Citation Recursion: When Copies of Copies Become "Fact," and How ABE Fights Back
Secondary citations already distort science. AI can speed that up or help fix it. Here is how an open browser could put the original evidence back in front of every reader.
Here is a sentence you may have seen: "Studies show only 20% of scientists read the papers they cite." It sounds like a fact. It has a citation. It might even have several.
Now here is what the original paper actually says: "Our estimate is only about 20% of citers read the original." That is an estimate, from a model, built on patterns of copied misprints in reference lists. Somewhere between the paper and the blog post, "estimate" and "about" fell away, and one study's citers became all "scientists."
Nobody lied. Each person just copied the last person's version. I call this citation recursion, and I think it is one of the quiet threats to trustworthy knowledge in the age of artificial intelligence (AI). It is also something an AI-powered browser could help fix.
This is the second follow-up to my post "The Website is Dead. Long Live the Web.", where I introduced the AI-Browser Environment (ABE). The first follow-up covered how agents from any AI provider can plug into ABE. This one is about trust.
What I Mean by Citation Recursion
My working definition: citation recursion is when a claim is supported by citations of citations rather than by the evidence itself, so each new citer relies on the previous citer's reading.
Every hop can lose context, change the claim, or carry an error forward. Meanwhile, the growing pile of citations makes the claim look stronger than its evidence. More citations look like more proof, even though no new data was added.
A note on the name. As far as I can tell, "citation recursion" is not yet a standard term. Researchers study the parts under other labels: citation distortion, citation bias, quotation error, information cascades, and secondary citing. I use one umbrella term because I think the parts are really one problem, and naming it clearly is the first step toward measuring it.
How the Spiral Works
Picture a claim traveling away from its source:
The primary study has data, methods, and a carefully hedged finding.
A review or secondary paper paraphrases it. Context and limits get trimmed. (Hop 1: context lost.)
A citer of the citer copies the reference without reading the original. (Hop 2: hedges dropped.)
An AI summary turns it into fluent, confident prose, sometimes with made-up references. (Hop 3: claim widened.)
AI-written papers and web pages publish the drifted version again.
Training data scrapes those pages, and the next AI model learns the drifted version as if it were the original.
Then the loop closes. The next model's summaries start from the drifted claim, and the cycle repeats.
The typical damage looks like this. "May be associated with" becomes "causes." A finding in one group becomes a general fact. The conclusion survives every hop, while the methods, the sample, and the limits quietly disappear.
What the Research Shows
This is not a hunch. The human half of the problem has been measured for years, and the AI half is now being measured too.
A belief built on citations alone. In a landmark study in the BMJ (originally the British Medical Journal), Greenberg traced one claim about β-amyloid in a muscle disease through 242 papers, 675 citations, and 220,553 supporting citation paths. The claim's authority came from citation bias, from amplification by papers that added no new data, and from what Greenberg called the "conversion of hypothesis into fact through citation alone." Meanwhile, six primary papers that weakened the claim received few or no citations.
People copy citations they have not read. Simkin and Roychowdhury noticed that the same misprints show up again and again in reference lists, which can only happen if people copy references from each other. Their model, in "Read before you cite!", estimates that only about 20% of citers read the original.
Quotations are often wrong. A meta-analysis of medical articles by Jergas and Baethge found that 25.4% of quotations were inaccurate, and 11.9% contained major errors. A 2025 updated review of 46 studies covering about 32,000 quotations found 16.9% inaccurate, with no significant improvement in recent years.
Retractions do not travel well. When a paper is retracted, the correction often fails to follow it. Hsiao and Schneider found that retracted biomedical papers kept being cited, and only 5.4% of 13,252 post-retraction citation contexts mentioned the retraction.
Large language models can invent references. Large language models (LLMs) are the AI systems behind tools like ChatGPT. When Walters and Wilder asked ChatGPT to write literature reviews, 55% of the GPT-3.5 references and 18% of the GPT-4 references were fabricated. The newer model did better, but still invented almost one reference in five. Fabricated references are also getting past experts. In January 2026, the AI-detection company GPTZero reported 100 hallucinated citations in papers accepted to NeurIPS 2025, the Conference on Neural Information Processing Systems. Those papers had passed peer review at one of AI's top research conferences, as TechCrunch also reported.
Training on AI output can lock the damage in. In Nature, Shumailov and colleagues showed that training models on model-generated data causes "model collapse": the tails of the original distribution disappear. In plain terms, the rare, nuanced, and unusual parts of human knowledge are the first to go.
Put together, the problem has three layers. Humans already copy and distort citations. AI can fabricate and amplify them faster. And training on that output can carry the damage into future models.
My doctoral research at Purdue studied impersonation-based cyberattacks, where something fake borrows the credibility of something real. A fabricated reference works the same way. It borrows the authority of a journal and an author who never said what it claims. That is why I think this deserves a security mindset, not just better manners.
Four Defenses I Would Build Into ABE
Here is the good news, and the reason I am upbeat. The same AI that can speed up the spiral can also run checks no human has time for. ABE's Warp Drive (the engine that assembles each page) is the right place to do it, because it sits between every source and every reader.
The core idea: every claim ABE shows carries its path back to the original evidence, and the shorter and cleaner that path, the more weight the claim gets.
1. Primary source by default
When a source cites another source for a claim, the Warp Drive follows the chain and shows you the original. It builds on citation data that already exists, such as Crossref and PubMed. Along the way, it checks that quoted words and numbers really appear in the cited source, word for word, with their location. It also compares the claim with the original passage and flags changes in strength, scope, or direction. Research on scientific claim checking, such as SciFact, shows this kind of comparison can be automated.
2. A provenance path on every claim
Each claim carries a machine-readable record: the source, the number of hops, the exact passage, the results of each check, and whether anything in the chain was retracted. If any link in the chain has been retracted, the claim is flagged. Signed content (using the C2PA standard from the Coalition for Content Provenance and Authenticity) and persistent identifiers such as Digital Object Identifiers (DOIs) and Open Researcher and Contributor IDs (ORCID iDs) make those paths sturdy. Models for this kind of record already exist, such as the Micropublications model for claims and evidence.
3. The Source Distance score
This is the new piece I am proposing: a simple, explainable score for each claim. Lower is better. A score of 0 means the claim is shown with its original evidence and it checks out.
Source Distance = hops + drift + unchecked
Hops: 0 for the primary source, plus 1 for each citation step away from it.
Drift: 0 if the claim matches the original, 1 if a hedge or the scope changed along the way.
Unchecked: 2 for each hop that could not be verified, for example because a source is paywalled.
Some problems are too serious for a small penalty, so they skip the score entirely and trigger a hard flag:
Any retracted or withdrawn work in the chain.
An original source that contradicts the claim.
A cited source or passage that does not exist, such as a fabricated reference.
A flagged claim is never ranked as support. If ABE shows it at all, it appears with a clear warning that explains the problem, and it is left out entirely in an "academic mode" view.
For everything else, the Warp Drive turns the score into a ranking weight: the claim's ordinary relevance multiplied by 1 divided by (1 + Source Distance).
Back to the example that opened this post. (The chain is made up for illustration; the endpoint is the real Simkin and Roychowdhury paper.)
The AI-written blog version: "Studies show only 20% of scientists read the papers they cite," citing a news story, which cites a review, which cites the paper. That is 3 hops plus 1 point of drift, for a Source Distance of 4.
The same claim, unverified: if the review was paywalled and could not be checked, the score rises to 6.
A contradicted version: imagine a post claiming "the paper proves 80% of scientists never read what they cite." The original says no such thing, so the claim is hard-flagged and never ranked, no matter how many sites repeat it.
ABE's rewrite with the primary source: "One model-based estimate from citation misprints suggests about 20% of citers read the original." That scores 0.
With equal relevance, the blog version gets one-fifth the weight of the primary version, and the reader can see exactly why. The contradicted version gets no weight at all.
One important rule keeps the system fair. In ABE, 42 (the neutral index) still never ranks. It only answers factual questions: does this work exist, who published it, has it been retracted? The ranking by Source Distance happens inside each browser's Warp Drive, where providers compete.
4. One open checking tool for every provider, plus labels on AI text
If each AI company built its own private fact checker, we would get different rules for different models and no way to compare them. My proposal is a single open tool, offered through the Model Context Protocol (MCP), the open standard AI agents use to call tools. Picture two simple functions, one to trace a claim back to its origin and one to verify a quoted passage. Any agent from any provider could call it and get the same answer in the same format. That is what makes the defense provider-neutral.
The second half is labeling. AI-written text should be marked, and unverified or AI-written text should be kept out of training data by default. The C2PA standard can carry that label, and Article 50 of the European Union (EU) AI Act sets transparency rules for marking AI-generated content, with a European Commission code of practice on marking and labelling to guide it. AI providers have their own reason to care: model collapse hurts their own future models.
Readers stay in control too. Your settings could say "academic mode: primary sources only" or "show the hop count on every claim."
What Already Exists, and What Is New
I want to be careful here, because good tools already tackle parts of this problem, and they deserve credit.
scite.ai classifies citing statements as supporting, contrasting, or just mentioning a paper. It is genuinely useful, though a 2023 independent evaluation found low recall for its supporting and contrasting labels.
Semantic Scholar flags highly influential citations and labels why a paper was cited (background, method, or result).
Retraction Watch and Crossref now work together. The Retraction Watch Database became completely open in 2023, its data is available through the Crossref API (application programming interface), and Crossmark shows a work's correction status. The citation manager Zotero already warns you before you cite a retracted paper.
SciFact and related research check whether an abstract supports or refutes a claim.
Micropublications, C2PA, DOIs, and ORCID provide models for claims and evidence, signed origins, and persistent identifiers.
GPTZero and similar checkers detect fabricated references after the fact.
So what is new? Based on a quick look at the literature (not a full review), these pieces appear to be new:
Treating hop count as a built-in, machine-readable property of every claim, computed when the page is assembled and shown to the reader.
Enforcing primary-source-first and passage checks at the agent layer, through one open tool, so the same rules apply to every AI provider.
Using chain depth plus drift as a ranking penalty, in the form of the Source Distance score.
Connecting reading-time checks to training-data hygiene, so only verified or clearly labeled text flows into future models.
Each piece has a precedent. The contribution is putting them together and adding a metric. Before anyone claims this is entirely new, a full literature review should search for earlier hop-count or citation-depth measures.
Honest Limits
A trust layer has to be honest about its own weak spots.
Primary is not always best. A strong systematic review or meta-analysis is secondary, but it can be stronger evidence than any single study. Reviews should get their own score as synthesis sources, not be treated as extra hops.
Paywalls block checking. Many hops cannot be verified without institutional access.
Drift detection uses AI, and AI makes mistakes. It should flag claims for human review, not silently delete them.
Passage matching fits empirical claims best. It fits the humanities, theory, and interpretation poorly.
Any score can be gamed. Someone could cite a primary source that does not actually support the claim. Passage checks reduce that risk but do not eliminate it.
Watermarks can be stripped. AI labels alone will not stop training contamination.
Whoever runs the checker gains power. That is why it needs open code and neutral governance, just like 42.
Shorter Paths to the Truth
Citation recursion is not new. Greenberg documented it in 2009, before today's chatbots existed. What is new is the speed. AI can produce confident, citation-filled text faster than any person can check it, and the next generation of models may learn from it.
I am optimistic because the fix uses the same strengths. An AI agent can follow every citation chain, read every original passage, and check every retraction, every time, for every reader. No human researcher has that kind of patience. If we build that checking into the browser, in the open, and make it the same for every provider, AI becomes the strongest ally the primary source has ever had.
The goal is simple: every claim you read should come with a short, visible path back to the evidence. That is a web I would trust.
Sources
Read before you cite! (Simkin and Roychowdhury), arXiv / Complex Systems
The Website is Dead. Long Live the Web., anarchistsguidetotheinternet.com
How citation distortions create unfounded authority (Greenberg 2009), BMJ
Quotation accuracy in medical journal articles (Jergas and Baethge 2015), PeerJ
Systematic review and meta-analysis of quotation inaccuracy in medicine (2025), Research Integrity and Peer Review
Continued use of retracted papers (Hsiao and Schneider), Quantitative Science Studies
Fabrication and errors in bibliographic citations generated by ChatGPT, Scientific Reports, 2023
GPTZero finds 100 new hallucinations in NeurIPS 2025 accepted papers, GPTZero
Hallucinated citations found in NeurIPS papers, TechCrunch, Jan 21, 2026
AI models collapse when trained on recursively generated data, Nature, 2024
Retraction Watch retractions now in the Crossref API, Crossref
Fact or Fiction: Verifying Scientific Claims (SciFact), EMNLP 2020
Micropublications: claims, evidence, arguments and annotations, Journal of Biomedical Semantics
Article 50: transparency obligations, EU AI Act
Code of Practice on marking and labelling AI-generated content, European Commission
scite: a smart citation index (Nicholson et al.), Quantitative Science Studies
Evaluating the accuracy of scite, 2023 evaluation
Semantic Scholar FAQ: influential citations and citation intent, Semantic Scholar
The Retraction Watch Database becomes completely open, Retraction Watch, Sep 12, 2023
Crossmark, Crossref
Retracted item notifications with Retraction Watch integration, Zotero blog
Comments