---
title: "Hallucination de l'IA juridique : la fausse citation qui passe tous les contrôles"
url: https://www.haqq.ai/fr/blog/legal-ai-hallucination-fake-citation
locale: fr
type: blog
description: "Nous avons fait usurper à un modèle de pointe l'identité d'une base de données juridique. Il a refusé d'inventer de fausses affaires, puis a cité une loi réelle pour la mauvaise raison, l'hallucination qui survit à la relecture d'un avocat."
category: "ai-legal-tech"
keywords: "legal AI hallucination, AI legal citation hallucination, fake legal citations, do AI legal research tools hallucinate, name cite mismatch, retrieval grounded legal AI, AI fake case citations"
published: 2026-06-30
english_version: https://www.haqq.ai/blog/legal-ai-hallucination-fake-citation
source: src/data/blogPosts.ts
generator: scripts/generate-page-markdown.ts
---

# Hallucination de l'IA juridique : la fausse citation qui passe tous les contrôles

Nous avons fait usurper à un modèle de pointe l'identité d'une base de données juridique. Il a refusé d'inventer de fausses affaires, puis a cité une loi réelle pour la mauvaise raison, l'hallucination qui survit à la relecture d'un avocat.

> **Note:** TL;DR: We turned a frontier model (Opus 4.8) into a fake legal database and probed where it fabricates. It refused to invent citations that do not exist, 4 out of 4 impossible references came back not found. But on a plausible query it fabricated the search itself: a ranked result list with a made-up relevance score and a real, clickable citation pointing to the wrong law. The dangerous hallucination is not the fake case you can catch. It is the real citation attached to the wrong content, and the model insisting it looked something up when it did not.

## The number everyone quotes, and the one they miss

If you have shopped for legal AI in the last year, you know the scary statistic. A 2024 [Stanford study](https://arxiv.org/pdf/2405.20362) measured hallucination rates of 43% for GPT-4, 33% for Westlaw's AI research, and 17% for Lexis+. A separate database now tracks [over 1,000 court cases](https://www.llrx.com/2026/02/what-the-science-says-about-hallucinations-in-legal-research/) where lawyers filed AI-invented citations. Some have been fined.

**Hallucination rates in legal research tools**

Stanford, 2024, share of responses containing a hallucination

- GPT-4 (general LLM): 43%
- Westlaw AI research: 33%
- Lexis+ AI: 17%

_Retrieval-grounded legal tools beat general models, but the errors that remain get subtler, like a real citation attached to the wrong law._

The advice that follows is always the same: verify every citation. True, and nearly useless, because it assumes hallucinations look like obvious fakes. The ones that get lawyers sanctioned do not. So we ran an experiment to see what they actually look like.

## The experiment: make the model be the library

Most hallucination tests ask a model a legal question and grade the answer. We did something stranger. We used an open-source tool, [world-model-harness](https://github.com/experientiallabs/world-model-harness), to make Opus 4.8 impersonate the legal database itself, the search engine, not the lawyer.

Here is the setup in plain terms. We recorded real legal-research sessions against a genuine multi-jurisdiction database, statutes and case law across the EU, France, and Germany. Then we trained the model to play that database: you send a query, it returns what it thinks the database would return, search hits, document text, citations. A real database retrieves. Ours predicts. The gap between the two is the hallucination, isolated under a microscope.

> **Note:** A confession from the build: our first run showed almost 100% citation accuracy and we nearly celebrated. Then we realized we were testing the cache, not the model, we had fed it queries it had already seen, so it simply echoed the real answers back. Embarrassing for an afternoon, instructive forever: the only honest test of a legal AI is on questions it has not seen before.

## Finding 1: it will not invent a citation from nothing

We fed the fake library four references that do not exist anywhere: a made-up EU regulation, a fabricated French Code civil article, a nonsense CELEX document number, and a fake German constitutional case.

All four times, it answered the way a real database would: resolved false, not found. It did not invent a plausible case to fill the gap. This is genuinely good news and worth saying out loud. Frontier models in 2026 are far better calibrated than the they-just-make-things-up reputation suggests. Asked for something impossible, Opus 4.8 declined. If that were the whole story, you could relax. It is not.

## Finding 2: it fabricates the search itself

Then we asked it something plausible but unseen, about the EU AI Act's rules for high-risk systems. The model produced a complete, ranked database response: multiple hits, official-journal dates, a relevance score of 0.71243286, all framed as if retrieved from EUR-Lex. None of that retrieval happened. There was no index, no search, no score. It manufactured the act of looking something up. And the citations it returned are the masterclass in why this matters:

- CELEX 32024R1689, real and correct. That genuinely is the EU AI Act. Impressive recall.
- CELEX 32024R0900, real, clickable, and wrong. That citation resolves perfectly on EUR-Lex, to Regulation (EU) 2024/900 on the transparency of political advertising. It has nothing to do with AI conformity.

That second one is the whole point. It is not a fake citation, a fake citation is easy to catch because it does not resolve. This is a real citation attached to the wrong law. Citation-checkers call it a name/cite mismatch. You click it, an official EU regulation loads, everything looks legitimate, and you have just grounded your argument in advertising law while thinking it is AI law. Every link resolves. Nothing looks wrong.

> **Note:** Honest caveat: this is a sharp proof-of-concept, not a benchmark, a small corpus and one fabrication caught across a handful of probes. We are not claiming a rate. We are showing you a shape.

## Why this is the failure that survives review

Go back to the standard advice: verify every citation. Now watch it fail. The lawyer checks the AI Act cite, real. Checks the next one, also real, loads fine. Verification passed, because verification usually means does this link work, and every link worked. The error is not whether the citation exists. It is whether it says what the model claimed, a much harder thing to catch, and exactly the kind of mistake that ends up in a filing.

The deeper lesson is about where the lie lives. We instinctively worry the model will get the law wrong. The real risk is that it gets the retrieval wrong, that the sentence I searched the database and here is what I found is itself the hallucination, complete with a fake confidence score to sell it.

## HAQQ's take: the lookup is the hallucination

This is the reason we ground HAQQ's answers in real retrieval instead of asking a model to recall the law from memory. Not because models are dumb, this one refused every impossible citation and nailed the AI Act's number. But a model asked to be the source will eventually perform the retrieval it did not do. The only fix is to make the retrieval real: pull the actual document from an actual index, then verify the citation resolves to the case the model named, not just to a case.

A buyer's question follows from this. Do not ask a legal-AI vendor whether they hallucinate, everyone says no. Ask: when you show me a citation, did you retrieve that exact document, and do you check that the case name matches the cite? The name/cite mismatch is the tell. A tool that only checks whether citations are well-formed will wave 32024R0900 straight through.

## Key takeaways

- A well-calibrated frontier model will not invent citations from thin air, it refused 4 of 4 impossible references.
- It will fabricate the retrieval itself on plausible queries, including a fake relevance score and a real citation pointing to the wrong law.
- The dangerous hallucination is the one that passes verification, every link resolves, only the meaning is wrong.
- Ground legal AI in real retrieval plus citation-matching, and judge vendors on whether they check that a cite says what they claim, not just that it exists.

## Sources and further reading

- [Stanford, Assessing the Reliability of Leading AI Legal Research Tools (2024)](https://arxiv.org/pdf/2405.20362)
- [What the Science Says About Hallucinations in Legal Research (LLRX)](https://www.llrx.com/2026/02/what-the-science-says-about-hallucinations-in-legal-research/)
- [world-model-harness, the open-source tool we used](https://github.com/experientiallabs/world-model-harness)

## Questions fréquentes

### Les outils de recherche juridique par IA hallucinent-ils encore en 2026 ?

Oui, mais le taux dépend fortement de l'architecture. Les modèles généralistes hallucinent le plus ; les outils qui ancrent leurs réponses dans de vrais documents (génération augmentée par la recherche, RAG) hallucinent beaucoup moins. Une étude de Stanford de 2024 a mesuré 43 % pour GPT-4 contre 17 à 33 % pour les outils juridiques fondés sur la recherche documentaire. Les erreurs qui subsistent sont plus subtiles, comme une vraie citation rattachée à la mauvaise loi.

### Qu'est-ce qu'une discordance nom/citation ?

C'est lorsqu'une citation est réelle et renvoie à un document existant, mais à une affaire ou une loi différente de celle qui est nommée. Elle passe une vérification naïve du type « ce lien fonctionne-t-il ? », ce qui la rend plus dangereuse qu'une citation manifestement fausse, car rien ne paraît anormal en surface.

### La génération augmentée par la recherche (RAG) suffit-elle à empêcher les hallucinations ?

Elle les réduit considérablement mais n'a rien de magique. Le RAG peut encore faire remonter un vrai document et l'attribuer à tort. L'étape qui compte est la vérification : confirmer que la citation récupérée correspond à l'affirmation, et pas seulement qu'elle existe.

### Comment un avocat doit-il vérifier les citations générées par l'IA ?

Vérifiez que chaque citation renvoie à un document et que ce document dit réellement ce que l'IA affirme : la solution retenue, l'objet du règlement, le numéro d'article. La discordance se cache dans la seconde étape, pas dans la première.

### Peut-on faire confiance à l'IA pour la recherche juridique ?

Seulement avec une vérification et la bonne architecture. Traitez le résultat de l'IA comme celui d'un chercheur rapide mais junior : utile pour la vitesse, jamais l'autorité finale. Privilégiez les outils qui puisent dans de vraies sources et vérifiez que les citations correspondent à l'affirmation avant de vous y fier.

---

Canonical page: https://www.haqq.ai/fr/blog/legal-ai-hallucination-fake-citation
Whole-site context for agents: https://www.haqq.ai/llms.txt · https://www.haqq.ai/agents.md

Generated at build time from the site's own content data. Do not edit by hand.
