pattern playground

Citation

Added · Last updated

There is a piece of information the actor needs to use. They did not write it, and they cannot see where it came from. Checking it properly means re-researching it, cost of which might be too big. Also, fluent wrong output is accepted exactly as readily as right output, because the surface gives nothing to tell them apart.

Forces

  • Mark vs. follow. The mark works on trust whether or not it is followed.
  • Document vs. passage. A pointer to a whole document is cheap to produce and expensive to use: the actor has to find the supporting sentence themselves, and citation links “offer little guidance about what evidence they contain”. A pointer to the passage is what makes the check short.
  • Staying vs. leaving. Professional fact-checkers verify by leaving the page and reading laterally; hover cards and in-place excerpts keep the actor inside the answer.
  • Coverage vs. correctness. Citing during generation buys precision and loses coverage; attaching citations afterwards covers more and supports less.
  • Every claim vs. the claims that need it. Marking everything makes the marks noise; marking nothing leaves the contentious claim indistinguishable from the banal one.

Solution

Citation attaches a pointer to each claim — a marker, a chip, a footnote — that resolves to the material the claim rests on. Render the marker so it can be read at a glance, previewed without leaving the answer, and opened into the source at the cited span. Place the sources where the actor meets them before any explanation of the answer, since a convincing explanation lowers the chance they look. Where a claim has nothing to point at, say so.

Variants

Ordered by how far the pointer sits from the evidence.

Passage at the claim

The marker resolves to the exact span, highlighted in the source, with the quoted text previewed on hover. NotebookLM’s numbered markers open the document scrolled to the highlighted paragraph; Anthropic’s Citations API makes the span the unit of the contract — each claim carries the cited text and its character range in a named document. Adobe Acrobat’s inline highlights and Granola’s transcript quotations are the same form on PDFs and recordings. The cost of following is lowest here; the risk is premature convergence on one source.

Chip with popover

A domain chip at the end of the claim (“northjersey +3”), opening an anchored card with title, snippet, favicon, and 1/N paging where several sources back one claim. Perplexity and ChatGPT both use it. The “+N” collapses multiplicity without cluttering the prose; the snippet is scent, not the evidence.

Source panel

A favicon stack in the answer bar opening a sidebar that lists every source as a card while the answer stays in view. This is the overview of the evidence base, complementary to the inline marks; on its own it gives the actor nothing to follow from.

Source list

Full URLs listed after the answer, unanchored to any claim.

Letting the user select a passage and ask for its sources to be re-checked without re-running the request

The Reference component carries a source document and a character range per reference, which is the data shape the passage-at-the-claim variant needs.

Research on this pattern

Consequences

  • each claim now has somewhere to look, so checking one claim costs a follow rather than a search — and a checkable claim is what a deliberate pause before acceptance needs in order to land on something
  • the mark raises trust whether or not anyone follows it, and most actors don't: the sign of checkability gets read as the fact of having checked, including when the citation is irrelevant or wrong
  • the characteristic failure is not a dead link but a real source that doesn't say what the claim says, so a pointer to a document under-specifies the check — the passage is what makes it checkable
  • citing narrows what gets said: the system leans toward what it can point at, and stricter citing trades coverage for correctness

Resources & references

Appears in

Calibrating reliance

Related patterns

Precedes

  • Cognitive forcing functions — each claim now has somewhere to look, so checking one claim costs a follow rather than a search — and a checkable claim is what a deliberate pause before acceptance needs in order to land on something

Enacts

  • Agency — gives the actor a place to check each claim, so trust can be selective rather than all-or-nothing

Complements

  • Transparent reasoning — the evidence for a claim versus the path that produced it
  • Explanation — where it came from versus why the system said it — and an explanation the actor finds convincing reduces their inclination to open the sources beside it
  • Expressed uncertainty — how sure it is versus where it came from
  • Provenance marking — who made it versus where it came from

Related to

  • Annotation — a citation is a system-authored mark anchored to the claim it supports, where annotation is the actor's own layer over content they can't change
  • Text lens — a summary whose sentences point back into the long document lets the actor go deeper at the relevant passage instead of re-reading
  • Prose — a citation is the hyperlink in its citation mode — the anchor names its target and the reader may follow it for verification

Preceded by

  • Generated content — the generated output is the material a citation attaches pointers to
  • Reference material — the output is anchored in named material, and can point back to it — the attachment that went in is what a citation can later point out

Enabled by

  • Link preview — hover previews let a source be weighed without leaving the claim