Why AI answers need a source reference

AI answers need a source reference because a language model writes what sounds likely, not what is documented. Without a document and a page you cannot check an answer. For deadlines, quantities, amounts and editions of standards, it must not go into a contract or a letter without this evidence.

Why models invent things

Specialists call invented but convincing-sounding statements hallucinations. A 2025 paper by Kalai and colleagues explains them like this: training and evaluation reward guessing over admitting uncertainty. The widely used evaluations only grade right or wrong. They give no points for “I don’t know”, but sometimes do for a guessed answer. The model becomes a good test-taker who would rather guess.

Even the 2020 publication on which search followed by an answer (retrieval-augmented generation, RAG) is based names the provenance of a statement as an open problem of pretrained models. A search in your documents helps, because the model is given real passages of text (what a language model can and cannot do in general is covered in What language models can and cannot do). But it does not make errors impossible. In its documentation on this topic, Anthropic writes that the methods described there reduce hallucinations considerably but do not eliminate them. Critical statements must always be checked.

The following example uses invented data. The question is: “Which deadline applies to the final payment on the Musterstraße project?”

• Answer A: “17 working days.” It names no source. It may be right, but nobody knows.
• Answer B: “17 working days after receipt of the final account. Source: building contract Musterstraße, page 9, extract: ‘…final payment within 17 working days after receipt…’.”

With B, someone opens page 9 and sees at once whether it is right. With A, the search starts from scratch.

How to recognise a good source reference

A usable source reference contains four details: the document name, the page, the passage in its exact wording, and the date or version of the document. A knowledge base supplies it from its index.

For long documents (the documentation names more than 20,000 tokens), Anthropic recommends a similar pattern. The model first extracts verbatim quotations and then answers only on the basis of those quotations. If it finds no supporting quotation for a statement, it should withdraw the statement. It should also be allowed to say that it does not know something. According to the documentation, this permission alone can sharply reduce false statements. “Not found” is therefore a valid and desirable answer from a knowledge base.

At Microsoft, the metric for this is called groundedness. It measures whether the answer is based completely on the retrieved texts. The architecture documentation also names completeness, utilisation of the retrieved hits, relevance and correctness. It recommends documenting the results and aggregating them across many queries.

Where the source reference comes from

In a knowledge base, the model does not produce the source reference from memory. It comes from the index, which stored the document name, page and section when the document was read in, and it is passed on with the hit. Microsoft describes this for its document services as source grounding: every extracted value can be traced back to its position in the document. The answer remains a suggestion, but the source reference is a pointer you can check. How such a knowledge base is built is described in Searchable project files.

The check before use

The four questions in the graphic are quickly asked. What matters is that they become routine and that someone records them: who, when, on which document.

For the spot check: in the start-up period, the named contact person checks every answer with a number, deadline or standard, and a selection of the rest, as a starting assumption about one in ten. Keep count while doing so. The share of answers that the original does not support is your own measure. If it does not fall, look at the filing and at how the questions are put.

Record four things for each answer checked: the question itself, whether the source reference is correct, whether the original supports the statement, and whether it was the current version. The errors show what kind of countermeasure is needed. If the source reference is wrong, the fault lies in reading in or in the index. If the original does not support the statement, the model summarised wrongly. If the hit comes from an old version, the filing lacks a marking. A hit can also be too narrow, for example when an exception appears two paragraphs further on. You only notice this if someone reads the surroundings of the passage and not just the extract.

When you must not use an answer

Some statements are expensive when wrong, so an answer without evidence counts for nothing:

The five-year period is named here only as an example. Whether and when it applies in your case is for your legal adviser to clarify.

Limits of the source reference

A source reference proves that the passage exists. It does not prove that it fits the question, that it was summarised correctly or that it still applies. A model can misrepresent a real passage, and in many filing systems an old version sits next to the new one. That is why date and path belong in every hit, and the final decision stays with your people.

How much can be avoided depends on well-maintained data and established routines. I deliberately give no percentage of avoided wrong answers, because I have no study that transfers to construction records. How many wrong answers remain in your case is shown by your own sample.

What you can do now

1. Introduce the rule: no number, deadline or standard from an AI answer goes into a letter or contract until the source reference has been read at the original.
2. Ask ten questions whose answers you know, and count how often the answer and the source reference are right.
3. Define a check note: who checked which document, and when.
4. Decide which statements are always checked by a second person at the original.