Regulation is running behind. Accountability cannot wait.

IThe week we are in

Today, 26 August 2026, Bill Gates published the most alarmed text of his public life: "The turbulent AI era is here. The choices we make now are critical." The man who spent five decades betting on software now writes that, for the first time in his life, he wishes a technology would advance more slowly. His fears: jobs that disappear for good, starting with the young; artificial intelligence in the hands of bad actors; children raised by artificial companions. His proposals: national coordination bodies, an unprecedented international AI organization, occupations deliberately reserved for humans, and taxes on tokens and robots.

Notice what these proposals have in common: they are all state machinery. They require legislatures, treaties and — as Gates himself concedes — cooperation between Washington and Beijing. He also admits he would support a credible plan to slow AI down globally, and does not believe such a plan is possible.

Three weeks ago, on the other side of the ocean, something just as telling happened, almost unnoticed. 2 August 2026 was supposed to be the day the obligations for "high-risk" AI systems in the EU's Artificial Intelligence Act — the centrepiece of the whole construction — became applicable. They did not. Through the "Digital Omnibus", adopted six days before the deadline, they were postponed by sixteen months: December 2027 for standalone systems, August 2028 for those embedded in regulated products. In July 2025, when more than forty European CEOs had asked for precisely this pause, the Commission's spokesperson replied publicly: "there is no stop the clock, there is no grace period". Sixteen months later, the clock was stopped by the Commission itself.

I am not writing this to mock Brussels. Part of the rulebook did enter into force on 2 August: AI systems must now say they are AI, synthetic content must carry machine-readable marking, deepfakes must be disclosed. And the postponement has real causes — the promised technical standards were not ready, the classification guidance came late, many member states had not designated their authorities. That is exactly the point: this is not an accident. It is a structural symptom.

IIBehind by construction

The AI Act was proposed on 21 April 2021 — nineteen months before ChatGPT existed. It was conceived as product-safety law for fixed-purpose systems; the chapter on general-purpose models was bolted on mid-legislation, in 2023, in reaction to ChatGPT. By the time its centrepiece came due, the ground had shifted three times: the era of chatbots (2023), the era of reasoning and the first agents (2024–2025), the era of autonomous work at scale (2025–2026).

Frontier labs now ship a new model every 4–10 weeks. Agents operate computers, purchase goods, and complete tasks measured in hours of human work — and the length of the tasks they can finish doubles roughly every seven months, by METR's measurements. In November 2025, the first cyber-espionage campaign executed 80–90% by an AI agent was documented. Open-weight models within a few points of the frontier can be downloaded and run outside any compliance gate whatsoever.

No parliament legislates at that cadence. The United States has no comprehensive federal AI statute — it has, instead, a revoked executive order, a Justice Department suing the states that legislate on their own, and 109 state AI laws passed in 2026 alone, the strictest of them delayed to 2027. The United Kingdom has no bill before Parliament. The UN's international scientific panel wrote in July 2026 that the technology is advancing faster than scientific understanding and faster than governments' capacity to adapt.

Regulation of the AI frontier is behind, and will remain behind for years — by construction, not by bad faith. Law consolidates consensus; consensus requires stable ground; the ground is moving.

This is not an argument against the AI Act. It is an argument against waiting for it. If accountability for how we use AI depended exclusively on the law, it would be suspended for years to come. We cannot afford that — and we do not need to.

IIIWhat the gap costs

Because, in the meantime, the effects do not wait. The labour data of summer 2026 is split in two halves — and both matter. In the aggregate, calm: US layoffs hit a two-year low in July, and occupational statistics show no shift attributable to AI yet. Underneath, exactly what everyone warned about: Stanford's research on ADP payroll data shows that in the occupations most exposed to AI, employment of 22-to-25-year-olds is shrinking by 3.8% a year — and accelerating — not through layoffs, but through a collapse in hiring. In parallel, AI has been the most-cited reason in US job-cut announcements for five consecutive months. The bottom rungs of the professions — precisely where judgment used to be formed — are being quietly sawn off. Gates is right about who gets hit first.

There is also the other cost, the more European one: non-adoption out of fear or helplessness. An MIT study in 2025 found that 95% of enterprise generative-AI pilots produce no measurable effect on profit and loss — for organizational reasons, not technological ones: adoption without governance. In the European Union, only one company in five was using AI in 2025. Romania sits last, at 5.2% — the same Romania that won the hosting of a European AI Factory in Bucharest, operational from 2027, and is bidding for a €5-billion AI gigafactory on the Black Sea. We are buying the compute and refusing the practice.

A firm that adopts AI without governance produces waste and risk. A firm that refuses AI produces uncompetitiveness. Both are ways of failing — and both are already visible in the data.

IVThe unit of accountability is the signature

Gates's proposals may yet come. But between today and the treaties there is a smaller, older unit of governance, available immediately: the named human who answers for an action. Not "human in the loop" as decoration — a real gate of accountability: the system proposes, a person with a name, a role and something to lose approves, and the record of that approval can never be edited afterwards, by anyone.

This can be installed today, at three levels:

• In companies. AI drafts, analyses, proposes; nothing takes effect until a named person approves; every proposal carries its evidence and its source; approvals land in a journal that cannot be rewritten and that auditors can read. Data stays inside the company's perimeter and trains nobody else's models.
• In professional bodies. Colleges of physicians, bar associations, veterinary and engineering orders can do what the legislator cannot: set the professional standard for AI use now. Tools must cite the profession's own authoritative sources or refuse to answer; the professional signs the outcome, so the professional keeps the decision. A professional body that does this protects both the public and its own young members' path to competence.
• In education. Schools and universities do not have to choose between banning AI and capitulating to it. The third way is gated use: the tool searches and drafts, the student explains, defends and signs. Attestation as pedagogy — because the alternative, on Stanford's data, is a generation that never gets to form.

"Human Reserved" occupations, Gates calls them — a good instinct, aimed at job categories, decided by governments. We would aim it lower: human-reserved decisions, decided by every organization that adopts AI. The list of decisions that require a human signature can be written, in any firm, this week. It requires no treaty.

VProof: it can be built. In Romania. Now.

This is not a thought experiment; it is our daily work.

Custos (custossapiens.org) is a Romanian-language reference for anaesthesia and intensive-care physicians: it answers questions about clinical guidelines with verbatim citations from the documents of the Romanian, European and American societies — and when no source supports an answer, it refuses to answer. It is, by design, not a decision tool: it says what the guidelines say; deciding remains the physician's act. It runs in production, physician-verified, as a free public good.

Lares Pecoris (larespecoris.vet) applies the same architecture to veterinary medicine: withdrawal periods, dosages, official SPCs, with mandatory citation and refusal when the source is missing. It protects the food chain — and the veterinarian remains the only one who signs.

Lex Vigilans (lexvigilans.ro) carries the architecture into fiscal law, for litigation lawyers: sources ranked by authority, citations at element level. It is in development — and obeys the same rule: the instrument orients, the lawyer argues the case.

And in our enterprise work we apply the same discipline at scale: systems that only propose — with no code path that could execute anything on its own; approval by a named human, with a recorded reason; an audit journal that not even an administrator can erase; data that never leaves the client.

None of these instruments waited for a law to require honesty of them. Mandatory citation, refusal without a source, human sign-off and complete audit are stricter than what the postponed rules would have demanded — and they run in production, in Romania, built by a three-person studio. If we can, far larger organizations can.

VIEurope's lane

Europe will not win the contest of the largest models; it does not have to. Draghi put it his own way: unlike earlier general-purpose technologies, AI improves as it is used — so the real contest is the contest of application. Today 43% of American workers use generative AI against 32% of Europeans, and the research says the gap comes from management practice, not talent. Europe's lane — and Romania's — is not waiting for the compute of 2027 or the obligations of 2028. It is governed adoption, now: companies that demand audit trails and data that remains theirs; professions that write their own AI codes; schools that teach the signature.

Regulation will catch up. When it does, it should find practice already better than the law — the way it happened with every technology we ever managed to civilize.

The Romans had an instrument for this: the groma. Before you build the road, you set the line. Artificial intelligence is the fastest road-building technology we have ever had. All the more reason to set the lines first: who decides, who signs, who answers. The frontier belongs to the labs. The line is ours.