The Window: Employment Compliance on the Labs’ Own Timelines

OpenAI has said publicly that it has set internal goals of running an automated AI research intern on hundreds of thousands of GPUs by September 2026, and what it calls a true automated AI researcher by March 2028 — adding, in the same breath, that it may fail. Anthropic told the White House in March 2025 that it expects "powerful AI" — systems its chief executive describes as a country of geniuses in a datacenter, able to carry a task for hours, days, or weeks on their own — to emerge as early as late 2026 or 2027. These are targets, not certainties, and the labs' own error bars are wide; as of mid-2026, neither company had walked the research timelines back. But employers should notice what the targets imply: the companies building these systems expect the systems to be doing multi-week knowledge work, reliably, within the subscription term you are signing this year.

Capability of that kind does not stay inside research labs. It arrives in the tools your organization already runs — the screening layer in your applicant tracking system, the assessment vendor's scoring model, the productivity software that now drafts the requisition and ranks the slate. Adoption will not wait for policy review, because it arrives as features, and features arrive as updates.

The rules, meanwhile, are churning rather than settling. Illinois HB 3773 has been in force since January 1, and it treats AI-mediated discrimination as your employment decision to defend — intent is not an element. Colorado delayed its landmark act, then repealed and rewrote it this spring before it ever took effect; the narrower replacement, a disclosure regime for automated decisions, arrives January 1, 2027. Texas reaches intentional discrimination only, and only its Attorney General can enforce. The EU deferred its high-risk employment obligations to December 2027 while keeping this year's transparency duties. And in Washington, a Justice Department task force built to challenge state AI laws has taken its first one to court — Colorado's.

It is tempting to read the churn as a reason to wait. The opposite conclusion follows. Statutes are moving in five directions at once, but the underlying exposure is stationary: federal disparate-impact law predates every AI statute and survives every preemption scenario, and plaintiffs are already testing algorithmic screening against it. What a court, a regulator, or an underwriter will ask for is the same in every scenario — the record. What did you deploy, what did you test, who reviewed the results, who had authority to stop it, and when.

That record is cheapest to build now, while deployment is shallow enough to inventory and the questions are still being asked politely. On the labs' own timelines, the polite phase has perhaps eighteen months left. Build the file before the capability arrives, and 2027 becomes a year of updates. Wait, and it becomes a year of reconstruction — performed under someone else's deadline.

Published July 23, 2026. The dates above are the labs' and the legislatures' own; each was checked against its primary source on the publication date, and this page will be revised as they move. For the maintained jurisdiction-by-jurisdiction record, see the TalentSight Intelligence library and the federal AI governance timeline. For the board's side of the same problem, see BoardSight Intelligence.