Blog · Data custody
Why AI answers need sources: grounding and citations explained
AI answers need sources because a language model produces fluent, confident text whether or not the content is true, and the reader cannot tell the difference from the wording alone. Grounding gives the model real documents to answer from, and citations let a person check each claim against them.
This guide explains why models get things wrong, what grounding and citation mean in practice, and how to judge whether a tool that "shows its sources" is doing so properly.
Why a model can be confidently wrong
A large language model is trained to predict the next piece of text given what came before. Through training on a very large body of writing it absorbs a great deal of general knowledge, but that knowledge is stored as patterns in the model's weights, not as a database of facts with references.
Several consequences follow:
- No built-in link to a source. When the model states something from memory, there is no record of where it learnt it.
- A training cutoff. The model knows nothing after its training data ends unless you supply it.
- No access to your private data. It has never seen your contracts, your field records or your customer files.
- Plausible gap-filling. When the pattern calls for a name, a number or a reference and the model does not have one, it may produce something that looks right. This is commonly called hallucination. Invented references that look entirely genuine are a well-known example.
None of this makes models useless. It means an unsupported answer is a draft, not a finding.
What grounding means
Grounding means the answer is based on specific documents supplied to the model at the time of the question, not on the model's memory. The usual method is retrieval-augmented generation:
- Your question arrives.
- The system searches a collection of documents and picks the most relevant passages. The search may use keywords, meaning-based similarity, or both.
- Those passages are placed in the request, with an instruction to answer only from them.
- The model writes an answer and marks which passage supports each statement.
The model is now doing a job it is good at, which is reading and summarising text in front of it. It is no longer being asked to recall.
Grounding reduces errors. It does not remove them. If the search returns the wrong passages, the answer will be built on the wrong material. If the passages do not contain the answer, a poorly instructed model may fall back on memory without saying so.
What a good citation looks like
A citation is only useful if a person can follow it in a few seconds and see the supporting words.
| Good citation | Weak citation |
|---|---|
| Points to a specific passage, page or record | Points to a whole website or a long document |
| Refers to a stored copy that will not change | Refers to a live page that may be edited or removed |
| Is attached to the individual claim | Is a list of links at the end of the answer |
| Says who published the source and when it was fetched | Gives a title only |
The stored-copy point matters more than it first seems. If the evidence can change after the answer is written, the citation can silently stop supporting the claim. Keeping the fetched file, with a checksum and its origin, is covered in the guide on data provenance.
A citation is not proof
A model can attach a real source to a claim the source does not make. It can also overstate, turning "may be associated with" into "causes". So a system that takes citations seriously adds checks after the model has written:
- Does the cited passage exist in the store? A reference to something that was never retrieved is rejected.
- Does the passage support the claim? This can be tested by matching quoted text, or by a second automated pass that compares claim and passage.
- What happens to unsupported claims? The strict approach is to remove or refuse them, not to show them with a warning nobody reads.
- Can the system say "I do not know"? An honest "the documents do not cover this" is a feature. A tool that always produces an answer is a tool that sometimes invents one.
The sources have to be good
Grounding moves the question of truth from the model to the collection it reads. An answer can be perfectly cited and still wrong, because the source was wrong, out of date, or about a different entity with a similar name.
That is why the unglamorous work before the question matters: gathering from known publishers, recording where each record came from, working out which records refer to the same thing, and having a person review doubtful matches. The guides on entity resolution and the seven steps of a data pipeline cover that groundwork.
How to test a tool that claims to cite
You can do this in ten minutes with any product.
- Ask a question you already know the answer to, and follow every citation. Does the passage say what the answer says?
- Ask about something that is not in the documents. Does it say so, or does it improvise?
- Ask the same question twice in different words. Do the answers agree?
- Check whether the citation opens the exact passage, or leaves you to search a long file.
- Ask where the underlying documents came from and when they were last updated.
- Ask what the system does when the model makes a claim with no source.
When sources matter most
For brainstorming, drafting and rewording, unsupported output is fine, because you are the judge of whether it is useful. Sources become essential when the answer will be relied upon by someone else: a report to a funder, a finding in a review, a figure in a board paper, a claim about a patent or a regulation. In those settings the test is simple. Could you defend this sentence if someone asked where it came from?
There is also a custody angle. A cited answer keeps a visible trail from claim back to record, and from record back to publisher. That is the same chain-of-custody idea that applies to your own data, described in data custody in AI.
Where Prism fits
Prism's engine turns a field's data into answers that show their sources. Every record keeps its publisher, request and file checksum, a person approves or rejects links, and in the Answer step the AI must cite stored evidence; uncited claims are refused automatically. Prism's own patent corpus holds 109 million claims, each traceable to source text, although the search product is still being built. See industries for where this applies.