CiteOwl

Blog/Getting citations right

I cited a paper I never read

The CiteOwl Team· Updated 29 August 2026· 13 min read

You took it from the abstract, or from the way another paper described it, and now you are wondering how much trouble you are in. Almost certainly none. Citing from an abstract or from somebody else's reference list is ordinary behaviour, and the peer-reviewed literature is full of it: when 2,648 researchers were each shown a real citation of their own work, one in six judged it inappropriate. The risk is not that a marker will catch you reading too little, because nobody audits reading. It is that an unread citation is the doorway a wrong claim and a wrong identifier walk into your paper. There are two honest ways out, and neither involves pretending.

The first exit is to read enough of the paper to cite it honestly, which is less reading than the guilt suggests. The second is to cite it as a secondary source and say where you got it. Both are below.

Everybody does this, including the people you are citing

In 2025 a team of information scientists tried a new way of measuring citation accuracy: instead of having reviewers judge a sample, they asked authors. They found a real citation of a researcher's own work, showed it to that researcher, and asked whether it was appropriate. 2,648 corresponding authors replied. Just over seven in ten agreed the citation was appropriate. 16.6% disagreed or strongly disagreed, which the authors read as a quotation error rate, and the rate did not differ between disciplines by any amount the study could detect.

The shape of the disagreement matters more than the headline. 24.2% of respondents agreed to some degree that the citing article had oversimplified their work. 15.3% said their work had been generalised inappropriately. 12.7% said the citation was misleading for the reader. These are not authors reporting invented findings. They are authors reporting that somebody described their paper without reading past the summary.

The respondents named the cause themselves: citing articles "without reading the entire article, but from an abstract or just adding a reference that another article has". That is the thing you did, described by the professionals it was done to.

It has been counted directly too. In 2010 four marine biologists took 198 papers from 33 marine biology journals, picked one citation at random from each, got the cited paper, and read it. Their title is the result: one in four citations in marine biology papers is inappropriate. The assertion was clearly supported 75.8% of the time, ambiguously supported 10.6% of the time, and not supported at all 6.0% of the time. The remaining 7.6% they called empty: the cited paper did not have the finding, it was citing somebody else for it. A 2021 study of 222 citations in polar sciences journals found 26% to 30% inaccurate, with empty citations at 9.1% in its most recent sample.

The marine biologists also name a cause that will sound familiar to anyone who has hit a paywall at midnight. Not every institution, they write, "can afford full access to all relevant academic journals, impelling authors to rely on imperfect data such as online abstracts". If that is where you are, the paper I need is behind a paywall lists the routes that usually work before you settle for the abstract.

What actually goes wrong, which is not what you fear

The fear is being found out. That is the least likely outcome here. There is no record of what you read and no tool that detects it. The marine biologists name why the habit persists: authors are "cavalier about how they cite, knowing that readers rarely retrieve the cited work and check its relevance".

Three things ride in on an unread citation, and every one is checkable by somebody who never asks what you read.

• A claim the source does not make. This is the one a marker actually tests. They open a reference, look for the sentence you attributed to it, and either find it or do not. The fix is in the paper says something else.
• An identifier that points to a different article. Copied with the reference, invisible afterwards, and it survives every check most students are taught. Section five.
• Somebody else's reading of the paper. Their compression of it, their emphasis, and any mistake they made. Section six.

The survey numbers tell you which is likeliest, and it is not fabrication. It is drift. A citation nobody opened tends to be too strong rather than false, the error that survives a skim and dies under a check. The exposure is not "did you read it". It is "does it say that".

What reading enough to cite it actually means

You do not have to read the paper. You have to be able to defend one sentence. Four things, in the order they take:

1. The abstract. Not to cite from, to orient. You are checking that the paper is about your claim at all and that its conclusion runs in the direction you assumed.
2. The result itself, where it is reported. Search the PDF for your number or term and read it where it lives, in the table or paragraph that reports it. An abstract gives the conclusion; the results section gives the finding, and they are not the same sentence.
3. The methods line that qualifies it. Sample, place, period, units. Say you want to write that street trees cut summer temperatures by two degrees. The methods line may tell you the study measured surface temperature by satellite in one city across three summers, which is a real finding and not the one your sentence claimed.
4. The limitations paragraph. The authors saying what they will not claim. If your sentence goes further than that paragraph allows, it is wrong no matter how much of the rest you read.

On a paper you already have open this is a few minutes; the mechanics of hunting down a single fact are in how to read a research paper. The four steps also give you one check free. If the sentence you found carries a citation of its own, the paper is repeating somebody else's result. That is the empty citation the marine biologists counted, and there are two clean responses: follow the reference and cite the original, or use the second exit.

The other exit, which is to say where you got it

Sometimes you cannot get the paper: behind a wall your library does not hold, in a language you do not read, a 1961 conference proceeding nobody scanned. Citing it from the description in the paper that led you to it is permitted in every major style, on one condition. You say that is what you did.

All three major styles agree on the principle and disagree on the bookkeeping, which is where the marks are actually lost. APA puts only the work you read in the reference list, so the original's identifier appears nowhere in your paper. Chicago wants both works listed, joined in the note by the words quoted in. MLA gives you the in-text form and leaves the works-cited question to your handbook. Getting that backwards produces a list entry for a paper you never opened, which is the exact failure the convention exists to prevent. The mechanics, worked through in all three styles with an example carried from start to finish, are in how to cite a source you found in another paper.

The clause costs four words. It buys a sentence that is true about what one paper reported another as finding, rather than false about what you know. Markers do not penalise this. They penalise the version where you claimed the original and it turned out not to say that.

You inherit their identifier

Copy a citation rather than opening the source and you copy its identifier too. If theirs was wrong, yours is now wrong, and the only way to find out is to open the thing it points at.

This would be a small problem if a wrong identifier failed loudly. It does not, and we measured how quietly it fails. We drew 600 journal articles at random from CrossRef, took the last run of digits in each DOI, and changed it by one in each direction. 58.7% of the time the altered string was still a registered article, and in 350 of those 352 cases the article sat in the same journal. The dataset is public with the method and forty worked examples.

So an inherited wrong DOI resolves. It carries the correct publisher prefix, because the neighbouring article has the same publisher. It opens a genuine paper in the right journal, often the same issue. It passes every check short of reading the landing page and noticing that the title is not your paper. That failure has its own page, the DOI links to the wrong paper, and our free DOI lookup shows the record an identifier is actually registered to.

Two things it does not say. It does not say how often anyone mistypes a DOI, only what happens when they do, and a resolving neighbour is not evidence anybody made that error. Read it as the reason this mistake is silent rather than loud.

You inherit their reading

The heavier inheritance is not the identifier. It is the sentence. Take a reference from another paper's list and you take that paper's account of what it shows, so their compression becomes yours.

How much of the literature is assembled this way? Two researchers at UCLA found a way to measure it without asking anyone anything: misprints. Look at a famous paper's citations and you find typing errors in the page numbers, many of them identical. The odds of two people independently mistyping the same digit of a four-digit page number are one in ten thousand, so a repeated misprint is a copy. For one heavily cited 1973 physics paper with 4,300 citations they found 196 misprints of which only 45 were distinct, and the most popular misprint had propagated 78 times. Fitting a model to that distribution, they estimate about 20% of citers read the original.

They repeated it across twelve high-profile papers and concluded that "about 70-90% of scientific citations are copied from the lists of references used in other papers". Their simplest estimator gave a wide spread across the twelve, from 12% to 58% with an average of 27%, so treat the headline as an order of magnitude rather than a rate. They concede the inferential step in plain language: it is "most natural to assume that a copying citer also failed to read the paper in question (albeit this can not be rigorously proved)".

What a copied reading does over twenty years is documented. In 2002 Anne-Wil Harzing traced a citation network of 60 references behind a number everyone in her field quoted, the rate at which employees sent to work abroad were said to fail. Her conclusion is the sentence this page is about: "miscitation has promoted a firmly entrenched myth unsubstantiated by any empirical evidence." Nobody in that chain invented anything. Each cited a predecessor who cited a predecessor, and the number at the far end was never supported by data.

Here is what nobody has measured. The misprint work counts copying but cannot tell whether the copied claim was wrong. The quotation-accuracy work counts wrong claims but never asks whether they were inherited or invented on the spot. We could not find a study that follows one incorrect claim through a chain of reference lists and counts the copies it made. The two halves have been measured separately and never joined, so anyone quoting you a rate for how often errors spread this way is quoting something that has not been established.

What to do right now

Still drafting. List the references you have not opened; nobody else will see the list. Then give each one of three fates: run the four checks from section three, convert it to a secondary citation naming where you found it, or cut it. Cutting is underrated. A shorter list you have read beats a longer one you have not, and as how many sources a paper needs sets out, the number is less fixed than most briefs imply.

Submitted, not yet marked. Split the question in two. Have you got a claim wrong, or did you simply read less than you would have liked? The second is not reportable and nobody wants that email. The first is: one sentence naming the reference, what the source says, and your correction. Report the error, not your reading habits. Volunteering that you skimmed invites a conversation about the rest of the list, and that conversation is not about anything.

Already marked. Check before you act. If the paper supports the claim and the identifier opens the right article, there is nothing to correct. If the identifier is wrong, that is a citation error and not misconduct, a distinction how to avoid plagiarism draws in detail. If the claim is wrong and carries part of your argument, say so anyway, because the alternative is somebody citing you for it and the chain in section six starting again with your paper at the top.

For a whole list rather than one entry, work in bulk. How to check if a citation is real is the manual routine, and our citation checker runs a pasted list through the existence and identifier checks at once. Neither can tell you whether a paper supports your sentence. Only opening it does.

The pattern under all of this is one thing. A citation assembled from text rather than from a source you opened carries whatever was wrong with the text, and carries it invisibly. That is the pairing CiteOwl is built on: the agent searches OpenAlex, CrossRef, Unpaywall and the web, reads the papers it retrieves, and writes each claim from a passage it keeps beside the sentence, so the quote supporting the line and the record the identifier came from are both there when you review it. Every edit arrives as a diff you accept or reject.