How summaries are made — and where they can fail.
Here’s exactly how the site works.
Where the opinions come from
All opinion text comes from CourtListener, a free public archive maintained by the nonprofit Free Law Project. Each opinion linked here is hosted by CourtListener and remains the authoritative source.
We pull new opinions filed in the U.S. District Court for the District of Minnesota daily and store a copy in our own database. The original opinion text is not modified.
One honest caveat: there is a lag between when the court issues an opinion and when it appears on CourtListener. The Free Law Project is a small nonprofit, and their indexing of district-court opinions can run anywhere from a few days to several weeks behind the court’s actual filings. Our pipeline checks for new opinions daily, but on any given day the most recent opinion we have may still be a few weeks old. The browse page shows the actual most-recent filing date so you can judge for yourself.
How summaries are generated
For each opinion, we send the full text to Anthropic’s Claude model with a structured prompt asking for three summaries of different lengths — one sentence, three paragraphs, and a detailed version — along with topic tags and a “who this affects” line.
The prompt tells the model to:
- Translate legal terminology into plain English.
- Never state a fact or holding that isn't in the opinion text.
- Flag its own uncertainty.
- Never offer legal advice.
The model can also leave a note for human reviewers when it notices something ambiguous.
Generated summaries go directly to the site without human review at this stage of the project. A review workflow is planned.
How summaries are checked
Before a summary is trusted, an automated verification system compares it against the source opinion. For each summary, a separate AI process re-reads the opinion and checks the summary’s factual claims — party names, case outcomes, dates, and holdings — against what the opinion actually says, flagging anything it can’t support.
When the checker finds a serious problem — a wrong party, a reversed outcome, a fabricated detail — the summary is automatically withheld from the site and marked “under review.” You’ll see a placeholder instead of the summary on those pages until a person can correct it. This happens before the page is served, so a summary the system flags as seriously wrong is not shown to you at all.
This checking is itself AI-based, so it isn’t perfect — it can miss errors, and it sometimes withholds summaries that are actually fine. It’s a safety net, not a guarantee. That’s why the next section matters, and why every page links to the original opinion.
What can go wrong
AI summaries of legal text can be wrong in ways that look right. A summary might:
- 01.Misstate which side won or what was decided.
- 02.Miss a key procedural detail (with vs. without prejudice).
- 03.Oversimplify the legal reasoning in a way that changes the meaning.
- 04.Mistakenly attribute a quote or holding.
- 05.Categorize a case under the wrong topic.
For this reason, always read the full opinion (linked from each summary page) before relying on anything here.
Reporting errors
If you find a summary that’s wrong or misleading, email corrections@courtexplained.org with the case ID and a brief description of the problem. We read every report.