A programme director in the Gulf sent me a headline in the second week of August and asked whether it meant what it appeared to mean. The headline said Saudi Arabia had legalised training on copyrighted material. What he wanted to know was whether he could stop having a particular argument with his legal team.

The instrument is real, it is dated, and it commenced on 12 August 2026. Royal Decree No. M/169, approved by Cabinet Decision No. 560/1447 of 27 January 2026, published in the Official Gazette on 13 February 2026, in force 180 days later. It carries a statutory exception permitting reproduction of protected works — without the rights holder's permission, without compensation — for the purpose of developing artificial intelligence products and algorithms, subject to conditions protecting rights holders, with the scope of those conditions delegated to implementing regulations.

The last clause of that sentence is the one his legal team was right about.

Three columns, and the error of reading them as one

The instrument is best understood by refusing to summarise it. There are three distinct things in play and they have three different epistemic statuses: one is law, one is a blank that regulations will fill, and one was never in scope at all.

FIGURE 1 · ANATOMY OF THE EXCEPTION A permission whose boundary is delegated is a permission with an unknown edge. ENACTED · IN FORCE 12 AUG 2026 What the statute grants reproduction of protected works without the rights holder’s permission without compensation for developing AI products and algorithms Royal Decree No. M/169 Cabinet Decision No. 560/1447, 27 January 2026 Official Gazette, 13 February 2026 in force 180 days later DELEGATED · NOT YET PUBLISHED What the regulations will fix the conditions protecting rights holders — and therefore: is lawful access required? may a rights holder opt out? commercial and non-commercial development treated alike? which categories of work? what records must be kept? Unknown as of 19 Aug 2026. Every practical question is in here. NEVER IN SCOPE What it does not touch at all personal data — a separate regime, untouched by copyright contract and terms-of-service restrictions on the source the law of any other jurisdiction holding the corpus, the training run, or the deployed model infringement on the output side, which is a different question from reproduction on the input side Most procurement risk lives here. One column is law. One is a blank the regulations will fill. One was never in scope. Reading all three as a single permission is the error.

Column one is enacted, and it is genuinely significant. A sovereign legislature has answered, by statute, a question that most of the world is answering by litigation or not answering at all. That is a real thing to have. It gives a programme a dated legal basis with a citation, which is a different kind of object from a legal opinion about how a court might rule.

Column two is drawn open on purpose, because its contents are unknown. The statute says the exception is subject to conditions protecting rights holders and hands the definition of those conditions to implementing regulations that had not been published as of 19 August 2026. The open questions are not marginal ones. Is lawful access to the work required, or does the exception reach material obtained however it was obtained? May a rights holder opt out, and if so by what mechanism? Are commercial and non-commercial development treated alike? Which categories of work are in and which are out? What records must a developer keep? Every one of those determines whether the exception is broad or narrow in practice, and none of them has an answer yet.

Column three was never in scope, and it is where most procurement risk actually lives. A copyright exception governs copyright. It does not govern personal data, which sits under a separate regime with its own lawful bases and its own consequences. It does not override the contract or terms of service you accepted to obtain the source — a permission from the state to reproduce a work is not a permission from a counterparty to breach an agreement about it. It does not reach the law of any other jurisdiction in which the corpus, the training run or the deployed model happens to sit. And it addresses reproduction on the input side, which is a different question from infringement on the output side, where a model reproduces something recognisable and someone objects.

So the answer to the programme director was that he could stop having one argument and would need to start having three smaller ones, which is not the answer he wanted but is the one that survives contact with a regulator.

A delegated boundary is a commercial fact

There is a habit of treating delegation as a legal technicality — the substance is enacted, the details will follow, get on with it. For a compliance function that is sometimes fair. For a programme making capital allocation decisions it is not, and the reason is specific.

A programme plans against a boundary. It decides which corpora to acquire, which suppliers to contract with, what to build in-house, and how much of the result it can carry into other markets — and every one of those decisions is a bet on where the edge of the permission sits. A boundary that is enacted but undefined does not remove the bet. It moves the bet from a question a lawyer can research to a question nobody can research, because the answer does not exist yet in any form.

The failure mode is a corpus assembled to the wrong edge. Suppose a programme acquires and processes a large body of material during the interval on the reading that the exception is broad, and the implementing regulations arrive requiring lawful access, or providing an opt-out with retroactive effect, or excluding a category the corpus is full of. The corpus does not become illegal in some dramatic sense. It becomes a thing whose basis is contested, in a model whose weights already carry it, at a point where retraining costs what retraining costs. The bet is settled after the money is spent.

That is not an argument for waiting. It is an argument for building the one asset that pays off under every resolution of the delegation, which is the subject of the last section.

Three postures on the same question

The comparison worth drawing is not between permissive and restrictive jurisdictions. It is between kinds of answer.

FIGURE 2 · THREE POSTURES Same question. Three different kinds of answer. The question: may protected works be reproduced, without permission, in order to train a model? STATUTE Saudi Arabia Answered by legislation. A statutory exception in force since 12 August 2026, with the conditions delegated to implementing regulations that have not been published. WHAT A BUYER GETS a dated legal basis to point at — and an unknown boundary LITIGATION United States Answered case by case. No statutory text-and-data-mining exception. The question is contested through fair-use litigation, turning on the corpus, the use and the court. WHAT A BUYER GETS no general permission, no general prohibition — certainty costs years SILENCE India Not yet answered. No statutory exception of this kind has been enacted, and the question sits outside the instruments the financial and data-protection regulators issue. WHAT A BUYER GETS nothing to cite in either direction The European position is a further variant and is out of scope here. One jurisdiction answered by legislation. The others are answering by litigation, or not yet at all — and only one of the three gives a programme something it can write down.

In the United States there is no statutory text-and-data-mining exception of this kind, and the question is contested through fair-use litigation whose outcomes turn on the specific corpus, the specific use and the specific court. A programme there has neither a general permission nor a general prohibition, and the cost of certainty is measured in years of proceedings. In India no statutory exception of this kind has been enacted, and the question sits outside the instruments the financial and data-protection regulators have issued — which means a programme has nothing to cite in either direction. The European position is a further variant and is out of scope here.

Held that way, the Saudi advantage stops being about permissiveness, which is unproven, and becomes about citability, which is available immediately. A programme operating there can write a sentence in a governance document naming an instrument, a decree number and a commencement date. A programme in a litigation jurisdiction writes a sentence naming a legal opinion. A programme in a silent jurisdiction writes a sentence naming nothing.

Drawn as instrument status rather than as territory, and deliberately: this is a comparison of what kind of answer each jurisdiction has produced, not a ranking of jurisdictions and not a recommendation about where to site a programme. The three postures are also unstable in different ways — a statute can be narrowed by its own regulations, a litigation position can consolidate quickly, and a silence can end without notice.

The honest caveat on all of this is that the characterisation of the Saudi exception as among the more permissive regimes globally comes from law-firm analyses rather than from the implementing regulations, because the implementing regulations do not exist yet. Those analyses are informed and I have relied on them for the instrument's mechanics. I would not rely on them for its breadth, and neither should a programme.

The record that pays off under every outcome

The practical question a programme has is what to do during a delegation. The answer is the thing that is valuable whether the regulations turn out broad or narrow: a provenance record specific enough to make a claim per source rather than once for the whole corpus.

FIGURE 3 · THE PROVENANCE RECORD An exception is a defence. A defence you cannot evidence is a position. THE FIELD THE QUESTION IT ANSWERS — WHICH THE EXCEPTION DOES NOT Source identifier corpus, version or snapshot date What exactly did you reproduce? A permission to copy is worth nothing if you cannot say what you copied. Claimed basis, per source exception · licence · public domain · consent Under which permission? Forcing the claim per source stops one broad exception being asserted over a mixed corpus. Jurisdiction of the reproduction where the copying physically occurred Whose law applied to the act? An exception governs acts within its reach, not acts performed somewhere else. Date, against 12 Aug 2026 when the reproduction happened Was the exception in force when you did it? A corpus assembled before commencement is not covered by it. Personal-data screening result what was found, and what was done Was any of it personal data? The copyright exception does not address this at all — a different regime governs it. Model bill of materials weights bound to the corpus that made them Which model actually carries this data? The only field that survives a fine-tune, a merge or a distillation. Retention of the claim itself with a named owner and a review date Can you still produce this in three years, when the implementing regulations exist and somebody asks? None of these seven fields depends on how the implementing regulations turn out. All seven get harder to assemble the longer you wait.

Take the fields in order, because each exists for a reason the exception itself does not supply.

  1. Source identifier, with a version or snapshot date. Answers what exactly was reproduced. No exception answers this for you, and "a large public corpus" is not an answer anyone can act on.
  2. Claimed basis, per source. Statutory exception, licence, public domain, or consent. Recording it per source rather than once for the corpus is the whole point: a single claim over a heterogeneous corpus is a claim that will fail wherever the corpus is heterogeneous.
  3. Jurisdiction of the reproduction — where the copying physically occurred. A Saudi exception governs acts within its reach. It does not travel with the data to a bucket in another region because the programme is Saudi.
  4. Date of the reproduction relative to 12 August 2026. A corpus assembled before commencement was assembled under a law that had not commenced. That is not a subtlety; it is the difference between a defence and no defence.
  5. Personal-data screening result. The copyright exception does not address personal data at all, and a corpus large enough to be worth arguing about is large enough to contain it.
  6. A model bill of materials linking weights to corpus. This is the only field that survives a fine-tune, a merge or a distillation — and the moment a programme has three derived models, it is the only field that can answer which of them actually carries the disputed material.
  7. Retention of the claim itself, with a named owner and a review date. The question arrives when the implementing regulations exist, which is to say later, which is to say after whoever assembled the corpus has changed teams.

This is where I will disclose my own bias plainly, because the recommendation happens to point toward the kind of thing I build. The system I am building treats an artefact's identity as its content digest rather than its name, and binds the digest that actually served an action into the record of that action — which is the sixth field, generalised. I would make the same recommendation without that, because the field is forced by the structure of the problem rather than by any product. But a reader is entitled to know that the person recommending the record has an interest in records.

What I would not claim

Four limits.

The instrument's mechanics — decree number, cabinet decision, gazette date, commencement — are high-confidence and individually checkable. Its practical breadth is not, and anyone quoting this piece as evidence that Saudi Arabia has broadly permitted AI training has quoted it wrongly.

The Saudi government's own pages were unreachable when this material was gathered, so the instrument's particulars trace to law-firm analyses of it rather than to the primary text. That is a real weakness in the citation chain and it is disclosed in the sources note rather than papered over. A programme relying on this for a decision should pull the primary text.

The comparison jurisdictions are characterised at posture level only. Nothing here is a legal opinion about American fair use or about Indian copyright, and neither summary is fine-grained enough to act on.

And the delegation could resolve in a direction that makes most of the anxiety in this piece look overwrought. If the implementing regulations arrive broad and clear, the correct response will be to say so. The seven-field record will have cost a programme very little and will still be the thing it needs when somebody asks which model carries what.