A citation demonstration

Can an AI show where its answer came from?

Not from the answer alone. But a harness can record which source it retrieved, what passage it gave the model, and which claim points back to that passage.

Question: How many control and Alzheimer’s disease cases were used in the Visium experiment?

Ask a model by itself

Model-only example · no retrieved source

“The experiment used three control cases and three Alzheimer’s disease cases.”

SOURCE: UNKNOWN. The words alone cannot tell us whether the model remembered the paper, saw the fact elsewhere, inferred it, or guessed.

Ask through a citation-aware harness

Fetch paper PMC9773466 Select the relevant Results passage Add it to model context as Source 1 Require the answer to cite Source 1

Evidence-aware answer · source recorded

The Visium experiment used three control cases and three Alzheimer’s disease cases. The control cases were Braak stages I–II; the AD cases were Braak stages III–IV.^[1]

Now we can show that the paper was retrieved and that its passage supports the answer. We still cannot measure how much the model’s hidden training influenced its wording.

Researchers call the broader task attributed question answering: generated text should point to identified sources.^[2] ALCE makes the implementation explicit—retrieve supporting evidence, then generate an answer with citations.^[3]

The safest quotations do not depend on the model copying text correctly. The model can select a source location while ordinary code returns the exact characters.^[4] That still requires a harness to retrieve documents and put them into the prompt.^[5]

Sources

Select a citation to inspect the exact text used.

[1] Chen et al. · Results

“Human postmortem fresh frozen brain sections from the MTG of three CT (Braak stages I-II, male, 58, 72 and 82 years old) and three AD (Braak III-IV, male, 86, 86 and 89 years old) cases (see detailed case demographics, clinical and neuropathological information in Additional file 2: Table S1) were chosen to perform the 10 × Genomics Visium experiments.”

PMC9773466 · Last reverified 3 September 2026

[2] Attributed Question Answering · Abstract

“We believe the ability of an LLM to attribute the text that it generates is likely to be crucial in this setting.”

Paper · Dataset repository

[3] ALCE · Abstract

“ALCE collects a diverse set of questions and retrieval corpora and requires building end-to-end systems to retrieve supporting evidence and generate answers with citations.”

Paper · Repository

[4] Deterministic Quoting · Technical article

“The AI chooses which section of source material to quote, but the retrieval of that text is a traditional non-AI database lookup.”

Read the article

[5] Simon Willison · Citations for RAG

“You still need to implement most of RAG—identifying potentially relevant documents, then feeding that content in as part of the prompt—but Claude’s API will then do the difficult work of extracting relevant citations and including them in the response that it sends back to you.”

Read the post