Alsup fined the download, not the training: the copyright risk in your stack is provenance

Copyright attaches to copying, not to reading or consuming a work, which is the hinge Cathy Gellis identifies in Alsup's reasoning ^[4]. Learning statistics from a corpus sits downstream of the moment the files were obtained, and it is that moment where a reproduction exists to be litigated ^[2]. For anyone reviewing an AI roadmap, that splits one compliance question into two unrelated ones: whether you can document how every corpus in the pipeline was acquired, and whether the thing you ship serves the same purpose as the material it learned from.

The first is a records problem, and it is the one most teams hold no artifact for. Model cards, eval suites and safety documentation all describe what happened after ingestion. Alsup penalised what happened before it ^[2].

The second is the question Judge Bibas answered against Ross Intelligence, holding that training on Thomson Reuters content to build a competing legal platform had no further purpose or different character than the original ^[8]. Jason Henderson's reading of the case law is that purpose-to-compete is what actually moves courts, and that non-competing uses tend to find a way through ^[9]. Market impact is already one of the enumerated fair use factors ^[10], so where a product sits relative to its training material is an input to the legal analysis rather than a positioning choice. Authors have tried the mirror-image version of the argument, that chatbots compete with them by generating synthetic books, and it has not yet won ^[11].

Then the price. The $1.5 billion measured against the roughly $200 billion in annual revenue Anthropic is projected to reach by 2028 works out at about three quarters of one percent ^[12]. That comparison flatters the defendant, and it is broadly why Gellis reads the outcome as good news for AI training ^[4]. It also shows what the number is not, which is a deterrent scaled to any buyer who inherits a corpus of unclear origin from a vendor and has no such revenue line ahead of it.

The harder part is planning on any of this. Henderson's description of the field is that the law is all over the place and has not caught up to the question ^[7], and the statute judges are interpreting has not been amended since 1976 ^[6]. Alsup's analogy, that an LLM ingesting trillions of words resembles a writer studying literature in order to turn a hard corner and create something different ^[3], is a persuasive sentence from one district court rather than a rule. What holds across both rulings is narrower than the headline figure and more usable: so far courts have been willing to bless the training, unwilling to bless the theft, and willing to bless the output only where it is not standing in the source's market ^[2][8].