Checking sources in AI output: what stays, what changes

An answer from Copilot, Claude or a legal AI tool often reads as if an experienced colleague wrote it: clearly structured, confident and with sources attached. In almost every training we get the same question: how do I know whether I can trust this? The honest answer is reassuring. You check AI output the same way you check any other source. But something does change, and that determines what you look out for.

What stays the same

A lawyer does not blindly trust a memo from a trainee either. You look up the ruling, read the passage that matters, check whether the rule still applies and ask whether it fits your case. Those are exactly the questions you ask of AI output.

1. Does the source exist? Look up every ruling, statute or publication yourself in a database you trust.
2. Does the source say what the answer claims? Read the passage yourself, not just the title or the summary.
3. Is it current? Check the date and whether anything has changed since.
4. Does it fit your situation? That judgement stays with you, and so does the responsibility.

If you are used to reading sources critically, you already have the most important skill.

What does change

In practice it still works differently, for three reasons. The first is that the errors look different. A trainee who is unsure lets you know. A language model predicts which words are likely to follow and fills in what it does not know just as fluently as what it does know. That is how you get a ruling with a real-looking citation that does not exist. You cannot tell the difference from the tone. In 2023, lawyers in the United States were fined for including such invented rulings in a court filing without checking them.

The second reason is volume. Where you used to review one memo, you now have five versions, a summary of the whole file and an overview of the case law within a few minutes. The more text, the greater the temptation to read only the first paragraph.

The third reason is that a source reference does not tell you everything. Tools that work on your own documents or a legal database often add a link. That helps, but the passage behind the link sometimes says something slightly different from what the summary claims. And when agents start carrying out steps on their own, the result moves further into your process before anyone looks at it. So you need to choose the moment of checking deliberately.

An example from practice

At a law firm, a lawyer who was writing an article on a doctrine in construction law brought that question in as an exercise. The group had an AI tool set out the state of case law and literature, and then went through the result together. An exercise like that, on your own question, shows better than any explanation where the work lies: which reference exists, which says something different from what the answer claims, and what is missing.

How to make it a standard step

Agree per work process when AI output is always checked, for example for everything that goes to a client, a court or an executive. Ask the tool to name the source for every claim and to say where it is unsure; that makes your check faster. And record who is responsible for the final text. Since 2 February 2025 the EU AI Act also requires organisations across the EU to make sure their staff are sufficiently AI literate. Being able to critically assess what a system produces is a natural part of that.

Want to start straight away? Take a piece you are writing this week anyway, have the tool you use at work do the research or a first draft, and go through it with the four questions. Want to practise this with your team on your own work, such as a memo, a court filing or a council proposal? We do that in our hands-on AI literacy training and in our training for lawyers, notaries and government. Feel free to get in touch.