When Knowledge Access Becomes Knowledge Extraction

The unresolved lesson from Aaron Swartz meets the age of foundation models: who gets punished, who gets protected, and who gets erased when law struggles to understand new forms of digital power?

The argument over artificial intelligence and copyright often sounds new because the technology is new. Foundation models are not printing presses, photocopiers, search engines, libraries, or peer-to-peer archives. They are systems that ingest enormous bodies of human work and convert that work into statistical capacity.

But the moral conflict is not entirely new.

A previous public article considered Aaron Swartz as a warning about open knowledge, institutional pressure, and the use of blunt legal tools against digital civil disobedience.[1] The AI copyright battles raise a related but inverted question: what happens when the actor crossing the knowledge boundary is not an individual activist, but a corporation with lawyers, lobbyists, infrastructure, and a business model built on scale?

That comparison should be handled carefully.

Swartz and modern AI companies are not the same. Their motives, tactics, legal posture, institutional power, and social consequences differ. But both sit on the same fault line: law built around older categories trying to respond to new forms of digital access, copying, distribution, and value creation.

Access and extraction are not the same thing

Swartz challenged the enclosure of knowledge. In the JSTOR matter, the U.S. Attorney’s Office alleged he downloaded more than four million documents through MIT’s network, while JSTOR’s archived statement said the downloaded content did not include personally identifying user information and that JSTOR’s interest had been securing the content rather than pursuing an ongoing legal matter.[3][4]

Whatever one thinks of the legality or wisdom of his tactics, the public-interest claim was legible: knowledge was being locked away, and access mattered.

Foundation-model training raises a different claim.

The issue is not simply whether people should be able to read. It is whether companies can take books, journalism, art, code, music, and public writing; use those works to create proprietary systems; and then describe the result as innovation rather than extraction.

That is a different moral terrain.

Access asks whether people can reach knowledge.

Extraction asks whether someone can convert other people’s knowledge into private infrastructure without consent, compensation, attribution, or durable memory of who created the work in the first place.

The scale changes the power relationship

The scale of foundation-model training changes the civic meaning of the act.

An individual downloading a corpus can be investigated, charged, pressured, and isolated. A corporation training a foundation model can litigate for years, retain elite counsel, negotiate settlements, shape industry norms, lobby policymakers, and continue deploying products while the legal theory catches up.

That does not mean every AI training claim is unlawful. It does mean the legal process lands differently depending on who is carrying the risk.

For an individual, the process can become punishment.

For a large platform company, the process can become a cost of market formation.

This is the David-and-Goliath inversion. Swartz faced institutions. In the AI cases, authors and artists often face institutions.

The lawsuits are about more than copying

The Authors Guild and a group of authors filed a class-action lawsuit against OpenAI alleging copyright infringement in the use of authors’ works to train GPT models.[5] The Guild’s public statement argues that professional books are especially valuable training material and that authors object to the use of their works without permission or compensation.[5]

Meta faces separate litigation by authors over AI training. WIRED reported that in Kadrey v. Meta, authors alleged Meta used books from shadow libraries such as LibGen, while Meta argued that its conduct was protected by fair use.[6]

These cases are not only about whether a model contains a copy in the old, intuitive sense.

They are about whether copyright law can understand a system that may not reproduce an entire book on command but may still have absorbed the expressive labour, structure, style, knowledge, and market value of millions of works.

They are about whether training is closer to reading, copying, indexing, transformation, or industrial-scale appropriation.

They are about whether a market can be changed by systems built from the very works whose market they may later disrupt.

Attribution is the quiet wound

Copyright discussions often focus on payment. Payment matters. But attribution matters too.

When a book is read by a person, the author remains visible. When a source is cited, the lineage of an idea can be followed. When a library preserves a text, the creator’s name travels with the work. Even piracy, for all its problems, often leaves the author’s name attached.

Foundation models can sever that lineage.

A model can absorb a style, a structure, a factual pattern, a genre habit, or a body of specialized knowledge, then return outputs without showing where the capacity came from. The user sees a generated answer. The original creators disappear into the machine.

That disappearance is not a minor matter.

Culture depends on memory. Scholarship depends on provenance. Journalism depends on traceability. Literature depends on voice. Public reasoning depends on being able to ask who said something, why, and from what evidence.

If training systems turn human work into unattributed capacity, the loss is not only financial. It is civic.

Fair use cannot carry every moral burden

Fair use is an important doctrine. It protects criticism, parody, teaching, research, news reporting, and many socially valuable uses of copyrighted work. It can also permit transformative uses that benefit the public.

But fair use is not a magic word that resolves every question of power.

A use can be technically transformative and still raise hard questions about extraction. A product can be innovative and still depend on uncompensated labour. A company can argue that model training does not substitute for a book in the ordinary sense while still building a system that competes with the human capacity the book represents.

The U.S. Copyright Office’s generative AI training report addresses the difficult fit between copyright law and AI training, including the role of fair use and the competing interests of creators, developers, and the public.[7]

That difficulty is the point. The law is being asked to decide whether industrial-scale ingestion of culture is closer to learning, copying, indexing, or something for which older categories were never designed.

The irony of the open internet

The open internet was built on a belief that information should move. That belief had democratic force. It challenged gatekeepers, widened participation, and let people outside formal institutions learn and publish.

But openness can be captured.

The same public web that allowed ordinary people to speak also created the training surface for companies with enough compute to ingest speech at planetary scale. The same culture of sharing that made digital public life possible also made creators vulnerable to having their work absorbed into systems they did not authorize and may never be credited by.

That is the irony.

The dream was access.

The business model became extraction.

The question of punishment

The Swartz comparison becomes sharpest around punishment and protection.

When an individual crossed a knowledge boundary, the criminal system could describe the act as theft, intrusion, and damage. The archived DOJ release described potential exposure of up to thirty-five years in prison and a one-million-dollar fine.[3]

When companies cross knowledge boundaries at much larger scale, the dispute is more often framed as civil litigation, licensing uncertainty, innovation policy, fair use doctrine, and market adjustment.

The facts differ, and criminal law is different from civil copyright litigation. But the contrast remains morally important.

Who is treated as dangerous?

Who is treated as innovative?

Who must ask permission?

Who can act first and litigate later?

Who bears the pressure while the law catches up?

A better standard for AI and knowledge

A humane digital policy should not collapse into either extreme.

It should not say that any use of copyrighted material in AI training is automatically theft in the ordinary sense. That would ignore real questions about learning, transformation, research, indexing, and public benefit.

It should also not say that because a model is complex and useful, the human work behind it becomes free raw material.

A better standard would ask:

• Was the training source lawfully obtained?
• Was the material public domain, licensed, opt-in, opt-out, or taken from known infringing repositories?
• Can creators discover whether their works were used?
• Is attribution possible at least for high-confidence or retrieval-assisted outputs?
• Does the model substitute for the market of the works used to build it?
• Are creators compensated when their work materially contributes to a commercial system?
• Are public-interest, research, accessibility, and nonprofit uses treated differently from proprietary enclosure?
• Does the law distinguish learning from laundering?

That last question may be the hardest.

If a person reads a book and learns from it, society calls that education.

If a corporation ingests millions of books and turns them into a commercial model without attribution, society has to decide whether “learning” is still the right word.

What should be carried forward

Aaron Swartz’s story should not be flattened into a slogan for AI companies or against them. He challenged enclosure. Foundation-model companies are often accused of practicing extraction. Those are not the same act.

But the unresolved lesson carries forward.

Digital law must be able to distinguish public-interest access from malicious intrusion. It must also be able to distinguish genuine learning from industrial-scale appropriation. It must protect creators without freezing research. It must allow public benefit without letting the largest actors privatize the labour of everyone else.

The question is not whether knowledge should move.

It should.

The question is whether movement becomes a path to shared understanding, or a pipeline that turns human culture into someone else’s private machine.

That is where the AI copyright debate belongs: not only in courtrooms, but in the broader civic argument over power, memory, credit, and the future of public knowledge.

Sources

[1] Pond: When Conviction Meets an Archaic System — https://pond.canuckduck.ca/node/42520

[3] Archived DOJ press release: Alleged Hacker Charged With Stealing Over Four Million Documents from MIT Network — https://web.archive.org/web/20120526080523/http://www.justice.gov/usao/ma/news/2011/July/SwartzAaronPR.html

[4] Archived JSTOR statement: Misuse Incident and Criminal Case — https://web.archive.org/web/20130112202533/http://about.jstor.org/news/jstor-statement-misuse-incident-and-criminal-case

[5] Authors Guild: class-action suit against OpenAI — https://authorsguild.org/news/ag-and-authors-file-class-action-suit-against-openai

[6] WIRED: Meta AI copyright case and shadow-library allegations — https://www.wired.com/story/meta-lawsuit-copyright-hearing-artificial-intelligence

[7] U.S. Copyright Office: Copyright and Artificial Intelligence, Part 3 — Generative AI Training — https://www.copyright.gov/ai/Copyright-and-Artificial-Intelligence-Part-3-Generative-AI-Training-Report-Pre-Publication-Version.pdf