Talks

Tokenization: Understanding AI's Data Transformation — Pamela Samuelson

27 min read

Pamela Samuelson · 2024 · 3:49 · ▶ Watch

Summary

Central question. When an AI model is trained on copyrighted works, is the copyrighted expression still in the model?

Thesis. The model takes works apart into “building blocks” and rebuilds them as something unrecognisable, so the model “does not embody the training data in a recognizable way”.

How she gets there.

  1. A Lego analogy: take apart a battleship, build the Eiffel Tower from the same bricks, and nobody can tell the bricks once made a battleship.
  2. A plain description of tokenizing: words or word segments get numbers, and related ones are grouped (colors, animals, philosophy, historical figures).
  3. A claim about the training data: it sits in the model in “a certain form” but not as recognisable expression.
  4. A claim about people: computer engineers see works as “bags of words”, not as “expressive elements” a reader enjoys, so they ask why copyright people care.

Close. She has given talks to computer scientists, and they are baffled: they think they are doing innovative, beneficial science and ask why they are getting grief.

Stance. She is explaining a gap in perspective (engineers vs copyright lawyers). She does not, in this clip, say who should win. She uses a metaphor that favours the “transformed, not copied” reading, but she also admits it is “a little hard to put your hand head around”. Treat the clip as an explainer, not a verdict. Her position in the full episode is unknown.

Map

#Point
1Lego analogy: battleship → Tower of London / Eiffel Tower, same bricks, different order
2Tokenizing defined: each word or word segment gets a number
3Grouping of related items: colors, animals, philosophy, historical figures
4Training data is in the copyrighted works, but the model does not embody it recognisably
5Engineers see “bags of words”, not expressive elements; “of course we’re not using the expression”
6Computer scientists are “baffled” by the copyright concern; they see their work as beneficial

Putting it together

Mental model. A language model works in two steps. First, text is chopped into numbered pieces (tokenization; reversible, like a zip file). Second, training turns those pieces into a learned geometry of use (embeddings and transformer layers), so the model keeps patterns and mostly not the pages. Samuelson’s Lego image says: same bricks, different structure, so the structure is not the original. The law asks a stricter question: not whether it looks like the original, but whether the original can be got back out and whether the use competes with it. The 2025–26 evidence says: mostly no, sometimes yes, and courts split on what “sometimes” means.

Open tensions

  1. Is a copy decided by what a person can see, or by what a machine can extract?
  2. Model-level vs output-level liability: one memorized work or a rule for all the weights?
  3. Transformative training vs competing products (Kadrey vs Ross).
  4. How much memorization is acceptable, and who proves it, work by work and model by model?
  5. Whether US, UK and German courts converge. They already differ (Munich vs Getty).

Reading path

  • Start: Part 1 above; this clip once more; the Getty and GEMA summaries (links in Part 1).
  • Then: Cooper et al., Extracting memorized pieces of (copyrighted) books (arXiv 2505.12546); Lemley and Casey, “Fair Learning” (2021); Sag on non-expressive use.
  • Deeper: Grimmelmann and Cooper, “The Files Are in the Computer”; Carlini et al. (2021, 2022); Sennrich et al. (2016) for tokenization; Delétang et al. (2023); the Copyright Office Part 3 report (read the original, not a summary).
  • Next step for this page: ingest the full episode (https://youtu.be/Js535xyiJOo) to see her position on outputs, licensing and fair use.

Source notes. Part 1 current-debate items come from live WebSearch on 30 Oct 2026 (secondary sources; cited inline). The Roots blocks and the ideas are † except where a link is given.

Part 1 — Before you read

Key terms

  • Token. A piece of text (a word, part of a word, or a punctuation mark) the model reads as one unit. Each token has an ID number. Words like “unbelievable” may be split into several tokens.
  • Tokenization. The splitting step. It does not by itself group anything by meaning.
  • Embedding. A list of numbers (a vector) the model learns for each token, so that tokens used in similar ways end up close together. This is the step that produces “colors near colors”. Watch out: she folds this into “tokenizing”.
  • Weights / parameters. The billions of numbers inside the trained model. They are what the company ships or serves. Whether the weights “contain” the works is the core legal fight.
  • Training data. The works the model learns from.
  • Memorization. A model that can output a stretch of a training work (nearly) verbatim. Flag: the word means different things in different fields. In machine-learning research it is an observed behaviour (the model outputs the passage). In copyright law it matters as possible evidence of a copy inside the model. These are not the same claim.
  • Expression vs idea (copyright). Copyright protects the particular way something is expressed (word choice, structure), not facts or ideas. Her phrase “expressive elements” refers to this.
  • Fair use (US). A defence that allows some copying without permission, judged case by case on four factors: purpose and character (including whether it is “transformative”), nature of the work, amount used, and effect on the market.

People and cases named

  • Pamela Samuelson is the only person named. No cases, thinkers or models are named in the clip. Everything below is added.

Her own frame (🎙)

Two worlds read the same file differently. Engineers see “bags of words”. Copyright lawyers see “expressive elements that you and I enjoy”. Her Lego metaphor is a bridge for the lawyers.

The current debate (searched 30 Oct 2026; WebSearch worked)

The clip is from January 2024. Since then the question “is the work in the model?” has moved from theory into rulings and experiments.

Rulings on training and fair use (US)

  • Thomson Reuters v. Ross (Third Circuit, 29 Sep 2026). Affirmed a ruling that copying Westlaw headnotes to train a legal-research AI was not fair use; reported as the first US appellate ruling on AI training. The tool was non-generative, so its reach is debated, and reports differ on whether the full opinion is unsealed (LawNext, IPWatchdog, Plagiarism Today).
  • Bartz v. Anthropic (N.D. Cal., Alsup, 23 Jun 2025) and Kadrey v. Meta (N.D. Cal., Chhabria, 25 Jun 2025). Both found training LLMs on books to be fair use, with different reasoning: Alsup called it “exceedingly transformative” but held that building a library from pirate sites was not fair use; Chhabria turned on market harm and said his ruling was limited (Public Knowledge, Taft).
  • Bartz settlement. A $1.5 billion class settlement over the pirated books got final approval in July 2026 (reports differ on the exact date, 20 or 22 July) (Jurist, Writer Beware).
  • NYT v. OpenAI (S.D.N.Y.). My search found no 2026 merits ruling on output regurgitation; it surfaced a March 2026 discovery dispute over the Times’ prompts (Bloomberg Law). Status beyond that is unverified.

Rulings on “is the work in the model?” (the clip’s real question)

  • GEMA v. OpenAI (Munich Regional Court, 11 Nov 2025). First instance, not final. The court found song lyrics were memorized in GPT-4 and GPT-4o, treated that as a reproduction, and rejected the text-and-data-mining defence; OpenAI said it would appeal (Bird & Bird, Verfassungsblog).
  • Getty Images v. Stability AI (UK High Court, 4 Nov 2025, [2025] EWHC 2863 (Ch)). Went the other way on the weights: the judge held the weights were not an “infringing copy” because they never contained or stored the Getty images. Getty did partly win on trade mark (watermarks) (Bird & Bird, AIPPI). Appeal status not checked.
  • US Copyright Office, Part 3 report (pre-publication, 9 May 2025). Per law-firm summaries (I did not read the report): if a model can output identical or nearly identical copies without the work being in the prompt, there is a strong argument that the weights implicate the reproduction right. It also says training is “often” transformative, case by case, and backs a “market dilution” harm theory (Sidley). Check copyright.gov for the current version.

Technical findings on memorization

  • Cooper, Lemley and co-authors (arXiv 2505.12546, 2025). Tested open-weight models on Books3. Most models do not reproduce most books, but Llama 3.1 70B reproduces some (Harry Potter and the Sorcerer’s Stone, 1984) almost entirely (arXiv, Stanford RegLab).
  • Ahmed, Cooper and co-authors (arXiv 2601.02671, Jan 2026). Report extracting long passages of books from four production models despite safety measures. A press figure (one model above 94% on some books) is secondhand; check the paper (arXiv, PPC Land).

How this lands on the clip. Her claim “does not embody the training data in a recognizable way” is true for most works and models, and false for some. The live legal question is whether the exceptions decide the case for the whole model (GEMA) or only for the outputs (Getty, Kadrey-style reasoning). Whether the position she states in the clip still holds for her is not known from the clip.

The ideas

The clip is one metaphor and two paragraphs. Behind it sit seven traditions. All of this section is 📚 added background, from memory unless a link is given; not checked line by line. None of it is claimed by Samuelson in the clip.

 LANGUAGE AS UNITS & STATISTICS ──┐                   ┌── COPYRIGHT'S IDEA/EXPRESSION LINE
 Gage · Sennrich (BPE) ·       │                   │   Baker v. Selden · Nichols v. Universal
 Firth · Harris · Salton          ▼                   ▼   Feist · merger
                        ┌────────────────────────────────┐
 EMBEDDINGS & GEOMETRY ─►│  "Are the works in the model?" │◄─ WHAT IS A 'COPY'?
 LSA · word2vec ·       │   (Lego bricks → Eiffel Tower) │   §101 · MAI v. Peak ·
 transformers           └────────────────────────────────┘   Cartoon Network
                         ▲             ▲              ▲
   MEMORIZATION & PRIVACY│   FAIR USE: COPY-RELIANT TECH   │  COMPRESSION & INFORMATION THEORY
   Carlini et al. ·      │   Sega · Perfect 10 · Google    │  Shannon · Kolmogorov ·
   Nasr · Cooper         │   Books · Sag · Lemley-Casey    │  "LM is compression"

Where each school of thought is covered:

  • Tokenization — from text compression to subwords: idea 2
  • Distributional semantics and embeddings — “you shall know a word by the company it keeps”: idea 2
  • Copyright’s idea/expression line — why “bags of words” is the crux: idea 4
  • What counts as a “copy”? The fixation tradition: idea 3
  • Memorization research — can you get the data back out?: idea 3
  • Compression and information theory — in what sense is a model “a copy”?: idea 1
  • Fair use for copy-reliant technology — the “non-expressive use” tradition: idea 4

1. The Lego analogy — rearranged blocks

In short. Samuelson offers a Lego or Tinker Toys analogy for how a model uses its training data. You build a battleship, take the bricks apart and build the Tower of London or the Eiffel Tower. The blocks are the same but they are put together so differently that you cannot tell it was once a battleship. She calls it a concrete metaphor people can use.

🎙 In the talk. “so I think actually a nice analogy is to Legos or Tinker Toys”. She builds a battleship, then takes the bricks apart and builds the Tower of London or the Eiffel Tower. “it’s the same building blocks as I used to make the the battleship”, put together so differently that “you can’t tell that that once was a battleship”. She calls it “a very concrete metaphor that people can use”.

📚 Deeper theory

  • What the analogy gets right. Training does break works into small units and relate them in new patterns (idea 2, idea 2).
  • What it hides. In Lego the bricks are conserved and the buildings are made by hand. In a model, nothing is “disassembled”: the works stay in the training corpus, and the model is a separate object tuned to predict them. A better analogy for some critics is a photocopier that sometimes jams and prints a page (idea 3, Roots).
  • Idea/expression. The analogy asserts that the “expression” does not survive reassembly (idea 4).
  • Dependence on the facts. The metaphor fits most outputs. It does not fit memorized cases, where the rebuilt object is the original (Part 1).

📚 Roots: Compression and information theory — in what sense is a model “a copy”?

The problem. A trained model is far smaller than its training set. So it cannot hold everything. But what does it hold?

The point. Two readings. Lossy-compression reading: the model is a lossy copy of the web, so a copy exists in law. Learning reading: it learned regularities and discarded the specifics, which is what a human reader does. Her Lego metaphor leans to the second. Neither reading is settled; the choice shapes the whole legal result.

Chain

  1. Claude Shannon (1948): prediction and compression are two views of the same thing. A good predictor of the next symbol is a good compressor.
  2. Kolmogorov complexity (1960s): the shortest program that reproduces a string.
  3. Delétang et al., “Language Modeling Is Compression” (2023): frames LLMs as compressors.
  4. “Blurry JPEG of the web” (Ted Chiang, 2023, New Yorker): the popular version of the same idea.

Where others push back. Grimmelmann and Cooper (idea 3) argue that what matters is whether the work can be extracted. The Munich court said memorized lyrics are a reproduction in the model. Defenders of the metaphor point to the UK Getty decision.

What she leaves out (in this clip). Memorization, outputs that resemble inputs, and who bears the cost. The clip may not be the whole argument; the full episode is not ingested.

Check yourself

  • Where does the Lego analogy break when a model reproduces a book nearly verbatim?
  • Which party in Getty v. Stability would be most helped by the analogy, and why?
  • What would a “better” analogy for memorization look like?

2. What tokenizing is, and where she blurs it

In short. She describes tokenizing as taking each word, or sometimes segments of words, and assigning a number to them. She then says the process looks for groups, such as the words that are colors, those about animals, and those about philosophy or historical figures. This is where she blurs two steps: the numbering and the grouping by meaning.

🎙 In the talk. “the process of um of tokenizing as the terminology is in the generative AI space is basically really taking each word or sometimes segments of words” and “essentially assigning a number to them”. Then she says the process looks for “here’s all the ones that basically kind of like are colors and here are the ones that are about um animals and these are the ones that are about um about philosophy um or about historical figures”.

📚 Deeper theory

  • Step 1, tokenization (Roots). Text → pieces → ID numbers. Reversible. Matches her first sentence.
  • Step 2, embeddings (Roots). The model learns a vector per token during training. Her “colors / animals / philosophy” grouping describes this step, not tokenization.
  • Step 3, transformer layers (Roots). Context decides which meaning of a token applies (“bank” the river vs the money).
  • Why the blur matters legally. A court asking “what did the defendant copy?” may treat step 1 (an exact lookup) very differently from steps 2–3 (statistical learning). Calling all three “tokenizing” makes the transformation sound complete from the first step.
  • Also: sorting by theme is a tidy picture. Real embedding dimensions rarely line up with human categories.

📚 Roots: Tokenization — from text compression to subwords

The problem. A neural network needs numbers, not letters. Whole words make a huge, brittle vocabulary (every misspelling and rare name is unknown). Single characters make sequences too long. The answer is in between: subwords.

The point. Her “assigning a number to them” is correct for this step: it is a reversible lookup. That matters legally: tokenization by itself loses nothing. Text → IDs → text gives back the exact text. If a lawyer asked “is the work in the token IDs?”, the honest answer is yes, in the same way it is in a zip file. The “unrecognisable” step comes later, in training.

Chain

  1. Philip Gage, “A New Algorithm for Data Compression” (1994). Byte-pair encoding (BPE): repeatedly replace the most frequent pair of adjacent symbols with a new symbol. It was a file compression trick.
  2. Rico Sennrich, Barry Haddow, Alexandra Birch (2016). Applied BPE to machine translation so rare words become pieces of known words.
  3. Taku Kudo and John Richardson, SentencePiece (2018), and the WordPiece family at Google. These make the tokenizer language-neutral and trained on text, not hand-written.
  4. Byte-level BPE (used by GPT-2 and later). Works at the byte level, so any text can be encoded.
  5. Practical consequence. Tokenizers are trained on a corpus, so their vocabulary reflects it. English gets short token sequences; many other languages (including Indic scripts) get longer ones, which costs more and may hurt quality. This is a known fairness issue.

📚 Roots: Distributional semantics and embeddings — “you shall know a word by the company it keeps”

The problem. Meaning is hard to define; use is easier to count. If words appearing in similar contexts have similar meaning, meaning can be computed from text alone.

The point. This is the tradition behind “here are the ones that … are colors and here are the ones that are about animals”. Her description fits the embedding idea, with one caveat: a real embedding space does not hold neat folders; it is a geometry, with thousands of fuzzy dimensions. And the “bags of words” phrase comes straight from Salton-era retrieval. It describes an old abstraction. It is useful as a description of how engineers think about text as data; it is not an accurate description of what a transformer does with order and context.

Chain

  1. Zellig Harris, “Distributional Structure” (1954) and J. R. Firth (1957): “You shall know a word by the company it keeps.”
  2. Wittgenstein, Philosophical Investigations (1953): meaning is use. A philosophical root of the same idea.
  3. Gerard Salton, vector space model (1970s). Documents as vectors. This is where the “bag of words” idea becomes standard: count words, ignore order.
  4. Latent Semantic Analysis (Deerwester et al., 1990). Compress word-document counts into a few “latent” dimensions.
  5. Bengio et al., neural language model (2003); word2vec (Mikolov et al., 2013). Learn dense vectors; “king − man + woman ≈ queen”.
  6. Transformers (Vaswani et al., 2017) and attention. Embeddings become contextual and order-aware, so modern models are not bags of words.

Where others push back. Technical writers point out that the tokenizer is not where “understanding” happens. Lawyers would add that a reversible encoding is, if anything, a copy.

What she leaves out. The training step itself (gradient updates), which is where the “learning” happens.

Check yourself

  • Is the text recoverable from token IDs alone? What does that mean for a “copy” argument?
  • Which step creates the “colors near colors” structure?
  • Why might an English-language corpus tokenize differently from a Tamil one?

3. “Does not embody the training data in a recognizable way”

In short. Her central claim is that the data is in a certain form in the copyrighted works, but the model does not embody the training data in a recognizable way. She admits this is hard to get your head around. It is not how we usually think about the component elements of copyrighted works.

🎙 In the talk. “the data basically is in a certain form in the in copyright works that are part of the training data but the model uh does not embody the training data in a recognizable way”. She adds: “it’s a little hard to put your hand head around that kind of because it’s just not the way that we think about um the component elements of copyright works”.

📚 Deeper theory

  • Why “recognizable” is the word to watch. It is a human-perception test. The statute’s test is broader: “perceived, reproduced, or otherwise communicated, either directly or with the aid of a machine” (Roots, from memory).
  • Evidence. The 2025–26 studies (Part 1, Roots) show a minority of works are recoverable. The legal question is how much recoverability is enough.
  • Two ways to count. Model-level: if anything can be extracted, the weights contain a copy (Munich; the Copyright Office’s reasoning). Output-level: only the infringing outputs matter (closer to Kadrey and Getty).
  • Remedies differ. If the weights are a copy, remedies could reach the model. If only outputs infringe, filters and takedowns may be enough.

📚 Roots: What counts as a “copy”? The fixation tradition

The problem. The reproduction right applies to “copies”. The statute (17 U.S.C. §101, from memory) defines copies as material objects in which a work is fixed and from which it can be “perceived, reproduced, or otherwise communicated, either directly or with the aid of a machine or device”.

The point. Her Lego metaphor argues the model is not a copy because nobody can tell. The fixation tradition says the test is not whether a human can tell by looking, but whether the work can be perceived or reproduced “with the aid of a machine”. If a prompt can bring the passage back, a court may call the weights a copy (the Munich reasoning) even though the weights look like noise. That is the sharpest point of tension with the clip’s metaphor, and the Getty decision shows the opposite result is available.

Chain

  1. White-Smith v. Apollo (1908): piano rolls are not copies of a song because humans cannot read them. Congress overturned this in 1909 and later added “with the aid of a machine”.
  2. MAI v. Peak (9th Cir., 1993): loading software into RAM can be a copy.
  3. Cartoon Network v. CSC (2d Cir., 2008): a copy must last more than a “transitory duration”.
  4. James Grimmelmann and A. Feder Cooper, “The Files Are in the Computer” (2024) (from memory): argues that the right question is whether the work can be extracted from the model, not whether the model “looks like” the work. Pairs with the memorization studies in Part 1.

📚 Roots: Memorization research — can you get the data back out?

The problem. Privacy and security researchers asked: does a trained model leak what it saw?

The point. The Lego analogy has one weak joint: in Lego, the bricks are rebuilt into a different object. Memorization is the case where the model rebuilds the same object. Because memorization rises with duplication, popular works are most at risk. So the honest reading of her claim is “most training data is not embodied recognisably”, with a long tail of exceptions.

Chain

  1. Carlini et al., “Extracting Training Data from Large Language Models” (2021): extracted names, phone numbers and verbatim text from GPT-2.
  2. Carlini et al., “Quantifying Memorization Across Neural Language Models” (2022): memorization grows with model size, with repeated data, and with longer prompts.
  3. Nasr et al., “Scalable Extraction of Training Data from (Production) Language Models” (2023): found a way to make ChatGPT emit training text.
  4. Cooper et al. (2025) and Ahmed et al. (2026) on books (see Part 1).
  5. Mitigations: deduplicating training data, filters on output, and training that discourages verbatim recall. Their effectiveness is contested.

Where others push back. Authors and publishers: the model is built from our work and can say it back. Model developers: extraction needs adversarial prompting, is rare, and is being engineered away.

What she leaves out (in this clip). Any figure or case. The claim is stated without evidence in the clip.

Check yourself

  • What is the difference between “not recognisable to a human” and “not reproducible by a machine”?
  • Why does the model-level vs output-level choice change possible remedies?
  • Which of the cases in Part 1 supports each view?

4. Engineers’ “bags of words” and the culture gap

In short. Samuelson says computer engineers think of copyrighted works as bags of words, not as the expressive elements that you and I enjoy in a novel. Their reaction is that of course they are not using the expression, so why should copyright people care? She has given talks to computer scientists who were baffled. They think they are doing important science that benefits humanity and ask why they are getting so much grief.

🎙 In the talk. “computer Engineers basically they think of of in copyright works as bags of words and they don’t think about it in terms of sort of the expressive elements that you and I enjoy whenever we pick up um a novel”. Their reaction: “of course we’re not using the expression why you know why why should copyright people be interested in this at all”. She has “given quite a few talks to uh to computer scientists” and “they’re just like baffled”. They think “we’re doing like really important science” and “something that actually is beneficial for Humanity” and ask “why are you giving me so much grief”.

📚 Deeper theory

  • A real disagreement of units. Engineers’ unit is a token or a statistic. Copyright’s unit is a “work” and its protected expression (Roots). Each side is consistent inside its own frame.
  • Science vs product. The “important science” defence mixes research and commercial deployment. Courts after the clip (Ross, Kadrey) focus on the product and its market effect (Roots).
  • “Benefit to humanity” is not a fair-use factor by name. It can matter through the first factor (purpose) and public-interest arguments, but copyright law weighs it against the author’s market.
  • Her observation is sociological, not legal. She reports baffled audiences. She does not say they are right or wrong in this clip.

📚 Roots: Copyright’s idea/expression line — why “bags of words” is the crux

The problem. Copyright cannot protect ideas, facts or methods, or every author would own the world. It protects expression. Where is the line?

The point. Her split — engineers’ “bags of words” versus “the expressive elements that you and I enjoy” — is the idea/expression split in another dress. The legal argument she is gesturing at: if training uses only unprotectable patterns (word statistics, facts about language), it takes no protected expression. The reply, from the other side: the statistics are extracted from the expression, and some models can regurgitate it (Part 1).

Chain

  1. Baker v. Selden (US, 1879). A book’s description of a bookkeeping system is protected; the system is not.
  2. Nichols v. Universal Pictures (2d Cir., 1930, Learned Hand). The “abstractions” test: the more general the pattern, the less protection. Stock characters and plot ideas are free.
  3. Feist v. Rural (1991). Facts are free; originality needs a modicum of creativity. A phone directory’s facts are unprotected.
  4. Merger and scènes à faire. When an idea can only be expressed one or very few ways, or a trope is standard, the expression is not protected either.
  5. The “non-literal similarity” cases (Hand, then later circuits) ask whether a work takes “protected expression” without copying word for word.

📚 Roots: Fair use for copy-reliant technology — the “non-expressive use” tradition

The problem. Machines often have to copy in order to analyse. Does intermediate copying that never reaches a human reader count as infringement?

The point. Her last paragraph — “of course we’re not using the expression” — is the lawyer’s version of the non-expressive use argument in plain words. The Ross decision shows where this tradition breaks: when the trained product competes with the source.

Chain

  1. Sega v. Accolade (9th Cir., 1992) and Sony v. Connectix (9th Cir., 2000): reverse-engineering copies were fair use because the end product did not reuse the expression.
  2. Perfect 10 v. Amazon (9th Cir., 2007) and A.V. v. iParadigms (4th Cir., 2009): thumbnails and plagiarism-checker copies were fair.
  3. Authors Guild v. Google (2d Cir., 2015): scanning millions of books to build a search index was fair use.
  4. Matthew Sag, “Copyright and Copy-Reliant Technology” (2009) and later work: the “non-expressive use” idea. Lemley and Casey, “Fair Learning” (2021): machine learning should be treated like this.
  5. Campbell v. Acuff-Rose (1994) introduced “transformative use”; Warhol v. Goldsmith (2023) narrowed it when the use has the same purpose as the original and competes. Ross (Part 1) leaned on this.

Where the seven traditions meet. Tokenization (idea 2) gives a reversible number encoding. Embeddings (idea 2) build a learned geometry from text. Idea/expression (Roots) asks whether what is taken is protected. The fixation tradition (idea 3) asks whether the weights are a “copy”. Memorization studies (idea 3) give the evidence, and compression theory (idea 1) gives two ways to read it. Fair use (Roots) decides what happens if the answer is “yes, some copying”.

Where others push back. Creators say “we’re just doing science” does not answer “you used my book without asking”. Developers say a model that benefits many people should not be blocked by rare edge cases.

What she leaves out. Licensing, opt-outs, payment for training data, and outputs that compete with authors.

Check yourself

  • Why do the two sides each think the other is missing the point?
  • Does “beneficial for humanity” appear in the four fair-use factors? Where might it enter?
  • What would Samuelson need to add for the clip to cover the output side?