NO FAKES has two safe harbors. Only one of them is about takedown.
The NO FAKES Act of 2026 advanced unanimously out of Senate Judiciary in June, with a companion bill in the House. It is being read almost everywhere as a takedown bill: a platform removes on notice, a platform is safe.
That reading is correct as far as it goes. It also skips the half of the bill that matters most to anyone running a generative model.
Two safe harbors, two different defendants
The one everyone is reading. An online service that stores or links to an unauthorized digital replica is not liable if it removes the material on proper notice, as soon as is technologically and practically feasible, and maintains a policy for terminating repeat violators. The 2026 version adds a counter-notification process modeled on the DMCA. This is familiar ground, and it is part of why the bill has industry support. Platforms already run this machinery for copyright.
The one that gets skipped. The bill separately addresses products and services capable of producing digital replicas. That safe harbor covers manufacturing, importing, offering, providing and distributing such a product, and it is lost if the product is primarily designed to produce unauthorized digital replicas, has only limited commercially significant purpose or use other than producing them, or is marketed for that use.
Read the two side by side and notice what is absent from the second one. No notice. No takedown. No removal window. No response time. How fast you take something down is irrelevant to whether your tool qualifies.
That structure is not new. It is the Sony test, and it asks about the character of the product rather than the diligence of its operator.
What this means if you make files rather than store them
If you host user uploads, you have a compliance path with a known shape and a known cost. Build the notice queue, staff it, remove inside the window. Unpleasant, bounded, survivable.
If your product generates the media, that machinery does nothing for you. Your safe harbor turns on what your product is and what it is for, assessed after the fact, by someone with a completed harm in front of them.
“We removed it in fourteen days” is an answer to a question nobody asked you.
The question you will be asked is whether your product has substantial commercially significant use beyond producing unauthorized replicas. That is a question about design, not about operations, and the strongest available answer to it is a system that structurally cannot produce one.
Three honest caveats
The bill is not law. It cleared committee unanimously and went to the full Senate. Senators raised First Amendment concerns even about the takedown provisions, and agreed to keep working with the sponsors. It may change.
The takedown safe harbor may be weaker than it reads. At least one firm has noted that the remedies section suggests an online service could still face reduced damages even after complying in good faith, which would make that immunity partial rather than complete.
This is not legal advice and I am not a lawyer. It is a reading of published text. Check it with someone who is.
Why I am writing this
I built Faceward, a permission system for AI likeness generation. A platform asks before it generates, and refuses unless two things both hold: a credential naming a specific person, and a face match against that person. Someone who never enrolled is protected by default, because nobody can produce a credential in their name.
So I have an obvious interest in the second safe harbor mattering.
That interest is exactly why the reading is here rather than behind a pitch. The argument stands or falls on the statutory text, which is public, and you can go check it yourself. If it holds, it holds whether or not you ever look at what I built.
If you run a generation platform and you have been reading this bill as a takedown problem, the thing worth doing this week is reading the products and services section with your own counsel.
← Back