Every few weeks, a new “magic prompt” appears on LinkedIn.
It promises that ChatGPT, Claude or another large language model will stop hallucinating if users simply paste a few lines into the custom instructions:
“Never invent sources.”
“Say when you are uncertain.”
“Only answer if you know.”
“Prioritise truth over fluency.”
This sounds reasonable. It is also dangerously incomplete.
For casual work, such prompts can help. They may make an AI assistant more cautious. They may encourage it to use disclaimers. They may reduce some obvious mistakes.
But they do not solve the core problem.
A prompt does not turn a general-purpose language model into a reliable legal system.
Hallucinations Are Not a Prompting Problem Only
OpenAI defines hallucinations as plausible but false statements generated by language models. These errors can appear even in seemingly simple factual questions.
Anthropic says the same about Claude. Its own documentation states that even advanced models can generate text that is factually incorrect or inconsistent with the given context. The official guidance speaks about reducing hallucinations, not eliminating them.
This distinction matters.
A model can be instructed to behave more carefully. But it does not suddenly gain a perfect truth mechanism. It does not automatically know which legal source is current. It does not reliably know whether a citation exists. It does not always know when it is wrong.
That is not a small issue in law. It is the issue.
Legal Work Is Different
If an AI assistant invents a marketing claim, the result may be embarrassing.
If it invents a legal basis, misreads a statute, overlooks a cantonal rule or cites a non-existent case, the result can be costly.
Legal questions require more than fluent language. They require:
verified legal sources
jurisdiction-specific reasoning
current law
transparent citations
traceability
confidence assessment
clear limits when the answer is uncertain
A general-purpose chatbot was not built for this.
Research from Stanford has shown why caution is necessary. General-purpose chatbots hallucinated between 58% and 82% of the time on legal queries in earlier tests. Even specialised legal AI research tools using retrieval-augmented generation still produced hallucinations between 17% and 33% of the time.
That does not mean legal AI is impossible.
It means legal AI must be engineered differently.
RAG Alone Is Not Enough Either
Some providers claim that adding documents through retrieval-augmented generation solves the problem.
It helps. But it is not sufficient.
A system may retrieve the right document and still misinterpret it. It may cite the correct law and draw the wrong conclusion. It may miss an exception. It may confuse similar legal concepts. It may produce an answer that sounds precise but is legally incomplete.
This is why legal AI cannot be treated as a chatbot with a document folder attached.
In legal work, the system must be built around the legal domain itself.
Why Jurilo Takes a Different Approach
Jurilo was built because Swiss companies need legal answers that are fast, affordable and reliable — but not based on blind trust in a general chatbot.
The point is not to make a model sound more careful.
The point is to reduce legal risk by design.
That means working with controlled legal sources, Swiss legal structures, domain-specific reasoning, source transparency and clear safeguards. It also means recognising where a question requires caution, where the answer depends on facts, and where professional legal review may still be necessary.
A prompt can tell an AI system to “be careful.”
A legal AI system must be careful by architecture.
The Dangerous Illusion
The most dangerous sentence in AI today is not:
“AI makes mistakes.”
Most users know that by now.
The most dangerous sentence is:
“We fixed hallucinations with this prompt.”
That creates false confidence. And false confidence is exactly what companies cannot afford when dealing with employment law, contracts, data protection, tax, corporate obligations or regulatory risk.
For low-risk writing, summarising or brainstorming, general-purpose AI tools can be useful.
For legal questions, the standard must be higher.
A prompt is not a compliance system.
A disclaimer is not legal verification.
A fluent answer is not a correct answer.
That is why Jurilo exists.
Legal AI should not merely sound convincing.
It must be grounded, traceable and built for the law.
