AI Policy

The Only Binding AI Check Left Is European

OpenAI has staffed recursive self-improvement research and publishes dedicated roles for it. The US office its own blueprint nominated to monitor that risk has had no permanent director since 20 July. From 2 August the EU AI Office is the oversight that can actually reach your vendor.

AI PolicyBy Servola Tech Desk2026-08-025 min read

AI-assisted, edited by humans. | Editor-in-Chief: Leon Soliman | Editorial standards

The Only Binding AI Check Left Is European

Key takeaways

OpenAI's June 2026 federal blueprint calls recursive self-improvement an urgent priority and says existing institutions are not equipped to govern it. The US office it nominated for that job, CAISI, has had no permanent director since 20 July 2026, so the only regulator that can compel your AI vendor today is the EU AI Office.
Treat any vendor assurance that points at US federal oversight as a reference to a proposal, not to an institution. Ask instead which specific AI Act obligation binds the vendor, and get that answer written into the contract rather than into a slide.
From 2 August 2026 the AI Office and national market surveillance authorities enforce the AI Act. The AI Office can demand technical documentation, evaluate a general-purpose model, require corrective measures and fine non-compliance. That is the leverage you actually hold.
OpenAI's own careers site lists recursive self-improvement safety and preparedness researcher roles in San Francisco. Capability work here is a published, staffed function now, so date your vendor-risk review to that fact and not to a future rulemaking.

What OpenAI just staffed

Lilian Weng announced on 27 July 2026 that she was leaving Thinking Machines Lab, the company she co-founded with Mira Murati in February 2025. She gave health as the reason, writing that she did not feel able to continue at the pace a startup requires and that the workload had pushed her beyond what her health could sustain physically. Two days later, on 29 July, a company spokesperson confirmed she had returned to OpenAI to lead a top-level team accelerating internal research. The Information reported that the team's remit covers recursive self-improvement, the process by which an AI system iterates on itself to become more capable. Her departure leaves only Murati and chief scientist John Schulman from the original founding group of six.

The staffing is not only a matter of reporting about one hire. OpenAI's own careers site currently advertises a Researcher, Recursive Self-Improvement Safety role and a Researcher, Recursive Self-Improvement Preparedness role, both in San Francisco, alongside an abuse investigator post scoped to AI self-improvement risk. The preparedness listing describes the work as measuring AI research and development risk by tracking progress toward the automation of technical staff, combining coding time horizons, internal inference spend and staff surveys. That is a published organisational function with a hiring budget, not a research aspiration.

The office its own blueprint nominated has no permanent chief

In June 2026 OpenAI published Democratic Governance of Frontier AI: A Blueprint For A Federal Framework. The document states that the company sees early signs of recursive self-improvement in today's systems, where AI development is itself accelerated by AI, and that this creates governance challenges which existing institutions are not equipped to address. Its remedy is institutional. The blueprint asks the Center for AI Standards and Innovation, CAISI, to treat recursive self-improvement as an urgent priority, to conduct evaluations and recommend mitigations rather than approve or block deployments, and to build standards for independent technical assessments that give policymakers ongoing visibility into progress toward it.

That nominated monitor is currently without a permanent head. Chris Fall led CAISI from April 2026 and departed on 20 July after roughly three months, a departure a Commerce Department spokesperson confirmed to CyberScoop. NIST Director Arvind Raman is serving as Acting CAISI Director, and no permanent replacement has been announced. So within a single fortnight the capability side of the question gained a dedicated team and published job postings, while the oversight side lost its director. Both halves of OpenAI's own proposal are visible; only one of them is staffed.

Where enforcement actually reaches your vendor

This is the part that matters for a buyer in Europe, and it is not what the hiring headlines say. A blueprint is a lobbying document. It describes powers that CAISI does not yet exercise, addressed to a Congress that has not legislated them, at an office that currently has an acting director drawn from its parent agency. Nothing in it can be invoked by a customer. If your assurance that a frontier vendor is being watched rests on the American federal picture, it rests on a proposal.

The European position is the reverse. On 2 August 2026 the Commission's AI Office, working with national market surveillance authorities in each member state, began enforcing the AI Act, and the new transparency rules took effect: systems must disclose that a user is dealing with AI rather than a human, deepfakes must be labelled, and generated or altered content must carry machine-readable marks. Over general-purpose models the AI Office can request technical documentation, evaluate the model itself, require corrective measures and fine non-compliance. Those are exercisable powers held by a standing office, with a start date that has already passed. The oversight asymmetry runs the opposite way to the market's centre of gravity.

What this changes in your next AI contract

Stop accepting oversight as a category. When a vendor answers a governance question by pointing at frontier safety commitments, US federal engagement or a voluntary framework, the useful follow-up is which obligation binds them and who can impose a consequence. Today the answer that survives that question in Europe is the AI Act, enforced since 2 August by the AI Office together with your own national market surveillance authority. Name the obligation in the contract, name the evidence the vendor will produce, and name the notice period you get if the model behind your service changes.

For a buyer in the United Kingdom the calculation is sharper still, because the UK sits outside the AI Act and has no equivalent binding regime for general-purpose models. A British firm that wants enforceable assurance generally has to obtain it contractually, or buy through an entity that is itself inside scope. That is a procurement decision with a cost attached, and it is better made deliberately now than discovered during an incident. The date to work from is 2 August 2026, not the next rulemaking cycle.

Frequently asked questions

What is recursive self-improvement?

It is the process by which an AI system improves itself, so that AI development is accelerated by AI rather than only by people. OpenAI's June 2026 blueprint says it sees early signs of this in today's systems and calls the resulting governance problem one that existing institutions are not equipped to address.

Does the EU AI Office actually have powers over a model I buy?

Over general-purpose AI models it can request technical documentation, evaluate the model, require corrective measures and fine non-compliance. Enforcement began on 2 August 2026, carried out by the AI Office together with national market surveillance authorities in the member states.

Is CAISI shut down?

No. CAISI continues to operate, but it has no permanent director. Chris Fall left on 20 July 2026 after about three months and NIST Director Arvind Raman is acting director. The wider powers OpenAI's blueprint proposes for it are a legislative proposal, not current authority.

The company closest to the risk has told Washington that existing institutions cannot govern it, and has since put a team and a hiring budget behind the capability while the office it nominated waits for a director. Europe's answer is smaller and duller and it has one advantage the American one does not: it is already in force, and it can be cited in a contract this week.

OpenAIRecursive self-improvementCAISIEU AI ActAI OfficeVendor riskProcurement

Three Months In, the US AI Testing Chief Resigned

Chris Fall is leaving CAISI after about three months. NIST director Arvind Raman takes the role in an acting capacity, and no reason has been given.

4 min read

Capital

The Agent Paying Your Invoices Is Not Software

Freehand raised 75 million dollars to run supply chain spend with AI agents. It is priced against the labour budget, not the software budget.

6 min read

Law & Regulation

The EA Buyout Now Hinges on Brussels

The largest take-private in history has cleared EA shareholders and US antitrust. The last real gate is the EU - and not classic merger review, but the Foreign Subsidies Regulation, facing its first blockbuster test.

3 min read

← All articles