# Amicus briefs and precedent in *In re: OpenAI, Inc. Copyright Infringement Litigation*

Data behind Tech Policy Press's analysis of the 56 amicus briefs that US District Judge Sidney H.
Stein accepted on Nov. 6, 2026, in *In re: OpenAI, Inc. Copyright Infringement Litigation*,
No. 1:25-md-03143 (S.D.N.Y.). The court granted 55 motions for leave in ECF 2496 and one more
in ECF 2498.

## Files

| File | One row per | What it holds |
|---|---|---|
| `openai-amicus-briefs.csv` | brief (56) | Filer, declared side, the rule step that set the side, and the verbatim evidence for it |
| `openai-amicus-citations.csv` | brief-case citation (798) | Every case citation extracted from every brief, with how many times the brief's text refers to the case |
| `openai-amicus-cases.csv` | case (23) | The cases cited by at least eight briefs: briefs and references by side, lean, CourtListener link |
| `openai-amicus-quotes.csv` | quote (385) | The audited quotes shown in the Tech Policy Press amicus tracker; an ellipsis (...) at either end marks a quote that starts or ends mid-sentence |
| `openai-amicus-citation-decisions.csv` | individual decision (29) | Every citation Claude added, excluded or removed after reading it in context, with the reason |
| `openai-amicus-data.json` | | All of the above in one file |
| `openai-amicus-briefs-pdfs.zip` | brief (56) | All 56 accepted proposed briefs as PDFs, named by ECF number and filer, plus `openai-amicus-briefs.csv` |
| `openai-amicus-methods.txt` | | This file |

## How each brief's side was set

Each brief is assigned the side it formally declares, applying these steps in order and stopping
at the first that names a side:

1. **Caption.** The caption of the proposed brief, or of the filing's cover, states a side,
   including "in support of neither party."
2. **Self-statement.** The brief states in its own words which party it supports.
3. **Relief.** The brief explicitly asks the court to grant or deny a party's motion.
4. **Motion.** The filer's own motion for leave to file the brief describes its side. Added
   Nov. 11, 2026, after a recheck of the briefs that declared no side: Michael Moritz's motion
   (ECF 2434) says his brief is "in support of neither party," and the Center for American Rights'
   motion (ECF 2306) calls its brief "an amicus brief supporting the plaintiffs." For these two,
   `side_evidence_quote` and `side_evidence_page` are from the motion, not the brief. The motions for ECF 2381, 2404 and 2408
   are not in CourtListener's RECAP archive and could not be checked; the memorandum supporting
   ECF 2417 (ECF 2430) says the amici take no position on how the motions should be decided.

Briefs that do none of these are "Declares no side," with a `leans` value showing which way their
argument points. The `side_evidence_quote` column holds the verbatim words that decided each case,
and `side_basis` names the step. `side_note` adds the filer's own disclaimer where one exists
(ECF 2417).

## How citations were found

1. **Text.** Each proposed brief was downloaded from CourtListener's RECAP archive and its text
   extracted page by page. One brief (ECF 2394) had a corrupted text layer and was OCR'd; its quotes
   were checked against page images. Appendices that are not part of the brief (CVs, an appended
   paper, a declaration) were excluded.
2. **Two independent extractions.** Case citations were extracted with
   [eyecite](https://github.com/freelawproject/eyecite), the Free Law Project's citation parser,
   once from each brief's table of authorities and once from its body text, with short-form
   citations resolved to the case they refer to. 51 of the 56 briefs have a table of authorities.
   `source` records which extraction found each citation (`toa+body`, `toa`, `body`), or `manual` for a citation Claude added after reading the brief.
3. **Parallel citations merged.** Different reporter citations of the same decision were merged.
   The Third Circuit's Thomson Reuters v. Ross decision appears as 2026 WL 2932557 and as 2026 U.S.
   App. LEXIS 29958, and under its docket number, No. 25-2153.
4. **Name cross-check.** For every case cited by at least eight briefs, each brief's text was also
   searched for the case's distinctive name, and the result compared with the extraction. The
   extraction never claimed a citation the name search could not find. The name search found 23
   additional candidates; Claude read each in context and decided it: 8 were real citations the extractor missed
   (typos, docket-number citations, short-form-only citations) and were added, and 15 were other
   cases or passing mentions and were excluded. See `openai-amicus-citation-decisions.csv`. Citations of cases cited by
   fewer than eight briefs were not cross-checked by name (`crosschecked` = False).
5. **Artifacts removed.** Six extraction artifacts were removed: line numbers on pleading paper that
   the parser read as citations, and three mistyped citations duplicating a correct one in the same
   brief.
6. **Case identity.** Each of the 23 cases was looked up in CourtListener's citation API. 20 resolved
   to an opinion (`verified_by` = CourtListener). Three recent decisions (Kadrey v. Meta, Bartz v.
   Anthropic, and the Third Circuit's Ross decision) are not yet in CourtListener's citation index and
   were verified against the identical entries in many briefs' tables of authorities.

## Briefs vs. references

Two different counts appear in the data:

- **Briefs citing** (`briefs_citing`, `plaintiff_side`, `defense_side`): how many briefs cite a case at
  least once. This is what the network map shows.
- **References** (`references` and the side columns in `openai-amicus-cases.csv`; `references`, `full_citations`,
  `short_citations` and `id_citations` in `openai-amicus-citations.csv`): how many times a brief's text refers to the
  case, counting full citations, short-form citations ("Campbell, 510 U.S. at 579") and "Id." references.
  eyecite links each "Id." to the citation immediately before it; a random sample of 12 "Id." links to the
  most-cited cases was checked against the briefs' text and all were correct. Tables of authorities are not
  counted. References are blank (and counted in `briefs_not_counted`) for the 12 brief-case citations the
  parser never matched in a brief's text: the 8 added after the name check and 4 found only in a table of
  authorities.

## Caveats

- **Citing is not relying.** A brief may cite a case to argue it does not apply. The data records
  that a brief cites a case, not how it uses it.
- **Lean** (`defense_share` in `openai-amicus-cases.csv`) is the share of a case's citing briefs, among those
  declaring a side, that support the defendants. It describes who cites the case, not what the case
  holds.
- The proposed briefs are those the court deemed filed; some are stamped with member-case or
  attachment docket numbers. Briefs are identified here by the MDL motion number the court used in
  its Nov. 6 order (`ecf`), and `brief_url` links to the document itself.

## Reproducing

The scripts are in the project's `precedent/` folder, along with the briefs pipeline one level up.
With Python 3, `pip install eyecite` and a free CourtListener API token:

1. `fetch.py` downloads the briefs and extracts the page text.
2. `precedent/extract.py` runs both citation extractions.
3. `precedent/crosscheck.py` runs the name cross-check and prints any disagreements for review.
4. `precedent/build_data.py` applies the reviewed decisions and writes the case data.
5. `precedent/build_release.py` writes the data files, and `make_uploads.py` gives them these names.

## Use of AI

This analysis was conducted with Anthropic's Claude Opus 5.5, an AI model, with prompts from
Tech Policy Press editors, who set the rules for categorizing the briefs. Claude:

- wrote and ran the code that downloaded the briefs, extracted their text and citations, and built
  these files;
- read each brief in full to record its caption, its stated side, its requested relief and the
  quotes in `openai-amicus-quotes.csv`;
- checked every displayed quote against an image of the cited page, using a separate Claude
  Opus 5.5 instance from the one that chose it, and re-verified each quote word for word by script;
- decided each ambiguous citation found by the name cross-check and recorded its decision and
  reason in `openai-amicus-citation-decisions.csv`. These citation decisions were not reviewed by an editor.

Every quote in `openai-amicus-quotes.csv` was verified verbatim against the brief's text by script, and every
count can be recomputed from `openai-amicus-citations.csv`.

A Tech Policy Press editor checked a sample of the quotes in the downloadable data against the
briefs and found each one verbatim. That review also found a quote that ended mid-sentence without
any mark; as a result,
every quote that starts or ends mid-sentence now carries an ellipsis at that end (14 of 385). No
other human review was done: the citation decisions in `openai-amicus-citation-decisions.csv` and the remaining
quotes were checked by Claude and by script only.

## Credit and license

Analysis by Tech Policy Press. Court filings are public records via
[CourtListener](https://www.courtlistener.com/docket/69879510/in-re-openai-inc-copyright-infringement-litigation/).
*CC by 4.0*