The Roblox NoAI Case Was Dismissed, and the Reason Should Worry Artists

A federal judge dismissed a class action against Roblox on October 2, ruling that the plaintiff had not shown the platform intentionally stripped metadata from his 3D work. The case is a small procedural decision with an outsized signal for anyone relying on opt-out tags to keep their work out of AI training.

The artist, Austin Beaulier, filed in March 2026. His claim was that Roblox used his 3D models for AI training without consent, and that the path to that use ran through the removal of NoAI tags from the assets. The legal theory was a Digital Millennium Copyright Act violation: removing copyright management information from a work.

The case turned on a dataset most people outside the field have never heard of. Objaverse-XL houses millions of 3D works, and the court examined how raw assets get converted into numerical inputs through machine-learning preprocessing. That conversion is where metadata goes. A pipeline that ingests objects and produces tensors has no reason to carry a text tag forward, and dropping it happens as a matter of routine rather than decision.

Judge Beth Labson Freeman agreed with one part of that theory. The NoAI tag does qualify as content management information. Where the claim failed was on intent. The plaintiff could not show Roblox intentionally removed the metadata, and without that showing, the DMCA claim did not survive.

Why the intent requirement is the whole problem

Removing a copyright notice is only a DMCA violation if it was done knowingly and with the wrong state of mind. In a machine-learning pipeline that ingests millions of 3D assets, converting raw files into numerical inputs, metadata is typically dropped as a side effect of preprocessing rather than as a deliberate act.

That distinction is close to unprovable from the outside. A creator can see that their NoAI tag is gone from a dataset. What they cannot see is a document saying someone decided to remove it, or an internal message describing the plan. The pipeline does the removal, and pipelines do not have intent.

The ruling reasoned specifically about Objaverse-XL, the dataset that houses millions of 3D works, and how raw assets get converted into numerical inputs through preprocessing. The court's finding was narrow: the plaintiff had not pleaded facts showing that any removal was intentional.

So the case settled nothing about whether training on artist work without consent is lawful. It established that the tag-removal route is a hard one to walk, because the evidence a plaintiff needs sits inside the defendant's pipeline.

What creators should take from it

The takeaway is that metadata tags are a weak legal instrument compared to terms of service. A platform's own contract with its users is enforceable in a way that a tag embedded in a file is not, because the contract does not require the creator to prove what happened inside a preprocessing step.

That is an uncomfortable conclusion for the many platforms that have built "no-AI" toggles into their upload flows. Those toggles are a real expression of user preference and they may shape the culture of a platform. As litigation strategy, they are doing less work than the people clicking them probably assume.

The ruling also puts a high evidentiary bar in front of artists generally. Cases involving automated training pipelines will usually turn on what the pipeline did, and the logs of what it did are held by the party being sued. A creator can see the output and the absence of their tag. Everything between those two observations sits behind a system they cannot inspect.

That asymmetry runs through every case of this kind. It is the structural reality of suing over an automated process: the defendant possesses the record, and the plaintiff has to reconstruct intent from outside. Courts can compel discovery, but discovery requires surviving a motion to dismiss first, and the motion to dismiss is where this case ended.

The other side of the same week

The same week produced rulings moving in the other direction on adjacent questions. The Third Circuit held in Thomson Reuters v. ROSS Intelligence that copying Westlaw headnotes to train a competing legal-research product was not fair use, marking the first federal appellate decision to adjudicate a fair-use defense aimed at AI training. The court was careful to note its limits, because the model in that case was not generative and the use substituted directly for the original product.

A California judge separately refused to dismiss a suit accusing ByteDance of circumventing YouTube's technological protections to obtain videos for training. In that case, the plaintiffs cleared the pleading bar by alleging ByteDance extracted and reused proof-of-origin token parameters outside the authorized playback environment.

Read together, the three decisions say the outcomes depend heavily on which mechanism the plaintiff uses to reach the training data. Circumventing an access control alleges a discrete act. Removing a metadata tag inside a preprocessing pipeline has to allege a state of mind.

That difference shapes which cases get filed, and therefore which legal questions get answered first. Plaintiffs gravitate toward theories where the evidence is available, which means the growing body of AI copyright litigation is being shaped less by the strongest arguments than by the most provable ones. An artist whose tag was stripped may have the clearer grievance and the harder case.

There is a second structural difference in the ByteDance ruling worth noting. The court relied on the Ninth Circuit's 2017 decision in Disney v. VidAngel, which held that a technological protection measure can operate as both an access control and a copy control, and that authorized methods of accessing a work do not necessarily prevent a measure from qualifying as an access control. That precedent gives plaintiffs a path that does not depend on proving intent, because circumvention is an act rather than a state of mind.

For artists, the practical advice is to read the terms of service, not the tags. The tag is what you can see. The contract is what a court can enforce. And for anyone building a training pipeline, the rulings are a reminder that legal exposure builds up at the preprocessing step, whether or not anyone wrote down an intention.

What the industry should hear

The Roblox ruling is also a signal about how platforms structure their training datasets. If removing a tag is routine in preprocessing, then any platform with a no-AI toggle and a training pipeline has the same exposure profile as Roblox, and the exposure depends on whether a record exists showing the removal was deliberate.

Platforms that want to reduce that exposure have a straightforward option: log the decision. A pipeline that documents its metadata handling, and that honors opt-out signals where it can, is in a different evidentiary position than one that silently drops everything. That does not settle the underlying question of whether the training is lawful. It does change which arguments a plaintiff can bring.

The counterpressure is that documenting intent creates a record. A platform that writes down "we remove metadata tags" has written down a fact an opposing lawyer would like to have. The rational response is to document the policy rather than the outcome, which is what responsible data governance looks like in most other contexts.