CourtListener vs. LawDiver: Can Free Caselaw Data Power a Real Legal Research Product?
· LawDiver
CourtListener and the Hugging Face caselaw datasets are the best free legal data in existence — and they still are not a legal research product. Here is exactly what is missing: good-law verification, controllable search, non-case authorities, and real cite checking.
Every few weeks, someone building a legal application arrives at the same happy discovery: American case law is free. Not paywalled, not licensed, not gated behind a six-figure contract. It is sitting on a public S3 bucket, in bulk, in CSV, with a permissive public-domain mark, from courts that were deciding cases before there was a United States.
That discovery is real, and the people responsible for it deserve enormous credit. CourtListener, run by the nonprofit Free Law Project, is the root data source for a striking share of the legal-technology ecosystem — including products that will never admit it. The Caselaw Access Project corpus mirrored on Hugging Face put 6.7 million digitized decisions into the hands of anyone with a laptop.
So the question every builder eventually asks is fair, and we get asked it constantly:
If the cases are free, why does anyone pay for legal research?
The honest answer is that the cases were never the expensive part. This post is a candid, source-linked walk through what CourtListener and the Hugging Face datasets actually give you, where they stop, and what has to exist on top of them before a lawyer can safely file something. We build LawDiver, so we are not neutral. We have written the bias disclosure and a test protocol into the end of this article so you can check every claim yourself.
What CourtListener actually gives you (and it is a lot)
Let us be precise about the strengths, because vague praise followed by a sales pitch is not an argument.
Bulk case law from 1658 to the present. Free Law Project publishes complete bulk exports — courts, dockets, opinions, opinion clusters, citations, and parentheticals — as CSV snapshots. They are regenerated quarterly, on the last day of March, June, September, and December, which means the current public export runs through June 30, 2026. This is not a teaser dataset. It is the database.
A genuinely excellent API. The CourtListener REST API exposes essentially everything the site knows: opinions, clusters, dockets, courts, judges, oral arguments, RECAP federal filings, and alerts. It is well-documented, versioned, and as of May 2026 full API access is included with membership. If you need programmatic access to American case law, this is the front door.
The citation graph, in both directions. For a given opinion you can retrieve what it cites and what cites it, with citation depth counts and parentheticals — the one-line descriptions later courts used when citing the case. Free Law Project built this by analyzing more than 50 million citations spanning two centuries, using eyecite, the open-source citation parser they developed with Harvard's Library Innovation Lab. Eyecite is so good that much of the industry, ourselves included, builds on it.
A citation lookup and verification API. Since April 2024, CourtListener has offered a citation lookup endpoint that parses a block of text and tells you whether each citation resolves to a real case in their database — up to 250 citations per request. As a first-line guardrail against a hallucinated cite, it works.
And now, semantic search. This is the part most comparison articles get wrong because they were written a year ago. Free Law Project shipped a Semantic Search API in November 2025 and brought semantic search to the CourtListener website in May 2026, powered by a domain-adapted embedding model they trained themselves, with hybrid keyword-plus-meaning queries. CourtListener is no longer keyword-only. Anyone still telling you it is has not looked since last year.
That is a formidable stack, offered free, by a nonprofit, for the public good. Use it. Fund it.
Now the hard part.
What the Hugging Face caselaw datasets give you — and exactly where they stop
The Hugging Face mirrors of the Caselaw Access Project are the single fastest way to get bulk American case law onto a GPU. The free-law/Caselaw_Access_Project dataset holds roughly 6.7 million state and federal decisions, CC0-licensed, derived from Harvard Law School Library's digitization of nearly 40 million pages from 39,796 volumes. There are companion embedding and FAISS-index datasets, and the common-pile/caselaw_access_project build combines CAP with roughly 900,000 CourtListener opinions into about 6.9 million documents and 78 GB of text.
Here is the problem, and it is not subtle.
The Caselaw Access Project corpus ends in 2018.
Harvard's scanning ran from 2013 to 2018, covering every volume of official, book-published U.S. case law the library held — from 1658 through the last volumes received in 2018. It is a closed historical corpus, not a living one. CAP handed off ongoing collection to CourtListener in March 2024 and retired its own API.
So a model or product trained on the Hugging Face CAP snapshot alone is reasoning about American law as it stood roughly eight years ago. Eight years is not a rounding error in law. It is Dobbs. It is Bruen. It is Loper Bright ending Chevron deference and unsettling four decades of administrative law. A research tool that has never heard of the case that overruled your case is not a slow research tool. It is a wrong one.
For pretraining, benchmarking, and NLP research, the Hugging Face datasets are excellent. As the retrieval corpus behind a product a lawyer files from, a 2018 cutoff is disqualifying on its own — before we even get to the deeper issue that both CAP and CourtListener share.
The deeper issue: retrieval is not research
Here is the thing that takes builders longest to internalize.
Finding the case is maybe forty percent of legal research. Everything expensive happens after you find it.
A first-year associate who hands a partner a beautifully on-point case that was overruled in 2021 has not done forty percent of the job. They have done negative work — they have created risk, and a partner now has to check behind them. The entire value of a commercial research platform lives in the layer that sits between "here is a relevant case" and "you may safely cite this."
There are five pieces to that layer. Free data ships with none of them fully in place.
1. Good-law verification (what old-timers call Shepardizing)
This is the big one, and everything else is a footnote next to it.
When you pull a case from CourtListener, you learn what cites it. You do not learn how it was treated. Followed? Distinguished? Criticized? Questioned? Limited on other grounds? Overruled? CourtListener shows you the citing opinions and their parentheticals and leaves the interpretation to you, which is to say it leaves you to read them.
To be scrupulously fair: Free Law Project knows this and is actively building a citator. They published a progress report in May 2025 on an AI-assisted approach and a status update in September 2025 on their expert-annotation project, where volunteer attorneys label treatments with at least two reviewers per citation. It is serious, careful work by people we admire, and when it lands the free ecosystem will be meaningfully better.
But it has not landed. Today, if you build on CourtListener bulk data or the Hugging Face datasets, your product has no good-law signal. Your users will assume it does — that is the deadly part. Silence reads as approval. A search result with no red flag looks exactly like a search result that passed a check nobody ran.
LawDiver ships a GoodLaw status on every case, inline, in every search result and every API response: status, negative treatments and their count, the basis for the determination, and — critically — a separate unknown flag published apart from negative, so an application can never mistake not yet determined for cleared. Bad law is flagged, never hidden; an overruled case is still the case the researcher went looking for, and a goodLawOnly filter exists but stays off unless you ask for it.
2. Search you can actually control
CourtListener now has keyword search, semantic search, and a hybrid mode. That is real progress and we will say so plainly.
What it does not have is control and comparison. A lawyer hunting the perfect block quote for a brief does not want one blended relevance list. They want to run the same idea five ways and look at the differences, because the differences are the research.
LawDiver runs four independent retrieval engines — citation, case-name, boolean keyword with proximity connectors (/s, /p, w/n), and semantic vector search — plus weighted fusion over them, exposed as five selectable search modes and an auto router that plans engines from the query itself rather than stamping it with a single label. A citation goes to deterministic reporter-key lookup. A half-remembered caption goes to the name engine. A question phrased in words no opinion uses goes to the vector engine. Miranda v. Arizona qualified immunity /s clearly established runs the name engine and the boolean engine simultaneously and lets fusion decide which mattered.
You can pin a mode. You can compare modes side by side on the same query. Jurisdiction and date phrases are lifted out of the query text into filters that run inside every engine — because leaving "5th Circuit" in the search string gets you opinions that discuss the Fifth Circuit, which is not what anyone meant.
For the builder-facing version of this stack, see the caselaw API for legal AI builders.
3. Citing authority that is not a case
CourtListener's citation graph does a great job on case-to-case citation. But an opinion is not only cited by other opinions, and it does not only cite other opinions.
Real research follows a statute to the cases construing it, a regulation to the decisions reviewing it, and a rule of civil procedure to the opinions applying it. LawDiver's citation extraction classifies authorities by kind — case, statute, regulation, and court rule — so the graph includes the sources lawyers actually cite in the sections of a brief where cases are the minority. A table of authorities that lists only cases is not a table of authorities.
4. Cite checking as a document workflow
CourtListener's citation lookup answers one question well: does this citation string resolve to a real case?
That is the first question a cite checker asks, not the last. The rest of them: Does the case name match the reporter cite, or is this the classic hallucination shape where a real citation has been welded to a real-but-different case name? Is the year right? Is the court right? Is the Bluebook form correct? Are the parallel cites right? Did the quoted language actually appear in the opinion, at the pinpoint page given? And is the thing still good law?
LawDiver's CiteChecker takes a single citation, a batch of up to fifty, or an uploaded PDF or Word brief, and returns a report on every citation in the document: resolution with a confidence score that reaches 1.0 only on a deterministic reporter-key match, a name_mismatch verdict for the hallucinated-cite shape, corrected Bluebook form assembled from stored metadata, and good-law status attached to each authority. That is a workflow a lawyer runs before filing, not an existence check an application runs in the background.
If you have been told an LLM can do this, read why LLMs cannot cite-check your brief.
5. Currency measured in minutes, not quarters
Quarterly bulk snapshots are perfectly correct for research, archiving, and model training. They are not how a research platform stays current. If the last export ran on June 30 and the case that overrules yours came down on July 12, your product is confidently wrong for up to three months — and it will not know it, because silence reads as approval.
LawDiver harvests continuously from court sites, public feeds, and government publications so yesterday's opinion is in the index today — the same currency model behind free caselaw research on CaseDiver.
Side by side
| Capability | Hugging Face CAP datasets | CourtListener / Free Law Project | LawDiver |
|---|---|---|---|
| Bulk case law download | Yes, CC0, ~6.7M cases | Yes, quarterly CSV exports, public domain | Not a bulk-data publisher |
| Coverage end date | 2018 (closed corpus) | Current, snapshotted quarterly | Continuously updated |
| Public API | Dataset download only | Yes — broad and well-documented | Yes — search, retrieval, cite check |
| Keyword / boolean search | Bring your own | Yes | Yes, with proximity connectors |
| Semantic search | Embeddings provided, no product | Yes (API 2025, site 2026) | Yes, plus hybrid fusion |
| User-selectable engines & comparison | No | No | Yes — five modes plus auto routing |
| Cases citing a given case | Limited | Yes, with depth and parentheticals | Yes |
| Statutes, regulations, court rules in the graph | No | Limited | Yes |
| Good-law / treatment signal | No | In development, not shipped | Yes — GoodLaw on every result |
| Document-level cite checking | No | Citation existence lookup only | Yes — PDF/DOCX brief reports |
| Cost | Free | Free / membership | Commercial |
So: can you build a legal app on CourtListener?
As a data source, absolutely — and you probably should. It is the best free legal corpus in the world, maintained by people doing the work for the right reasons. If you are training a model, running research, building a citation tool, or standing up a public-interest service, start there.
As a finished product, no. Not because anything is wrong with it, but because it was never trying to be one. Free Law Project's mission is open access to legal data. Deciding whether a case is still good law is a different discipline with different economics — annotation, editorial judgment, and continuous maintenance — and they are only now, carefully, beginning to build it.
The gap between "here are the cases" and "you may cite this" is where Westlaw's KeyCite and Lexis's Shepard's have lived for a century, and it is where LawDiver lives now. That gap is not a feature list. It is the entire reason legal research is a profession and not a search box.
We are biased. Here is how to test us anyway.
We build LawDiver. Treat everything above as an interested party's argument and go verify it. The good news is that this is one of the easiest product categories in software to benchmark, because the answers are knowable.
A protocol you can run in an afternoon:
- Build a ground-truth set. Pull 25 cases you know were overruled, abrogated, or superseded, and 25 you know remain good law. Mix in a few limited-on-other-grounds cases — that is where citators actually differ.
- Run every candidate. Westlaw KeyCite, Lexis Shepard's, LawDiver GoodLaw, and any newer entrant. Record: correct flag, missed negative treatment, false alarm, and how many clicks and seconds it took to reach an answer.
- Score the misses asymmetrically. A false negative — telling a lawyer a dead case is alive — is not the same class of error as a false positive. Weight it accordingly.
- Cite-check a real brief. Take a filed brief, corrupt ten citations in realistic ways (transposed volume and page, right cite with the wrong case name, wrong year, quote moved to the wrong pinpoint), and see which tools catch which. Include a hallucinated cite of the shape an LLM actually produces.
- Search the same question five ways. One boolean query, one natural-language question, one citation, one half-remembered case name, one mixed query. Judge whether you could steer the tool when the first answer was wrong — which is what research actually feels like.
Publish what you find, including where we lose. A benchmark that only the vendor can run is marketing. A benchmark you can run is evidence.
Frequently asked questions
Is CourtListener free to use?
Yes. CourtListener is operated by the Free Law Project, a 501(c)(3) nonprofit. The site, the bulk data, and the API are free, with full API access included in membership as of May 2026. The bulk data carries a public domain mark.
Does CourtListener tell you if a case is still good law?
Not yet. CourtListener shows citing opinions, citation depth, and parentheticals, but it does not classify treatment as followed, distinguished, criticized, or overruled. Free Law Project announced a citator project in 2025 combining AI with attorney annotation, but as of September 2026 there is no shipped good-law signal. You must read the citing cases yourself.
Does CourtListener have semantic search?
Yes. Free Law Project launched a semantic search API in November 2025 and brought semantic search to the CourtListener website in May 2026, using a domain-adapted embedding model, with hybrid keyword-plus-semantic queries supported.
What year does the Hugging Face Caselaw Access Project dataset end?
- Harvard Law School Library digitized official published U.S. case law from 1658 through the last volumes received in 2018, and the project handed ongoing collection to CourtListener in 2024. The Hugging Face mirrors reflect that closed corpus and do not include cases decided after 2018.
Can I build a legal research product on CourtListener data?
You can build on it, and many products do. You cannot ship it as-is to lawyers, because it lacks good-law verification and document-level cite checking. Whatever you build must supply the validation layer between "relevant case" and "safe to cite."
What is Shepardizing?
Shepardizing is checking whether a case is still good law by reviewing how later courts treated it — following, distinguishing, criticizing, limiting, or overruling it. The term comes from Shepard's Citations, now part of LexisNexis. Westlaw's equivalent is KeyCite; LawDiver's is GoodLaw.
How is LawDiver different from CourtListener?
CourtListener is an open-data source; LawDiver is a research platform built for filing. The differences that matter: GoodLaw status on every result, five user-selectable search modes with side-by-side comparison, statutes, regulations, and court rules in the citation graph, and document-level cite checking that reads an entire brief.
Is LawDiver a CourtListener alternative?
For bulk open data and public-interest access, CourtListener remains the right starting point. For lawyers and builders who need good-law verification, controllable multi-engine search, and cite checking before filing, LawDiver is the product layer on top of what free data alone cannot do.
Sources
- Free Law Project — Semantic Search Is Now on CourtListener (May 2026)
- Free Law Project — Semantic Search API Now Live (November 2025)
- Free Law Project — Building a Citator with AI, A Progress Report (May 2025)
- Free Law Project — Citator Expert Annotation Project: A Status Update (September 2025)
- Free Law Project — Combat Hallucinations and Look Up Citations with our New API (April 2024)
- CourtListener — Bulk Legal Data documentation
- CourtListener — Citation Lookup and Verification API
- Harvard Library Innovation Lab — Transitions for the Caselaw Access Project (March 2024)
- Hugging Face — free-law/Caselaw_Access_Project
- Hugging Face — common-pile/caselaw_access_project
Last verified September 9, 2026. Free Law Project ships quickly; if their citator has launched since publication, this article is out of date and we will update it.
Try the product layer yourself
Free data gets you the cases. The rest of research is verifying them, searching them under your control, and checking every cite in the brief before you file.
Search caselaw on CaseDiver · CiteChecker · Caselaw API docs · Sources and methods