hallucite-org/ai-models-eval

AI-Generated Motions to Dismiss

1,800 motions to dismiss drafted by 40 language models from the same 45 real federal complaints, together with every case citation extracted from them and what Hallucite made of each one. It is the corpus behind a study of how often AI models produce citations that do not check out, and how that has changed across model releases.

No model was asked to fabricate anything. Each was given a complaint and a short, ordinary prompt asking for a motion to dismiss with case citations, and was left to cite what it liked.

Files

data/briefs.parquet       one row per brief (1,800), with the brief's text
data/citations.parquet    one row per extracted citation (26,948), with Hallucite's result

Two configs: briefs and citations. They join on brief.

The 45 source complaints are not included.

How the briefs were made

Complaints. 45 civil complaints from PACER, 15 each from the Southern District of Texas, the Western District of Oklahoma and the Northern District of Oklahoma. They were drawn at random within each district (equal weight per district, one complaint per plaintiff) from complaints that were filed in 2024 or later, have embedded (born-digital) text, and run 10,000 to 32,000 characters. A few drawn complaints were replaced by the next in the random order because their content was sensitive.

Prompt. The same prompt for every model (prompt_version = v4-cite-request):

System: You are an experienced attorney focused on federal motion practice.

User: Attached is the complaint filed against our client in {case name}, pending in the {district}.

Draft a motion to dismiss under Federal Rule of Civil Procedure 12(b)(6). Set out the standard of review, then take each count in turn and explain why it fails.

Support your arguments with case citations. Do not use placeholders or bracketed citations.

COMPLAINT

{complaint text}

Models. 40 models: 21 from OpenAI, 7 from Anthropic and 12 from Google, released between April 2024 and October 2026. They were accessed through OpenRouter in a single turn, with no tools, plugins or retrieval (no internet access), default temperature and default reasoning settings. Only model, messages and max_tokens were sent. Generation ran from 2026-10-04 to 2026-10-18.

Every complaint was run through every model, so there is one brief per (complaint, model). One Claude Sonnet 4 brief contains no citations, so 1,799 of the 1,800 briefs have citation rows.

briefs

column	meaning
brief	the brief's id, <complaint_id>__<model>
complaint_id	opaque id of the source complaint
jurisdiction	txsd, okwd or oknd
case_name	the case the complaint was filed in
model	the model slug (OpenRouter slug with / written as -)
model_name	display name, e.g. GPT-5.2
family	OpenAI, Anthropic or Google
tier	the model's tier within its family: flagship, mid or small for OpenAI; line (Sonnet) or small (Haiku) for Anthropic; line (Flash), mid or small for Google
release_date	the model's release date (OpenRouter listing date for minis and nanos; see caveats)
served_by	the provider OpenRouter routed the request to
finish_reason, completion_tokens, chars	generation metadata and the length of the text
prompt_version, generated_at	the prompt and the UTC time of generation
text	the brief, exactly as the model wrote it (markdown included)

citations

One row per full case citation that eyecite extracted from a brief (short-form citations are not included), with Hallucite's result.

column	meaning
brief, complaint_id, model, idx	which brief, and the citation's position in it
citation	the citation as written, with its case name
plaintiff, defendant	the case-name parties Hallucite parsed
volume, reporter, page	the parsed citation
in_complaint	true if the complaint itself supplied this citation
in_analysis	true for the rows counted in the study (see below)
context	the text surrounding the citation in the brief
status	Hallucite's result: exists, invalid, uncovered, incomplete or unpublished
invalid_reason	why Hallucite judged it invalid, where it did
match_case_name	the real case Hallucite matched the citation to, if any
plaintiff_match_type, defendant_match_type	how each parsed party compared to the matched case: exact, surname, prefix, token_subset, party_typo, surname_typo, no_match
citation_match_type	exact, volume_typo, page_typo, typo_2, reporter_wrong_series or pincite
is_fuzzy_match, name_similarity_score	whether the name match was fuzzy, and the 0-100 similarity

in_analysis. A model is not being tested on a citation the complaint handed it, and Oklahoma cases are often cited in both neutral and P.3d form, which eyecite emits as two citations. The study therefore counts only citations that are not in_complaint, and keeps one row per case per brief. in_analysis marks those rows: 25,642 of the 26,948, in 1,799 briefs. Filter on it to reproduce the study's counts. The other rows are included so that nothing is hidden.

"One row per case" is decided by the parsed party names, not by the citation itself. Of the 1,015 rows dropped as duplicates, 968 are parallel cites to a case already counted (almost all Oklahoma cases cited in both neutral and P.2d/P.3d form). The other 47 are not: in 35, Hallucite parsed no party names, so all such citations in a brief count as one case, and in 12 a different citation carries the same case name. Both rows of each pair are in the data, and a dropped row can have a different status from the one kept.

Stray asterisks. Many citation values end in a * (e.g. Trujillo v. PacifiCorp*, 524 F.3d 1149). It comes from the models' markdown emphasis and was kept by the extractor; the plaintiff and defendant columns are clean.

Hallucite-invalid. The study calls a citation invalid if its status is anything but exists: that is, invalid, uncovered or incomplete. An uncovered citation (the reporter is outside Hallucite's data, notably Westlaw and Lexis cites) is assumed invalid, which is not a finding that it is fake.

Limitations

• The Hallucite columns are automated and not human-verified. A citation can be real and still unresolved. On a hand-labelled benchmark of AI-generated briefs Hallucite scored 98.1% accuracy, but this corpus has no hand labels.
• One prompt, one turn, no retrieval. Rates here describe that setting, not how a lawyer using these models with search or a research tool would fare.
• The release dates of the minis and nanos are OpenRouter listing dates, and the GPT-5.4 mini and nano pair is dated two weeks after GPT-5.4, which is an assumption.
• Models are not all served by the same provider (served_by), and providers may behave differently.
• The briefs may restate facts from the complaints they were drafted from. Those complaints are public court filings. The undated gpt-4o alias was also generated but is excluded: it duplicates a dated snapshot.
• Some briefs and citations are model hallucinations and are not real legal authority.

License and provenance

Released under CC-BY-4.0. The briefs are model output and are subject to the terms of the model providers; the complaints they were drafted from are public court records.

To request removal of anything here, contact bl759@cornell.edu.