OpenAI's research-intern claim needs a board proof file

Define the milestone before accepting it

Write the claim as a testable record: provider, announcement date, target date, exact label, definition, task population, human role, system and model version, environment, evaluation owner, threshold, exclusions, and evidence location. OpenAI defines its research-intern milestone as performing well-defined research tasks under human direction, including work that would take a skilled researcher a few days. That definition does not say the system independently selects a valuable research question, validates a scientific conclusion, runs every step of a project, or operates without steering. Preserve those nonclaims. The board should require management to translate any comparable external or internal milestone into observable pass conditions before using it in strategy, competitive positioning, workforce assumptions, or a timetable. A named milestone is decision-useful only when directors can see exactly what crossed a threshold and what remains outside it.

Reconcile task success with human intervention

For every success rate, retain the eligible sessions, users, task source, estimated human duration, ground-truth rule, uncertain-outcome exclusions, minimum sample rule, observation period, and the people and systems that classified completion. Then join each successful task to the number, timing, and substance of human interventions. OpenAI reports that more than half of successful four-to-eight-hour tasks in its recent analysis involved one or more interventions; the board should not hear successful as autonomous. Classify steering that clarifies an instruction, supplies missing context, corrects an approach, recovers an environment, changes a result, or stops unsafe work. Report failure and uncertainty beside success, and show the distribution by task type and difficulty. A board proof file should make visible whether the system's apparent horizon increased, whether people performed the hardest judgment, and whether results reproduce across users and model versions.

Distinguish activity acceleration from research progress

Keep agent runtime, token use, code contributions, experiment count, support-channel traffic, task completion, validated research findings, and product or safety impact as separate measures. The article says researcher usage, code contribution, and experiment activity increased, while also noting that compute grew, bottlenecks can shift, and activity metrics are hard to interpret. Require a bridge that explains how a measured activity changed a research decision or produced a result that survived expert review, replication, and downstream integration. Attribute environment changes, staffing, model updates, process redesign, and additional compute rather than assigning the full movement to agents. Track discarded experiments, rework, invalid results, safety restrictions, and the scarce human steps that became the new constraint. The CEO's role is to prevent a high-velocity operating measure from being reported to the board as validated scientific acceleration without the intervening evidence.

Set a target-verification cadence

OpenAI says it is making progress toward an automated AI researcher by March 2028. Treat that as an attributable provider target, not a forecast the board can copy into its own plan. Schedule reviews that preserve the prior definition, current definition, disclosed methods, new evidence, model and environment changes, intervention burden, failed and uncertain tasks, bottlenecks, safety constraints, and any external replication. Use an evidence ladder: stated target, defined milestone, method disclosed, internal result, reproducible result, independently challenged result, and demonstrated relevance to the company's own decisions. Record who verified each stage and what strategic assumption it is allowed to update. If the definition or denominator moves, show the discontinuity rather than drawing a smooth progress line. This cadence gives directors a durable way to challenge milestone claims without pretending that one company's internal measure proves a market-wide timetable.

Turn this source into a reviewable decision

For AI for CEOs, use this briefing as a dated decision record rather than a substitute for the source. Preserve Research acceleration: The view inside OpenAI, the exact URL, the September 7, 2026 review date, the supported facts above, the editorial interpretation, the limitations, and any buyer-specific evidence. Link that record to the decisions most directly affected: Board governance and oversight; Leadership capability and decision practice; Strategy and scenario intelligence; Operating-model redesign. State whether the source changes the scope, evidence requirement, control, sequence, or only the language used to describe the decision.

Before action, name the accountable owner, affected population and workflow, exact offering or configuration, source data and rights, human decision point, exception and appeal path, complete cost, expected benefit, failure and stop conditions, retained evidence, and next review date. Keep official facts, provider statements, buyer observations, representative tests, measured outcomes, editorial inferences, and unknowns visibly separate. Reopen the record when the source, offer, model, integration, data, policy, population, responsible person, or measured result changes.

Limitations and unknowns

This briefing uses OpenAI's official September 6, 2026 publication, checked September 7, 2026. It is the lane's one verified post-cutoff material change: a new attributable OpenAI statement about its internal research-agent progress. The page reports OpenAI's own measurements, definitions, interpretation, incidents, and intentions. It does not independently validate task success, establish general customer availability, prove autonomous scientific discovery, quantify total research productivity or ROI, show transfer to another organization or domain, or establish that capabilities, alignment, security, monitoring, or democratic governance will advance together. OpenAI calls the measures preliminary and identifies interpretive limits. Buyer-specific strategy, scientific validation, architecture, security testing, risk assessment, board judgment, and qualified research, safety, cyber, privacy, finance, IP, regulatory, procurement, records, accessibility, and legal review control.