Separate the citation from the claim.
You ask an AI assistant for background on a proposal. It returns a tidy reference list. Before that list goes into a presentation, document or tutorial, check one citation independently. This guide uses Crossref to inspect a DOI and its registered title, followed by a human source review.
There are three different questions: did this registry return a record, do the citation details agree, and does the actual source support the sentence you want to write? Keep those answers separate in a review ledger. A successful lookup answers only part of the first two.
This is an established research workflow, not a new AI feature. The examples below were authored for the guide. We did not ask a model to produce citations or measure how often models make citation errors.
Start with one public DOI.
- Copy the original citation and the exact claim into your notes before editing either.
- Open Crossref Metadata Search and search the DOI. With no DOI, search the title; treat search results as candidates to inspect.
- Compare title, authors, venue and year with the citation. Open the publisher record from the DOI and inspect any differences.
- Read the relevant source passage before deciding whether to keep the claim.
The fixture uses DOI 10.1145/3442188.3445922. At review time its Crossref title field reads On the Dangers of Stochastic Parrots. That is a metadata field, not a transcription of every subtitle or character the publisher may display. This guide makes no substantive claim about the paper’s findings.
Try a small, inspectable check.
Prerequisites: Windows PowerShell, Python 3.13 available as python, and internet access to api.crossref.org. No third-party packages or AI API key are required for this exercise. Download the practice ZIP or use these commands from a folder where you can create a new directory:
$work = Join-Path $PWD 'citation-practice-2026-10-22'
New-Item -ItemType Directory -Path $work -ErrorAction Stop | Out-Null
Set-Location $work
Invoke-WebRequest -Uri 'https://jabsystems.io/ai-anonymous/Social/check-the-citation-practice.zip' -OutFile '.\practice.zip'
Expand-Archive -LiteralPath '.\practice.zip' -DestinationPath '.\practice' -ErrorAction Stop
Set-Location '.\practice'
Get-Content '.\check-citations.py'
Get-Content '.\citation-fixtures.json'
Review the script before running it. It sends only the fixture DOI strings to Crossref. It does not upload an AI answer, draft document or claim text. The command below saves the report locally:
python .\check-citations.py | Set-Content -Encoding utf8 .\my-citation-results.json
Get-Content .\my-citation-results.json
The complete Python script uses the standard library’s Request and urlopen. It URL-encodes each DOI, uses a timeout, and reuses the repeated DOI response within the run. There are two unique requests, separated by a one-second pause. The core comparison is deliberately simple:
def normalize(title):
return " ".join(title.split()).casefold()
agrees = any(normalize(claimed_title) == normalize(t) for t in returned_titles)
Only whitespace and case are normalized. Punctuation, subtitles, translations and incomplete metadata can cause a difference that a person must resolve. The script prints returned author surnames and publication date parts for inspection; it does not automatically validate them.
What actually happened in our run.
On October 22, 2026, we executed the downloaded code’s authored fixtures on Windows with Python 3.13.15 against the live Crossref API. The observed JSON report records the UTC check time and returned details.
- Matching metadata title: the real DOI returned a record and the title comparison reported
title_agrees. - Deliberately changed title, same real DOI: the lookup still returned a record, but the comparison reported
title_differs. - Deliberately missing identifier: the authored identifier returned HTTP 404 and
not_in_crossref. Its title was not checked.
All three results retain claim_support: not_checked. The changed title and missing identifier are intentional examples, not mistakes observed from an AI model. A process exit of 0 means the report was produced, not that its citations passed review.
Keep “unresolved” as a real outcome.
The Crossref API documentation describes DOI lookup and registration-agency coverage. A 404 from this endpoint is scoped to Crossref; it is not a universal verdict about DOI existence. Check for a transcription error, try the DOI resolver and investigate the relevant registry or publisher. Do not turn an outage, timeout or missing field into an accusation.
On a service error, including rate limiting, this script reports unresolved and stops. It does not retry automatically. Follow Crossref’s API tips, back off before retrying and keep responses you already retrieved. For ongoing use, add your own valid contact information according to its etiquette guidance and use persistent caching. This tiny exercise is not a bulk citation service.
Tested: two live registry requests, the three authored outcomes above, and extraction/rerun of the practice files. Not tested: an AI model, citation-error frequency, semantic claim support, research quality, comprehensive author/year validation, correction or retraction detection, network-failure simulation, large batches, other operating systems, or performance. Metadata and API availability can change. The example DOI returned registry metadata, but its publisher page blocked our automated fetch with HTTP 403; we did not verify the full paper or any claim from it.
Read the sentence behind the citation.
Once the record looks plausible, open the source. Find the section, page or table that addresses the exact claim. Record a short relevant passage and its location. Check whether the wording describes an experiment, an opinion, a limitation or a result with a narrower scope. Look at the publisher’s current version and any correction or update notice. If you cannot access enough of the source, leave support unresolved.
Use the ledger to decide: keep the sentence, narrow it, replace the reference or omit the claim. A real paper can still be the wrong support for a particular sentence. This human reading step is our proposed workflow, not something the supplied script performs.
An optional proposed extraction prompt asks an approved AI tool to separate claimed citation details from the claim without silently repairing them. It was not executed or evaluated. Verify the extraction yourself, and use only material appropriate for that tool.
Take the source-checking kit.
Practice ZIP · Complete script · Authored fixtures · Observed report · Human review ledger · Proposed prompt · Practice README
Primary sources reviewed October 22, 2026. Crossref’s REST API overview provides the starting point. The original portrait is AI-generated conceptual artwork, visually reviewed; its mock identifier is illustrative.

