaese. Chief research officer Mark Chen explained the change as a way to give safety an earlier, more direct role in model, product and launch decisions.[6]

Before these 2026 responsibilities, Glaese worked at Google DeepMind on pretraining, evaluations, factuality and reinforcement learning from human feedback for large models. By July 12, 2024, she was OpenAI’s Head of Human Data. Her research involved jointly optimizing human data and training algorithms, integrating human judgment to refine AI behavior. This earlier work provides context for her later explanations of expert review and the limits of automated evaluation.[2][1]