DMCA Section 1202 and AI 2026: The Claim That Does Not Need Infringement

How the Section Is Shaped

It is not an infringement claim, and that is the point

Section 1202 sits in the DMCA's integrity provisions rather than in the exclusive-rights sections, so a plaintiff pleads it alongside infringement or instead of it. The conduct it reaches is not copying — it is lying about who owns a work, or quietly deleting the part that said so. A defendant whose copying is ultimately excused still has to answer separately for what happened to the attribution.

Subsection (a) is about false information, (b) is about removal

The first half bars knowingly providing or distributing copyright management information that is false, where it is done with intent to induce, enable, facilitate or conceal infringement. The second half bars intentionally removing or altering CMI, and bars distributing works or copies knowing the CMI was removed — in each case where the actor knows, or has reasonable grounds to know, that it will induce, enable, facilitate or conceal infringement. Two different fact patterns, one statute.

CMI is broader than a copyright notice

The definition reaches the title, the author, the copyright owner, performers and writers in some cases, the terms and conditions for use of the work, identifying numbers or symbols referring to that information, and links to it — when conveyed in connection with copies of the work. A byline, a photo credit, a licence line in a source file header, an IPTC record inside a JPEG and a Creative Commons string all sit inside that definition.

The damages are statutory and per violation

A plaintiff may elect statutory damages for each violation in a range that starts at $2,500 and runs to $25,000, plus costs and attorney's fees at the court's discretion, and there is a criminal provision for wilful violations committed for commercial advantage or private financial gain. What a 'violation' is — per work, per file, per removal event — is exactly the thing litigants fight about, because the multiplier decides the case.

Registration is not a prerequisite

An infringement suit generally waits on the registration process. A civil action under the DMCA's remedies section is not an infringement action, and courts have entertained 1202 claims without the registration predicate that section 411(a) imposes. That removes the delay a defendant usually counts on, and it widens the set of plaintiffs who can sue you this quarter rather than next year.

Five Places Attribution Disappears

None of these were built by anyone thinking about copyright. All of them produce a copy of a work that no longer carries the information identifying it.

Retrieval chunking that drops the byline

A document goes into the index as 800-token chunks of body text. The author line, the publication, the licence footer and the canonical URL were structural elements in the source and are not body text, so the chunker discards them. Every chunk stored and every chunk served is a copy of the work travelling without the information that identified it — which is the literal shape of the removal conduct, whatever the intent behind the parser.

Image transforms that strip IPTC and XMP

Most resize, transcode and optimisation steps drop non-pixel metadata unless told otherwise, and many CDNs do it by default on the fly. Photographers embed creator, credit, copyright notice and usage terms in exactly those records, and stock-licence terms frequently require preserving them. A pipeline that accepts a licensed image and republishes it stripped has produced a distributed copy with the CMI removed, at scale, with a server-side log of every instance.

Summarisation and answer generation without attribution

An assistant that ingests a source with a visible credit and returns a close paraphrase with none is the fact pattern several pending AI cases are built on. The contested legal question is how similar the output has to be before the statute is engaged at all; the contested engineering question is whether your product could attribute and simply does not. The second one is answerable today and it is the one you control.

AI-assisted rewriting published under your own byline

This is the (a) half rather than the (b) half. Taking third-party material, running a rewrite, and publishing it with your organisation's author line and copyright footer supplies information about authorship and ownership that is not true. The statute's scienter element is doing real work there — the question is what the publisher knew and intended — but a content operation that does this systematically is building the record for the other side.

Features that remove watermarks and credits

A tool whose advertised purpose is erasing a visible credit, a watermark or a stamp is the hardest version of this to defend, because the intent element that saves an accidental parser is supplied by your own marketing copy. If the feature exists for a legitimate reason — removing your own customer's own mark from their own asset — say so in the product, gate it, and keep the record of who asserted rights over what.

What Is Genuinely Unsettled

This is an area where the honest answer is that courts disagree. Each of these is a live argument rather than a rule you can build on.

Whether the output has to be identical

Several courts have read the removal provision to require that the defendant distributed an identical copy of the work minus its CMI, which is a hard fit for a model that produces close-but-not-identical text. Other decisions have declined to impose that limit, and the appellate picture is not settled. A defendant relying on the identicality reading is relying on a rule that may not survive, in a forum it may not get to choose.

Whether the plaintiff has standing to be in federal court at all

At least one prominent AI case was dismissed because the plaintiffs could not articulate a concrete injury from the removal itself, separate from the copying. That is a real obstacle for some claimants and no obstacle for others — a publisher that can show lost attribution traffic, a broken licence-compliance chain or a specific downstream reuse is in a different position from one asserting the bare statutory violation.

What 'reasonable grounds to know' asks of an engineer

The double-scienter language is the main defence to an accidental strip: you must know or have reasonable grounds to know the removal will induce, enable, facilitate or conceal infringement. An internal ticket that says attribution is being dropped and is closed as won't-fix converts a parser bug into evidence about what the organisation knew. Write those tickets as if they will be read aloud.

How violations are counted

Per work, per infringing copy, or per act of removal produces wildly different exposure across a corpus of any size, and there is no single answer running through the cases. Assume the worst plausible multiplier when you are sizing the risk of a pipeline that processes millions of documents, because that is the number the demand letter will use.

Six Things to Put in Place

Find every place attribution is dropped

One pass over the ingestion, transform and publication paths, asking a single question at each hop: does the credit that arrived still exist on the thing that leaves? Chunkers, image resizers, PDF-to-text converters, OCR, scrapers, CDN transform rules, and the export path from your CMS. Most teams find at least one hop nobody had considered part of a copyright surface.

Carry CMI as a field, not as text that survives by luck

Attribution that lives inside the body of a document will be lost by any parser that normalises the body. Lift author, owner, source URL and licence terms into structured fields at ingestion and store them on the chunk, the embedding row and the asset record. The retrieval layer can then render them, and you can prove the information was never discarded.

Turn metadata preservation on where the library allows it

Image and document toolchains generally have a flag that keeps the metadata block, and it is generally off by default. Turn it on, then verify with a real file rather than the documentation — read the metadata off a processed asset in your own production bucket, because the CDN layer may be stripping what your application preserved.

Make the generated answer cite

Surfacing the source title, author and link next to a generated answer is a product improvement readers ask for and the cheapest available answer to the whole family of claims. It also converts a hard legal argument about what your model output resembles into an easy factual one about what your interface displayed.

Write the false-CMI rule down for the content team

A one-line editorial policy — we do not put our byline or copyright line on material we did not author, and AI-assisted work is labelled per our disclosure standard — costs nothing and addresses the half of the statute that engineering cannot fix. The exposure here comes from a person making a publishing decision, not from a parser.

Keep the licence terms you were given

Stock licences, Creative Commons grants and content agreements frequently make attribution a condition of the licence. Strip the credit and you have both a statutory problem and a breach that turns an otherwise licensed use into an unlicensed one, which is how a single dropped IPTC field becomes an infringement claim as well.

Questions Product and Legal Teams Ask

Does fair use protect us from a Section 1202 claim?

Not in the way teams assume. Fair use is a defence to infringement — it answers the question of whether the copying was lawful. Section 1202 asks a different question: what happened to the information identifying the work, and did you know that removing it or falsifying it would induce, enable, facilitate or conceal infringement. A use that is ultimately excused as fair can still leave a live dispute about a stripped byline, which is precisely why plaintiffs plead the two together. The practical consequence for a product team is that the fair-use analysis your counsel ran on the ingestion pipeline does not cover this, and the mitigation is different: fair use is argued after the fact, while attribution is a thing you either carried through the pipeline or did not.

Our chunker drops author lines by accident. Is that a violation?

Accident is a real defence, because the removal half of the statute carries two knowledge requirements: the removal or alteration must be intentional, and the actor must know or have reasonable grounds to know that it will induce, enable, facilitate or conceal infringement. A parser that discards structural elements without anyone deciding to is a weak fit for the first and often for the second. The risk is that accident has a shelf life. Once the behaviour is known internally — a ticket, a customer complaint, a rights-holder email — the organisation has notice, and continuing to ship the same pipeline is a different posture from having shipped it unknowingly. Treat the discovery as the start of a clock, fix the field handling, and keep the record showing when you learned and what you changed.

We resize licensed stock images. Does stripping EXIF create exposure?

It creates two kinds at once. The metadata records that photographers and agencies use — creator, credit line, copyright notice, usage terms — sit inside the definition of copyright management information, so distributing the resized file without them is the conduct the removal provision describes. Separately, many stock licences make preserving the credit or the metadata a condition of the grant, so the same act can breach the licence and take the use outside its scope, which reopens ordinary infringement. The fix is unglamorous: set the preserve-metadata flag in your image library, check whether your CDN's transform layer strips it downstream anyway, and verify by reading the metadata off a file served from production rather than from a local test.

Do we have to be registered with the Copyright Office to be sued under 1202?

The plaintiff does not need the registration that an ordinary infringement suit requires. The civil remedies provision creates an action for a person injured by a violation of the integrity provisions, and that is not an action for infringement of copyright, so the statutory precondition on infringement suits has not generally been applied to it. For a defendant this changes the timetable rather than the substance: a rights holder who would otherwise be waiting on an application can file now. It also changes who shows up, because many of the individual creators most likely to notice a stripped credit are exactly the ones who never registered the work.

What is the cheapest change that reduces this risk the most?

Make attribution a structured field rather than text that has to survive a parser, and then render it. Concretely: capture author, owner, source URL and licence terms at ingestion; store them on the chunk or asset row rather than relying on them staying inside the body; and show them next to whatever your product outputs. That single change addresses the removal provision at the point where removal actually happens, produces a record that the information was never discarded, and has the side effect readers reward — a generated answer that names its sources. Pair it with an editorial rule that your byline does not go on material you did not author, which covers the false-information half that no amount of engineering will reach.

The One-File Test

Take one licensed image that entered your system with a credit in its metadata. Fetch the version your production CDN serves today and read the metadata block off it.

If the creator, credit and copyright fields are gone, you have both a statutory question and — under most stock licences — a condition of the grant you are no longer meeting. The count of how many files it happened to is in your own object storage.

Your Pipeline Strips the Byline. That Is Its Own Claim.