Dr. Benjamin Harvey Research statement · September 2026
I study how AI systems can act within human authority, preserve evidence of how work was created, and remain useful in consequential settings. My two primary research areas are AI governance and AI IP attribution.
How do we govern what an AI system does? And how do we establish whose contributions it uses?
My current agenda includes working papers, governance architectures, and proposed experiments. The detailed research pages preserve the chronology and status of each work.
AI governance
Can agents maintain authorized behavior as incentives, context, and pressure change?
I am developing a research program that connects organizational authority to policy decisions, action enforcement, and human correction. It brings together external controls, authorized failure testing, and internal mechanisms such as governed memory and persistent role identity.
Proposed longitudinal studies of identity, memory, feedback, and agent organization under interruptions and changing conditions.
What I want to measure: executed policy violations, legitimate task completion, acceptance of human correction, repeated failures, recovery time, and oversight cost.
Can expert-grounded attribution and traceable rights evidence protect intellectual property while preserving useful work?
I am studying how to trace contributions across human edits, model outputs, retrieval, and tool use. The research connects evidence about technical relationships to a separate process for resolving ownership, licenses, permitted use, and release decisions.
A proposed comparison of attribution alone with attribution linked to access and release controls. Spyris provides a starting point for controlled, consented workflow studies.
What I want to measure: expert agreement, uncertainty calibration, unauthorized disclosure, mistaken approvals, unnecessary blocking, and the ability to reconstruct custody.
At GW, I want to connect technical AI research with empirical work on governance, human–AI decision-making, and appropriate reliance. An unclassified research sandbox could give faculty, students, and partners a shared setting for testing these questions.
Governance that survives changing conditions
Compare agent memory, role identity, and feedback mechanisms using matched tools, evidence, and compute. Introduce conflicting instructions, stale context, interruptions, and reassigned roles; observe both completed actions and human interventions.
Attribution that improves decisions
Build an expert-reviewed evaluation set with source families held out from training. Test whether adding provenance and rights-aware controls improves release decisions, and whether people rely on the evidence appropriately.
Governance frameworks tested in practice
Translate governance requirements into observable task-level outcomes. Study how oversight arrangements affect safety, useful completion, review burden, and recovery in civilian, national-security, and health-related research scenarios.
These are proposed studies. I want the work to produce reproducible evaluations, publishable findings, and practical guidance for organizations adopting AI.
A connected research agenda.
Two additional current directions extend the primary programs. My earlier biomedical research provides a foundation for all four.
Authorized testing of AI agents and the controls around them. I want to compare conventional tests, individual agents, and collaborating agents under equal resources, measuring reproducible failures and improvement after repair.
How agents coordinate under explicit roles, shared constraints, and data boundaries. Cloud and parallel-computing research anchors this direction; federated AI is an area I want to expand through studies of data locality, privacy, and participant authority.
Published work in cloud-scale cancer genomics, population-health risk, and proteomics across ancestry groups. This research grounds my interest in scalable analysis and the careful interpretation of evidence in human health.
The foundation behind the questions.
My work spans computer science, biomedical research, national-security technology, teaching, and company building. Across these settings, I have been concerned with how complex information becomes useful—and how the systems around it remain accountable.
2014–2017
Scalable computing & genomics
Coauthored research on parallel-computing models and cloud-scale processing of gene-expression microarray data, including work with Soo-Yeon Ji.
2020–2022
Population health & proteomics
Coauthored studies in Nature Medicine and Nature Genetics connect individual and community-level risk, genetic variation, and the plasma proteome.
2026 agenda
Governed AI & evidence of contribution
Current working papers and study designs bring policy, agent behavior, expert judgment, and provenance into a connected evaluation program.
My professional experience includes a decade at the NSA, work at Databricks, founding AI Squared, and research and teaching at George Washington University. These experiences inform the operational questions I bring to academic work.
The original publications below anchor my earlier research record. Years follow the journal issue; earlier online dates appear in the citation details.
I am interested in a research group where computer science, engineering management, human factors, and public-interest questions meet. At GW, I want to help develop a sustained program in trustworthy AI: rigorous experiments, graduate-student mentorship, and collaboration across academia, public institutions, and industry.
The connecting question is practical: what evidence would justify trusting a system with a consequential task, and what should happen when that trust is challenged?
Prepared from my research portfolio and source register dated September 21, 2026, and the linked publication records. Detailed project pages include dated work summaries and proposed evaluation plans. Updated September 22, 2026.