The source rule

Never accept a source you have not opened. Ask the model what to look for, then go and find it yourself. Five moves: Search, Open, Understand, Record, Cite. It is written for the specific failure a language model creates, which is a reference that looks perfect and does not exist, and it puts the existence check first because the established frameworks for evaluating sources were all written when every source under discussion was real.

The five moves#

Search. Use the model to find search terms, debates, authors and likely material. Ask it what to look for, not for the finished reading list. If a title, an author and a year do not resolve together anywhere, you have a plausible sentence rather than a source.

Open. Open the original yourself. Not the abstract, and not a summary of the abstract. Most misquotation is a correct citation of a document nobody opened.

Understand. Abstract, method, findings, limitations. All four. If you cannot say what the study measured, on how many people, and what it does not establish, you are not in a position to use the finding.

Record. DOI, page numbers, full reference, straight into a reference manager at the moment you find it. Two minutes now replaces an afternoon in third year.

Cite. Cite the original. Never the AI summary. A claim nobody can trace is an assertion with a footnote attached.

The frameworks this sits with, and one of them is better#

Source evaluation is a solved teaching problem with two dominant answers, and neither of them is this one.

The CRAAP test, from California State University Chico, asks about Currency, Relevance, Authority, Accuracy and Purpose. It is taught in schools and universities worldwide, and it has a structural weakness: every question invites you to stay on the page and judge the page by its own contents, which is the behaviour the research says fails.

The research is Wineburg and McGrew's, published in Teachers College Record in 2019. Forty-five experienced internet users evaluated unfamiliar websites while thinking aloud: ten PhD historians, ten professional fact-checkers and twenty-five Stanford undergraduates. The historians and the undergraduates read vertically, staying on the page. The fact-checkers left almost immediately, opened new tabs and read laterally, judging the source by what the rest of the web said about it. They reached sounder conclusions in less time, and doctoral expertise in reading documents conferred no advantage at all.

Mike Caulfield's SIFT turned that result into four moves anyone can do in under a minute: Stop, Investigate the source, Find better coverage, Trace claims to the original context. It is what university libraries now teach and it inherits real evidence from the lateral reading literature. For judging a web page or a claim in circulation, use SIFT. It is the stronger instrument and this page is not going to pretend otherwise.

The failure none of them was built for#

SIFT and CRAAP both assume the source exists and the only question is how good it is. A model introduces a prior question, because the reference may be a fluent construction with nothing behind it.

An audit across 2.5 million biomedical papers found 4,046 fabricated references across 2,810 papers, with the rate of papers carrying at least one rising from 1 in 2,828 in 2023 to 1 in 277 in the first seven weeks of 2026. Review articles ran 57 per cent higher than other paper types. These are published papers, written by researchers, that went through peer review.

The failure is not that a model lies badly. It fabricates in exactly the format a real citation takes, with a plausible author, a real journal and a year that fits, and every instinct a reader has for spotting a weak source is calibrated on sources that exist. That is why Search comes first here and Investigate comes second in SIFT. If you already run SIFT, the amendment is one line: before you investigate the source, establish that there is one.

Understand is the move that gets skipped#

Search, Record and Cite are mechanical and people do them once told. Understand is the one that gets dropped, and it matters most, because a correctly cited paper can still be used to support something it does not show.

The deck specifies four things to understand and the fourth is the one that does the work: abstract, method, findings, limitations. Every graded entry in this evidence base carries a field for what the study does not establish. That field exists because writing the sentence is the fastest way to discover you have not read the paper. Somebody who can say "nineteen endoscopists, observational, one procedure, one country" understands the colonoscopy result. Somebody who can say "AI makes doctors worse" has a headline.

Doing this as a student, in about twenty minutes#

Install a reference manager tonight. Zotero is free and it will save more hours across a degree than any tool on any list, because the cost of a reference is almost entirely in finding it again.

One collection per module rather than per essay, so everything you read this term lives in one place and is still there in third year. Save at the moment of discovery, with the DOI and the page number, because a reference you meant to record is a reference you will spend an evening reconstructing.

And ask the model for search terms rather than a reading list. It is better at naming the debate, the authors and the likely literature than at telling you what any of them said, and the difference between those two requests is the difference between a bibliography you can defend in a viva and one you cannot.

What this rule has not been shown to do#

Nobody has tested S·O·U·R·C against SIFT, against the CRAAP test or against nothing. It has no evidence base of its own and borrows all of it from Wineburg and McGrew, whose study is about web pages and predates generative models entirely.

The honest ranking for anyone choosing one instrument: use SIFT for a claim in circulation. Take two things from here. A model requires you to check that a source exists before you check whether it is any good, and the limitations line in Understand is the one that separates a person who has read the paper from a person who has read about it.

Key sources

• Wineburg, S. and McGrew, S. (2019). Lateral Reading and the Nature of Expertise. Teachers College Record, 121(11).
• Caulfield, M. (2019). SIFT (The Four Moves). Hapgood, 19 June 2019.
• Topaz, C. M. et al. (2026). Fabricated citations: an audit across 2.5 million biomedical papers.

Related SuperSkills research#

The frameworks this sits with, all six compared, and the sequence it belongs inside, think, AI, think. On the failure itself, what a hallucination is and why AI sounds so confident. On checking, how to know when AI is wrong. For students, how to use AI at university. On how this research grades its own sources, how this research works.

About this framework#

S·O·U·R·C is used by Rahim Hirji in teaching and appears in the Mastering AI deck, most recently in September 2026, on the slide headed "So here is the rule". No claim of first use is made: no dated first publication exists for it and the Box of Amazing archive carries none, so it anchors to SuperSkills (Kogan Page, 2026) and to the deck. Correction, 6 September 2026: an earlier version of this page stated that the framework did not appear in the deck and dated it to that day. That was wrong. It is on slide 52, and the error was in the search rather than in the deck. The CRAAP test belongs to California State University Chico, SIFT to Mike Caulfield, and the lateral reading result to Sam Wineburg and Sarah McGrew. The judgement that SIFT is the better instrument for a web page, and that a model requires an existence check before an evaluation, is an interpretation by Rahim Hirji and is marked as an interpretation and not a finding.