Editing Openai/69408af5-1bf8-800f-b5dc-86cd01e07ec0 (section)

=== Assistant: Short answer: no, frontier labs are not simply ignoring infringement risk and scraping everything indiscriminately, but… ===
Short answer: no, frontier labs are not simply ignoring infringement risk and scraping everything indiscriminately, but they are also not limited to the same ruleset people intuitively imagine. What changed after the infringement claims is how data is sourced, processed, and justified, not whether models continue to scale.

Below is the clean breakdown.

==== Large post-2023 increases in token counts come primarily from: ====

===== - Models generate their own training material: paraphrases, explanations, step-by-step reasoning, code variants, question–answer pairs =====
* Synthetic data can be orders of magnitude larger than human-written corpora
* This data is legally clean because it is model-generated, not copied

This is now a major component of frontier training.

===== - The same underlying datasets are: - Re-tokenized - Reweighted - Re-mixed - Re-paraphrased - Re-sequenced =====
* A single document can yield many distinct training samples without adding new source material

Token count ≠ number of unique documents.

===== - Explicit licensing deals (publishers, news orgs, code repositories, image libraries) =====
* Proprietary internal corpora (Google, Microsoft, Amazon have massive first-party text, logs, docs, manuals)
* Paid human-created datasets at scale

These sources expanded substantially after lawsuits began.

==== Frontier labs are relying heavily on transformative fair use, specifically: ====
* Training ≠ storing or redistributing text
* Models do not retain retrievable copies of works
* Outputs are probabilistic transformations, not reproductions

This position has not been conclusively tested at Supreme Court level, but:
* Multiple lower-court rulings
* Decades of precedent from search engines, indexing, and ML currently favor this interpretation.

Labs are betting that training is legally analogous to reading, not copying.

==== Compared to early GPT-3 era scraping: ====
* Copyright filters are stricter
* Deduplication is aggressive
* Memorization tests are routine
* Regurgitation benchmarks are enforced
* Opt-out signals are increasingly honored

The irony: models today are less likely to verbatim reproduce copyrighted text than earlier ones, despite being larger.

==== Modern token counts include: ====
* Image tokens
* Audio tokens
* Video frame tokens
* Sensorial embeddings
* Structured metadata

These are often:
* Licensed
* First-party
* Synthetic
* Or weakly copyrightable

So “15T tokens” is not “15T copyrighted paragraphs.”

==== They are not reckless, but they are accepting uncertainty because: ====
* The upside of frontier capability is enormous
* The legal landscape is unsettled
* Waiting for perfect clarity means losing the race
* Precedent historically favors transformative technologies

This is similar to:
* Search engines indexing the web
* DVRs recording TV
* Web browsers caching pages
* Music sampling before clear licensing norms

In each case, companies moved first, courts caught up later.

==== Because data scale is no longer bottlenecked by raw copyrighted text. ====

The bottlenecks now are:
* Compute
* Energy
* Alignment
* Data quality, not quantity

Once synthetic + licensed pipelines matured, token counts could continue growing without proportional legal exposure.

==== - No, labs are not just scraping everything and hoping for the best ====
* Yes, they are relying on fair use, licensing, synthetic data, and transformation
* Token growth today does not mean proportional growth in copyrighted human writing
* The legal fight is real, but the technical strategy has already adapted around it

If you want, next we can:
* Separate copyright risk by data type
* Analyze whether fair use will actually hold up
* Compare AI training to music sampling law
* Or examine what a “worst case legal loss” would realistically force labs to do

Just say the direction.