How this was made. The version number counts drafts of the text. It does not measure the inquiry behind it, which has run over days and across several AI systems, with argument between those systems and within them, directed, refused and repeatedly redirected by the author. The source material was AI-generated, and then adversarially and iteratively refined across a range of tools — systems built by different companies in different jurisdictions, set against each other and against the author. No one of them produced this text, and no one of them reviewed it alone. The plurality is deliberate rather than incidental. A single model carries a single set of priors about which sources are authoritative, and this series argues that an evidence base narrowed in exactly that way is how a contested question comes to look settled. Using one model to investigate that claim would have been the claim refuting itself. To name a single model on it would credit that model with work that was neither its own nor done in a single pass. The plurality was also necessary, and the record should say why. In drafting, the assisting model repeatedly led with United States institutional sources — a national laboratory, an industry association, a market study nineteen years old — and presented conclusions drawn from them as the state of knowledge. On one occasion European measured data contradicting those conclusions was present in the same research return and was placed below them. Framings were proposed that would have argued against this series’ own position using that evidence base, and offered as rigour. Each was refused by the author and the material rebuilt. That is the mechanism these documents describe, occurring in their own making, and it is recorded because a series arguing that evidence bases narrow without anyone deciding to narrow them cannot credibly claim its own production was exempt. The framing, the corrections and the judgements are the author’s, and so are the errors. How this site is written sets out what is declared on every piece, who checks it, and where the per-piece record lives.
It is an answer to a different question. “Human in the loop” says a person is present. It does not say what the person may refuse, what they see before deciding, how long they have, or what happens if they do nothing.
The piece is direct about where this fails: at the level most institutions actually choose, where a person approves items arriving faster than they can be examined. The gate exists, it is staffed, and it passes everything. Naming a level is not bureaucracy for its own sake — an undeclared level cannot be audited, because there is nothing to audit it against.
Possibly, and the strongest published objection to scales of this kind is engaged with rather than brushed aside. Bradshaw, Hoffman, Johnson and Woods argued in IEEE Intelligent Systems in May/June 2013, in “The Seven Deadly Myths of ‘Autonomous Systems’”, that autonomy is not a quantity that can be slid along a dial, and that treating it as one leads to exactly the failures it is meant to prevent.
That argument is serious and partly correct. The response is that a declared level is not a measurement of autonomy; it is a statement of what an institution has permitted. The permission is discrete even where the capability is not.
It is a solved regulatory problem in one industry. SEC Rule 15c3-5 has required pre-trade controls since 2010, and Commission Delegated Regulation (EU) 2017/589, known as RTS 6, requires them in Europe. Both exist because of what happens without them.
On 1 August 2012 Knight Capital sent more than four million orders, traded 397 million shares, and lost more than $460 million in forty-five minutes. The rule requiring those controls was already two years old, and Knight became the first enforcement action brought under it. That is the point worth taking: these are not a philosophy of automation but limits on what a system may do before a person is involved, written into binding rules — and a rule on the books is not the same as a control in the path of the order.
In Europe, to a degree. The CJEU’s SCHUFA decision (C-634/21, 7 December 2023) read the automated decision-making provisions more broadly than many had assumed, and the Article 29 Working Party guidance adopted in 2017 and revised in February 2018 remains the reference point.
In New Zealand there is no such rule. The word “automated” appears zero times in the Privacy Act 2020, and “algorithm” appears once. That is a statement about the statute’s text rather than about the quality of New Zealand privacy law, and it is checkable in a minute.
The institution that will answer for the result — which is the argument of the piece before this one. What the levels add is that the decision has to be written down before the fact, in terms someone can check afterwards.
The law already says something about this in places: where a decision produces legal or similarly significant effects for a person, the permitted level is not purely an operational preference.
An organisation operating at high autonomy with no declared level, no pre-set bounds, and a record of catching its own errors as reliably as one that declares them. If undeclared autonomy turns out to be just as auditable in practice, the instrument is not earning its place.
What this publication does not claim, and what is outstanding against it in the register.
Nothing outstanding in the register. Every claim in this publication has its evidence recorded, and no question against it is parked. That is a statement about this publication on the date shown above, generated from the register rather than asserted, and it will change when the register does.
Alongside: the publication · sources and provenance