Regulate the application, not the technique. Six words, and they are the whole principle underneath the AI Act, and it is a good principle: a system that screens job applicants is high risk because of what it decides, while the same mathematics inside a spam filter is not.
The exception is the thing the industry spent the last eighteen months building, which is a model with no application at all until somebody points it at one.
The European Parliament adopted the Act on Wednesday, 523 votes to 46 with 49 abstentions. The argument here is that its central architecture was designed for a world of purpose-built systems, that general-purpose models broke it, and that what was bolted on to accommodate them is the part of the law worth watching, because it is the part everybody else will copy.
The architecture, and why it was the right idea
The Act sorts systems into tiers by risk. A small set of uses is prohibited outright. A larger set is high risk and carries obligations: risk management, data governance, documentation, human oversight, accuracy and resilience requirements, conformity assessment before it goes on the market. Below that, transparency duties. Below that, nothing.
Tiering by use rather than by technology is the correct instinct and it is why this law is more serious than most attempts. It avoids the trap of naming techniques, which dates a statute within two years. It puts the obligation on the party that knows the context. And it scales the burden to the stakes, so a company sorting warehouse parcels is not carrying the same compliance load as one scoring loan applications.
A rule of that shape needs one thing to work: every system must have a use you can point at when you classify it.
First test: does a general purpose model have a use
It does not, and this is not a philosophical quibble.
A large model is trained on undifferentiated text and produces a capability. What that capability becomes is decided later, by a developer who builds on it, or by a user typing into a box. The same weights can draft a marketing email, summarize a medical record, or screen a job application, and the model’s builder has no way of knowing which.
Put that into a use-based tier system and the classification has nowhere to attach. The model itself is not high risk, because it does not do anything. The application built on it might be, but the obligations that matter most, about training data, documentation and testing, concern decisions made upstream by somebody who is not the deployer.
So the tiering was sound and had a hole in the middle of it precisely where the most consequential engineering was happening.
Second test: what was bolted on
The answer the Act arrived at is a separate regime for general-purpose models, and its trigger is the interesting part.
Baseline obligations apply to all such models: technical documentation, information for downstream developers, a policy on copyright, and a summary of the training data. Above a threshold, a model is treated as carrying systemic risk and acquires heavier duties, including evaluation, adversarial testing and incident reporting to a new European AI Office.
The threshold is training compute above ten to the twenty-fifth floating point operations.
Read that again, because it is a genuine departure. A law that deliberately avoided naming technologies has ended up regulating on the basis of how much arithmetic was performed during training. Not what the model does, not how it performs, not how many people use it. How much electricity went into making it.
Third test: is compute a sensible proxy
It has one large virtue and two obvious defects, and the virtue is why it was chosen.
The virtue is that it is countable in advance and hard to argue with. A developer knows the figure before training starts. A regulator can check it against hardware and energy records. It requires no judgment, which means it produces no litigation about the classification itself, and for a regulator that is worth a great deal.
The first defect is that it is a measure of cost rather than of capability. Compute and capability have correlated so far, which is the entire basis for using it, and the industry has spent two years working on doing more with less. A threshold set against today’s efficiency will, if that work succeeds, exempt tomorrow’s more capable models.
The second defect is that it is a fixed number in a statute governing a field where the relevant quantities move by orders of magnitude. The Act does provide for the threshold to be updated, which acknowledges the problem without solving it, because updating it is a process and the models will not wait for one.
There is a third consideration that cuts in the Act’s favor and deserves stating. A threshold that is too low is a much cheaper mistake than one that is too high, because the consequence is paperwork for a model that did not need it, rather than no paperwork for one that did.
The verdict
The Act is the most serious regulatory instrument anybody has produced on this subject and its central design is right. The general-purpose regime is the weakest part of it, and it is weakest for an honest reason: it was written to close a hole that the architecture created, under time pressure, against a moving target.
A proxy chosen for administrability rather than accuracy is a reasonable engineering compromise. It is also the kind of compromise that gets optimized against, and a threshold that can be engineered around is not a threshold for very long.
What to watch
Two things over the next two years, neither of which is the compliance deadline everybody will write about.
The first is whether anybody publicly reports training just under the threshold. A cluster of models arriving at a figure slightly below the line would be the clearest possible evidence that the proxy has become a target, and it would be visible in the documentation the Act itself requires.
The second is what other jurisdictions copy. The tiering is the good idea and the compute threshold is the expedient one, and expedient ideas travel faster because they are easier to draft. A legislature elsewhere wanting to be seen to act will find a number far simpler to adopt than a taxonomy of uses.
For anybody building on these models, the practical change arrives before either of those. The documentation and training data summaries flow downstream, which means a developer integrating somebody else’s model will for the first time have a legal basis for asking what is in it, and a supplier with an obligation to answer.
