The Regulator Draws the Same Line

Somewhere in a draft the European Commission opened for public comment this spring, past the sections on watermarks and chatbots and fabricated faces, there is a paragraph that decides when a human being has done enough.

The document is the AI Office’s draft guidelines on the transparency obligations in Article 50 of the EU AI Act, published on 8 May 2026 and open for consultation through early June. Most of it is about machine-readable marking, synthetic media, and the labelling of deepfakes. The part I kept returning to is narrower and, for a profession, far more consequential. It is the rule for AI-generated text, and the single exception carved out of it.

The rule is simple. When an AI system generates or manipulates text that is published to inform the public on a matter of public interest, the text has to be labelled as artificially produced. The exception is where it gets interesting. The label is not required when the text has undergone genuine human review or editorial control, and a named person holds editorial responsibility for it. To write that exception, the Commission had to do something it plainly did not set out to do — it had to define when a human’s involvement in machine output is real.

This series has spent five essays circling that same definition from the other side. I did not expect to find it being drafted into law by people whose concern is not professional accountability at all.

---

The exception that had to define a signature

Read the carve-out closely, because every clause in it is load-bearing.

A label is not required, the guidelines say, where two things are both true. First, the text has been through human review or editorial control. Second, a natural or legal person holds editorial responsibility for the publication. On their own those phrases could mean very little. A general counsel’s name on a masthead, a policy document in a drawer, a junior who clicked approve. The Commission anticipated exactly that reading and closed it off.

Human review, the guidance specifies, means deliberate examination of the substance of the content by a person with relevant competence and professional judgement in the subject matter. Editorial control means a responsible entity with the authority to approve, alter, or reject the substance on substantive grounds, including checking facts and the trustworthiness of sources. And then the sentence that does the real work: superficial or purely formal checks, the mere existence of an editorial policy, or cursory approval without substantive engagement cannot satisfy the exception. Editorial responsibility, treated separately, means ultimate legal responsibility, held by a person whose identity is meant to be publicly findable.

Strip the regulatory vocabulary away and look at what is left. To escape the warning label, AI-generated text needs a competent human who engaged with its substance, who had the authority to reject it, and who is identifiable and answerable for it. A rubber stamp does not count. A policy on the wall does not count. A name with no engagement behind it does not count.

I have read that paragraph more times than I can defend. It is a description, in the language of an administrative agency, of a professional signing an opinion.

---

Travelling with the purpose

Before I say where this lands, I want to credit what the regulator is actually after, because it is not what I am after, and the gap is the whole point.

The Commission’s stated aim, set out plainly in the guidelines, is to reduce impersonation, deception, misinformation, and manipulation, and to protect trust in the information ecosystem. The worry is a public square flooded with text that looks authored and reasoned but was generated at scale by no one in particular. The label is a defence against a reader being misled about whether a mind stood behind the words. That is a real problem, it is getting worse, and the instinct behind the rule is sound.

Notice that this has nothing to do with tax, or with professional standards, or with audit. The Commission is not trying to protect the integrity of an opinion. It is trying to protect a reader from a forgery. It reaches the question of meaningful human involvement not because it cares how professional work is made, but because it needs a fair line between text a person stands behind and text no one does. The motive is foreign to mine. The line is identical.

---

The same line, drawn from the other side

Earlier in this series I argued two things that I took to be claims about professional ethics, not about regulation.

The first, in the essay on the assembly line, was that a cross-border tax opinion is built on a converging line that narrows many parallel analyses to one position a named professional must adopt, and that the last station, the signature, can never be automated, because what the signature carries is responsibility, and responsibility has no machine equivalent. The second, in the essay on who has to be in the room, was sharper and easier to miss: signing is necessary but not sufficient. You can put your name on a conclusion a machine reached, read it, agree with it, and sign it, satisfying every rule about whose name goes on the file, while having had no real hand in it. A professional who merely ratifies inherited reasoning has met the letter of the requirement and missed the thing that decides the case.

That second claim is the one I thought was mine to argue. It turns out the European Commission argues it too, in almost the same words, for entirely different reasons. Cursory approval without substantive engagement cannot satisfy the exception. Ratification is not enough. The reviewer has to engage with the substance and be answerable for it. I wrote that a signature without ownership is hollow as a matter of professional honesty. The Commission wrote that a signature without engagement is hollow as a matter of law, and attached a fine to it.

This is the inversion of the move I have made in every prior essay. Each time, I took a strong outside position (the case that AI will replace professionals, the consultancies’ theory of the assembly line, the implementation playbooks) and travelled most of the way with it before marking the one place it breaks for a regulated profession. Here the outsider does not break against the profession. The outsider, working from a motive that has nothing to do with professional accountability, draws the profession’s own line and writes it into binding obligation. When two parties who share no premises reason to the same conclusion, the conclusion is usually not an accident of either one. It is a property of the thing itself.

---

Which line the law rewards

There is a second convergence underneath the first, and it is structural.

The assembly-line essay distinguished two kinds of machine. A diverging line produces autonomous, terminal output: millions of finished results rolling off the belt, each one done when the line is done with it. A converging line narrows toward a single position that a human adopts and owns. I argued that a tax opinion can only be built on the second kind, and that an assembly line which forgets this does not produce cheaper opinions, it produces faster documents no one is answerable for.

The Article 50 text rule sorts machine output into exactly those two bins, and treats them in opposite ways. Text that rolls off a diverging line, generated and published with no substantive human owner, must wear a label announcing that no one stands behind it. Text that comes off a converging line, substantively reviewed and owned by a named, answerable person, is exempt, because the label would be telling the reader something untrue. The regulation never says it is choosing between two architectures. But that is what the exemption does. It rewards the line that ends in genuine human ownership and flags the line that does not. The audit room in the earlier essay reached that verdict from inside the profession. The regulator reaches it from outside, and turns the verdict into a rule with a date.

---

How this gets built, and what it is not

I should be precise about scope before I say anything about my own system, because precision here is the difference between a fair point and an overclaim.

A private cross-border tax opinion is almost certainly not the kind of text Article 50(4) governs. The obligation bites on text published to inform the public, accessible to an indeterminate audience. The guidance is explicit that private professional correspondence and internal documents are not published in that sense, and its own example of out-of-scope output is a summary available only to the user who asked for it. An opinion delivered to one client is the second thing, not a public broadcast. So I am not claiming the Act regulates what ktAi produces, and I am not in the business of selling compliance with it. The overlap is not in the product. It is in the principle, and the principle is the part worth having.

ktAi’s report generation runs as a fully automated pipeline. Parallel agents analyse the question across Israeli law, US law, and the treaty between them; a layered quality gate sends any section that fails its faithfulness, citation, or rule checks back to be regenerated; and a final review pass runs over the assembled draft. None of those steps waits for a human, and the system does not refuse to produce a report. When its confidence is low it surfaces the gaps and flags them rather than dressing them over, but it returns content. The automation is real, and it runs the length of the line.

What the automation cannot do is release the work. For a formal report, the meaningful control sits after the machine is finished, and it is exercised by a person. An advisor with role-based access to the report submits it for review, requests revisions, or approves it, and the approval is blocked at the system level for as long as an unresolved revision request stands against it. That gate is the only point in the system where something can actually stop delivery, and it is operated by a named human, not by any layer of the automation. The machine assembles. A person releases, or declines to.

That is not a feature I added to satisfy a European regulation I had not read when I built it. It is the mechanical form of the two claims this series made on professional grounds: the converging line that ends in a signature, and the signature that has to be substantive ownership rather than a stamp. I built it that way because I could not, in honest practice, ship what a machine produced under no one’s authority. The Commission, with a different worry entirely, has now written down why.

---

The shape of the problem

The honest caveats matter, and they cut in my favour rather than against it. These are draft guidelines, open for consultation, and explicitly non-binding; only the Court of Justice of the European Union can give the Act an authoritative reading. The penalties (up to fifteen million euros or three percent of worldwide annual turnover) and the application date of 2 August 2026 are real, but the interpretation around them is still wet ink. None of that is the point. I am not citing the regulation as authority over my work. I am noting that two parties who share no premises — a profession protecting the meaning of its signature and a regulator protecting a reader from a forgery — have independently arrived at the same definition of when a human’s involvement in machine output is real.

I find that more persuasive than either argument alone. I built ktAi on the conviction that the line between an opinion and a document is a human being who engaged with the substance and will answer for it. I did not need Brussels to validate that conviction. But when the regulator, reaching for an unrelated goal, draws the same line I did, it stops looking like one practitioner’s preference and starts looking like the shape of the problem. Whether a system built on that line holds up, against real authority, in the room where the work is eventually read, is still not something a regulation or an essay can settle. That gets settled one signed report at a time.