Home / Research
Research program

Research

The Board publishes research to inform its agenda: what organizations actually do, where controls fail, and what good looks like. Research is advisory; it does not set requirements.

Flagship study

State of AI Ownership 2027

The first AIGSB benchmark will ask a simple question with a hard answer: for what share of an organization's AI systems can it name a current, accountable owner? It will also measure unaccountable AI spend, exposure-path prevalence among agents, and how often governance reports state their coverage.

Participation is open to Founding and Practitioner members. Results are shared with participants first, then published in aggregate.

In design · data collection planned for 2027
Members submitstandard measures Anonymizeidentifiers removed Aggregateminimum group size Publishranges and method No statistic is published from fewer than ten organizations
Benchmark method. Individual responses are never published.
Research agenda

Open questions the Board wants answered

Owner decay

How quickly do AI systems lose their owners after reorganizations and departures, and which controls slow it down?

Informs AIGS 100

Allocating shared AI spend

How should platform-wide capacity, shared models and agent-to-agent calls be allocated to owners fairly?

Informs AIGS 200

Calibrating risk ranking

Without labeled incidents, how far can usage-based risk ranking be trusted, and how should it be presented to avoid harm to individuals?

Informs AIGS 300

Retrieval exposure

How much previously unreachable data becomes reachable once retrieval indexes and connectors are added?

Informs AIGS 400

Agent chains

How common are exposure paths that only appear when agents pass data to one another, and how should chains be analyzed?

Informs AIGS 500

Assurance for AI

What would an independent assurance engagement over AI governance look like, and what evidence would an auditor accept?

Informs AIGS 600

Research principles

  • Method first. Every publication states its sample, its method and its limits before its findings.
  • No invented numbers. If a measure was not collected, it is reported as not measured.
  • Privacy by default. Responses are anonymized and aggregated. No individual organization is identifiable.
  • Separation from sponsors. Sponsors do not see responses, do not review findings before publication, and cannot veto results.
Join to participatePropose a research question