AI Hallucinations: How KAIUS is designed to reduce them

In legal matters, a hallucination — meaning an assertion produced by AI that sounds true but is inaccurate, fabricated, or poorly sourced — is not just a technical curiosity. It is an ethical, professional and, at times, financial risk. Widely publicized cases in 2023 and 2024, where lawyers filed submissions citing non-existent case law, have firmly established mistrust.

At KAIUS, we start from a simple principle: a legal assistant that hallucinates is unusable. Reliability is not just one feature among many; it is the condition of use. This article explains how our architecture is designed to reduce these errors and, above all, to make them visible to the lawyer before they leave the firm.

Where do hallucinations come from?

A language model does not "know" the law; it produces statistically plausible text based on its training data. Three main causes explain hallucinations in a legal context:

• Outdated or incomplete training data on certain jurisdictions, particularly Belgium and Luxembourg.
• Lack of context: the model does not know which case it is working on, or with which documents.
• Implicit incentive to "answer no matter what": a generalist model prefers to produce a fluid response rather than admit it doesn't know.

RAG: anchoring every response in a real source

KAIUS relies on a RAG (Retrieval-Augmented Generation) architecture. In concrete terms: before generating a response, the system first searches for relevant passages in verified sources — case files, indexed case law, current codes, firm content. Only from these excerpts is the response composed.

This constraint changes everything: the AI no longer "reinvents" freely, it reformulates existing material. When the material does not exist, it is explicitly instructed not to fill the void.

Clickable citations: verification in one click

Every assertion produced by the KAIUS assistant that is based on a source (article, ruling, case document) is accompanied by a clickable reference that opens the original source. The lawyer can therefore, in a few seconds, verify that the sentence they are about to use is indeed based on what is stated.

This mechanism restores a clear chain of responsibility: the AI proposes, the source attests, the lawyer decides. This traceability is what makes its use ethically compatible with professional requirements.

Safeguards: refusing rather than inventing

The KAIUS assistant is instructed to prefer an explicit "I don't know" over a plausible but unfounded answer. When the documentary base does not provide sufficient evidence, the interface clearly indicates: "No relevant sources found for this question in your file."

This stance is culturally counter-intuitive for a consumer tool, but it is essential in a professional context. No answer is better than a wrong answer.

Restricted scope: less, but better

Rather than aiming for universality, KAIUS targets specific legal use cases: document summaries, drafting submissions, assisted case law research, generation of letters and emails. Each use case has its own prompt, scope, safeguards, and regression tests.

This restricted scope mechanically reduces the hallucination surface: a system that knows exactly what is expected of it makes fewer mistakes than a generic assistant invited to do everything.

Human supervision: the final line

No document produced by KAIUS leaves the firm without human validation. The interface is designed to make review quick: highlighting of generated passages, confidence level indicators, explicit flagging of areas where the AI relied on a single excerpt. The goal is to focus the lawyer's attention where it is needed.

We state it clearly to our users: AI is a productivity tool, not a substitute for expertise. The final responsibility remains with the lawyer. Our role is to provide the means to exercise it quickly and with peace of mind.

Residual hallucinations: what to do?

No system reaches zero hallucinations. Our commitment is threefold: transparency about limitations, continuous updates to prompts and databases, and a feedback loop that allows users to report a suspicious case in one click. Every report is processed and feeds into regression tests.

A legal AI that never says "I don't know" is dangerous. An AI that says it, and cites its sources when it answers, becomes a true assistant.

Taking it further

The reliability of a legal AI is measured over time and in the field. We regularly publish our developments, and any user can test KAIUS for free for 14 days to evaluate the quality of responses on their own files — that's the only test that truly counts.