> Bron: https://neuralex.nl/en/blog/ai-en-jurisprudentie
> AI can search case law without inventing judgments when answers are grounded in real sources. How RAG, source links and verification rules reduce hallucinations in legal research.

[Back to Insights](/en/blog)

Legal ·23 August 2026 ·6 min read

# Can AI search case law without making things up?

AI can search case law without inventing judgments when answers are grounded in real sources. How RAG, source links and verification rules reduce hallucinations in legal research.

Yes. AI can search case law without having to invent decisions, but only when the system is designed to ground its answers in real, retrieved sources. A language model that answers directly from its internal model knowledge can produce convincing but non-existent judgments, citations or ECLI numbers.

Legal research therefore requires a different workflow: retrieve relevant case law from a trusted source first, and only then allow the language model to summarise, compare or answer questions. This approach is commonly known as [retrieval-augmented generation, or RAG](/en/blog/rag-uitgelegd).

## Why a language model can invent case law

A language model is not a legal database. It generates text by predicting which words are likely to follow one another based on patterns learned during training.

That makes it useful for explanation, summarisation and text generation. It does not mean the model automatically knows whether a particular judgment actually exists.

If a user asks for relevant case law on a legal topic, a language model may generate an answer that looks as though it came from a legal database. It can produce a plausible ECLI number, date, court name or even a quotation without there being a real judgment behind it.

This is known as hallucination: the system produces information that sounds credible but is not supported by a source.

The problem is particularly serious in legal information because an incorrect answer may have exactly the same form as a correct one. A fabricated reference can look structurally identical to a genuine legal citation.

## How RAG makes case law research more reliable

Retrieval-augmented generation changes the order in which the AI system works. Instead of immediately generating an answer, the system first searches for relevant documents.

For case law research, a simplified workflow may look like this:

1.  The user submits a legal research question.
2.  The system searches a collection of real judgments.
3.  The most relevant documents and passages are retrieved.
4.  The language model receives those sources together with the question.
5.  The model formulates an answer based on the retrieved material.

The model therefore does not need to invent which judgment might be relevant. It works with documents that actually exist in the underlying source collection.

For the user, the interaction may still look like an ordinary chat interface. Behind the scenes, however, document retrieval takes place before the model generates its answer.

## How judgments are retrieved technically

A RAG system can make legal documents searchable in several ways. Traditional search methods may look for exact words, phrases or metadata.

Modern retrieval systems can also search by meaning. Text is often converted into numerical representations known as embeddings.

This means a search query does not have to contain exactly the same wording as the judgment. The system can also retrieve passages that are semantically similar to the user's question.

For case law, metadata can also be used, such as:

-   court or judicial body;
-   date;
-   area of law;
-   ECLI number;
-   topic or keyword;
-   document type.

A combination of semantic search and traditional filtering is often more useful than relying on a single retrieval method. The objective is not simply to return many documents, but to retrieve sources that are relevant enough to support the final answer.

## RAG does not automatically eliminate hallucinations

Connecting a language model to a legal database does not automatically make the system error-free.

Even with RAG, the model can misinterpret information, summarise passages too broadly or draw conclusions that are not explicitly supported by the source. The retrieval component may also fail to find relevant case law.

The architecture must therefore be designed not only for useful answers, but also for verifiable answers.

A key principle is that the source should remain authoritative. The language model should not freely fill in missing information when the retrieval layer fails to provide sufficient evidence.

## Make source grounding mandatory

One practical safeguard is to allow the system to make substantive legal claims only when there is a retrieved source supporting them.

For example, each conclusion in the answer can refer to the judgment on which it is based. If the system cannot find a suitable source, it should state that explicitly instead of generating a likely answer anyway.

This changes the design principle from "give the best possible answer" to "only answer to the extent that the available source material supports the answer".

That distinction matters. A language model is naturally designed to produce text. A legal research system must also be good at deciding when not to produce an answer.

## Link citations to the actual judgment

An AI system for case law research should not merely display a source reference as text. The reference should ideally link directly to the underlying judgment.

From the answer, the user should be able to open the source and check:

-   whether the ECLI number is correct;
-   whether the judgment actually exists;
-   whether the quoted passage is accurate;
-   whether the context matches the AI-generated summary;
-   whether the judgment is genuinely relevant to the research question.

This turns AI into an interface on top of the source material rather than the final authority. That is a much safer model for legal research than a system in which users only see an AI-generated summary.

## Make the model refuse when no source is found

One of the most important safeguards is also one of the simplest: the model must be allowed to refuse.

If the retrieval layer cannot find a relevant judgment, the answer should state that no sufficiently supported answer can be given based on the available sources.

This is functionally different from the behaviour of a general-purpose language model. Such a model will usually try to provide a helpful answer anyway. For case law research, that behaviour is undesirable: if no source has been found, the model should not reconstruct a possible judgment from patterns in its training data.

A well-designed system therefore treats uncertainty as part of the answer.

## What additional checks reduce the risk?

Several additional controls can complement mandatory retrieval and source links. For example:

-   index only documents from approved legal sources;
-   extract ECLI numbers directly from document metadata rather than letting the language model generate them;
-   display quotations directly from the retrieved source passage;
-   verify that every judgment mentioned in the answer appears in the retrieval results;
-   clearly distinguish source text, summary and interpretation;
-   block answers when insufficient source material is available.

Logging can also be useful. If the system records which documents were retrieved for a particular answer, it becomes possible to reconstruct which sources the AI relied on. That makes errors easier to investigate and the overall system easier to audit.

## AI as a search and analysis tool, not as the source

The most useful role for AI in case law research is therefore not to replace legal sources. AI can help users search large collections of text more efficiently, identify relevant passages, summarise judgments and highlight similarities or differences between decisions.

The underlying judgment, however, remains the source.

That distinction is essential. The moment an AI-generated reference is treated as an independent legal source, the same risk reappears that RAG was intended to reduce.

## Conclusion

AI can search case law reliably when the system is required to connect its answers to real judgments. RAG supports this by retrieving relevant sources before generating text. Mandatory source references, links to the original judgment and refusal when no supporting source is found are more important than simply using a more powerful language model. For firms putting this into practice, that is exactly the approach behind [AI in legal practice](/en/blog/ai-in-de-juridische-praktijk).

Verifiable legal research

## Want to see how source-grounded case law search works?

Our case-law RAG retrieves judgments with a source reference per claim, including a link to the actual judgment. Curious what that could mean for your files?

[View the showcase](/en/showcase/rechtspraak) [Ask your question](/en/contact)

---
Volledige (opgemaakte) versie: https://neuralex.nl/en/blog/ai-en-jurisprudentie