GLOBAL PERSPECTIVE · INDEPENDENT THINKING
UNDERSTAND WHAT MOVES THE WORLD.
Technology & AI · News analysis

An AI Citation Check Is a Request to Read the Source

Wikimedia’s experimental source-checking update puts the relationship between a sentence and its citation under scrutiny. An automated suggestion should open an editorial inquiry.

BY GABRIEL NICOLAEV – CYGURU · · 4 MIN READ
READING TOOLS
Hands following printed lines in an open book on a wooden table; illustrative photograph.
THE READING GUIDE

Inside this article

01A newly reported experiment, not a machine verdict on truth ↓02Locate the proposition before accepting the flag ↓03An unavailable source is a limit on the check ↓

Event / source date: Wikimedia reported the experimental availability in its 9 November 2026 technical newsletter; linked documentation is supporting context, not a second new launch. · Published 10 NOVEMBER 2026

A newly reported experiment, not a machine verdict on truth

Wikimedia’s technical newsletter dated 9 November says Source Verification is available as an experimental suggestion for experienced editors on English and Russian Wikipedias. It describes an LLM hosted on LiftWing attempting to check whether a cited source supports the attached claim. That is an account of an experiment’s availability and purpose, not proof that its judgments are reliable in every case.

The practical editorial question is what should happen after a suggestion appears. A citation check concerns a relationship between two pieces of text: what an article asserts and what its cited material supports. The editor still needs to identify the actual statement, inspect the relevant evidence and decide whether the article should change.

Locate the proposition before accepting the flag

Imagine a hypothetical encyclopedia sentence stating that a bridge opened in June, cost a particular amount and shortened a journey. Its reference might support the opening date while saying nothing about the cost or travel time. A single judgment attached to the sentence would leave several possible editorial tasks unresolved.

The reviewer should split the statement into the propositions that matter. Which part has support? Which requires another source? Which is an inference the article has presented as established? This exercise need not make the published prose awkward. It makes the working question precise. The editor can then revise the unsupported clause rather than either discard a useful reference wholesale or allow its strongest supported detail to lend authority to everything nearby.

An unavailable source is a limit on the check

The linked Source Verifier documentation describes an experimental user script and warns about false positives and difficulties reading some sources. It says support checking does not determine source reliability. The script documentation is useful context; it should not be treated as a measured evaluation of every deployment mentioned in the newsletter.

For an editorial workflow, the important distinction is between a source that was examined and did not support a claim, and a source the checker could not examine. The second result creates a retrieval problem. An editor might obtain the original document through an appropriate route, ask for a better accessible reference or leave the question open while seeking evidence. Treating lack of access as a demonstrated contradiction would add a conclusion that the failed check did not establish.

Support and reliability require different questions

A hypothetical promotional statement might support the sentence that an organisation made a particular promise. The same statement should not automatically support the sentence that the promised result occurred. The editor needs to preserve attribution and the distinction between a declaration and an outcome.

Even a close textual match leaves another question: is this the right kind of evidence for this claim? A personal recollection, an official count and a company announcement can serve different purposes. The proposed editorial standard is to state what the source establishes within its limits. An automated match can help locate a relationship worth examining. It should not be allowed to decide the wider evidentiary role merely because the language aligns closely.

A suggested correction needs a reason another editor can inspect

When a reviewer changes a sentence, the explanation should identify the problem and the supporting material. “The tool flagged it” does not explain whether the source lacked the claim, the article overstated an estimate or the source could not be retrieved. Those distinctions determine whether another editor should accept the change, seek more evidence or challenge the interpretation.

In the fictional bridge example, an edit summary might explain that the opening date is supported while the cost figure needs a separate citation. The discussion can remain narrow enough to resolve. The software’s suggestion has then served as a starting point for editorial work, while the final reasoning remains available without requiring another person to trust the same software.

Preserve disagreement instead of converting it into a score

Two editors might read the same passage differently. One sees a direct assertion; another sees a qualified projection. The next useful step is to examine the wording and context that produce the disagreement. Running another model might offer an additional perspective, but it should not replace the question being disputed with a vote among outputs.

The working record should distinguish what the source says from what the article concludes. An editor may decide to retain the information with a qualification, replace the source or remove the disputed claim. Each option should be justified by the evidence and the publication’s standards. This keeps the correction process accountable to the article, rather than to the task of making every automated indicator appear satisfied.

The experiment should be judged through completed editorial work

A useful evaluation would ask what happened to a reviewed sample after human examination: which suggestions identified a genuine citation problem, which reflected access limitations and which should not have changed the text. The sample and review method should be described before any broad performance conclusion is drawn. This is a proposed evaluation approach, not a report of new findings.

CYPAPER’s earlier essay on the design of AI review addresses a wider workplace question. The Wikimedia update provides a specific editorial setting for it. A citation suggestion earns its place by helping a person make a better supported edit. The lasting product should be a clearer claim and an inspectable reason, not a reassuring symbol placed beside a source nobody has actually read.

Gabriel Nicolaev — CyGuru

Sources & further reading

Wikimedia Tech News 2026, Issue 46 ↗

Meta-Wiki: Source Verifier documentation ↗

Hands following printed lines in an open book on a wooden table; illustrative photograph.

FOLLOW THE CONNECTION

Another angle on this subject

Continue through related ideas, decisions and consequences.