Jurisvio Editorial Team ·

Citing employment case law with AI: how to avoid made-up decisions

An AI can invent decisions. The method: decisions retrieved from Judilibre, citation built by code, two checks, and their limits.

In brief

To avoid made-up decisions, an AI should cite only decisions retrieved from a database (here Judilibre), with a citation built by code, followed by a check that each decision exists and that the statement matches its extract. These checks do not replace reading the decision.

The problem

A language model writes plausible text. Asked to cite case law from memory, it can produce a well-formed reference (court, date, appeal number) that matches no decision, or a real decision to which it attributes something else. In a legal document, such a citation engages the lawyer’s responsibility.

The method: the AI cites only what the code has retrieved

  • The code queries Judilibre, the Cour de cassation’s API, in the Social Chamber, favouring published decisions, then in the courts of appeal. The Social Chamber’s code is read from Judilibre’s taxonomy.
  • The citation is built by the code from the date and the appeal or case number returned by Judilibre: “Cass. soc., [date], n° [number]”.
  • The AI receives only this closed list, with the summary or an extract of the reasoning of each decision. Court of appeal decisions are presented to it as decisions on their facts.
  • If no decision has been retrieved, the AI is instructed to cite none and to write “case law to be researched”.

Two checks after drafting

First check: the code rereads the document and checks every decision cited. A number that is not in the retrieved list, a date that does not match the number, or a dated citation that matches no decision in the list is highlighted in the document with the note “unverified decision”.

Second check: for each statement attributed to a decision, a second pass of the model compares the text of the document with the decision’s extract. Statements that the extract does not support, or supports only in part, are highlighted “to check”. This second opinion is automatic: it can be wrong.

What these checks do not guarantee

  • That the interpretation is correct: the consistency check is a second opinion, not a guarantee. The lawyer reads the decision; a link to read it is provided for the Cour de cassation.
  • That the decision is the most relevant or the most recent: selection depends on the search and the legal question asked.
  • That the ruling has not been called into question since: a change of case law is found by reading.
  • Completeness: in the test (sandbox) environment, Judilibre may return incomplete results; the tool says so.

What you need to use it

An access key to the Judilibre API, which the firm enters in Jurisvio’s integrations. Without a key, the document is drafted without case law and any decision cited anyway is flagged.

Try drafts with checked case law

Submissions for the employee or the employer with decisions retrieved from Judilibre and a citation check. 30-day free trial.

Related Articles

Frequently asked questions

Can an AI invent a court decision?
Yes: a language model can produce a well-formed reference that matches no decision. That is why it should only be allowed to cite decisions retrieved from a database.
What is Judilibre?
The Cour de cassation’s API giving access to court decisions. Jurisvio queries it in the Social Chamber, then in the courts of appeal; an access key is required.
Does the consistency check guarantee a citation is accurate?
No. It is an automatic second opinion that can be wrong: the lawyer reads the decision before any use.

See also: Jurisvio for French employment lawyers · Choosing software for employment lawyers

    Employment case law and AI: avoiding made-up citations | Jurisvio