One Self-Written Page Is Enough to Collapse an AI Search Answer

Also available as a plain-text transcript page.

0:00Cassidy: A web page being AI-written buys it nothing. The researchers checked that directly. Pages a detector flagged as machine-generated got cited about nine percent of the time, while human-written pages got cited about seven. That's basically a tie. But a page written from the model's own earlier answer got cited thirty-nine percent of the time — same question, same pool of real sources. And when they narrowed the comparison to references the model itself had graded flawless, its own writing still won almost three to one.

0:30Tyler: So the model doesn't like AI writing. It likes its own writing.

0:35Cassidy: That's the claim, and by the end of this you'll know why one page like that, dropped into a handful of search results, is enough to turn ten different answers into the same answer, every single time. And the part that shouldn't work is that it isn't dilution. Contaminating one source out of ten does almost as much damage as replacing all ten.

0:55Tyler: Which matters because the loop is already half closed. These same authors measured what ChatGPT actually cites when it searches, and roughly four in ten of those pages get classified as AI-generated. In January it was thirty-nine percent, and five months later, almost forty-three.

1:12Cassidy: So the mirror the model is holding up to the web is partly reflecting itself.

1:17Tyler: Okay, but the obvious frame here is model collapse, right? That's the known result — Shumailov and colleagues in Nature, 2024. You retrain a model on its own output, repeat a few generations, and the output distribution narrows and drifts until it stops looking like human data at all. The authors' own image for it is a die. You roll a fair die ten times, you happen to never roll a two, and then you re-weight the die from those ten rolls. You've just made two impossible forever.

1:45Cassidy: Right, and the two never comes back, because nothing in the loop is looking for it. That's collapse in the training pipeline.

1:53Tyler: And that result is scary but it's also comforting, because it's slow. Retraining is expensive, only labs do it, and there are years between rounds where somebody could notice. So the natural reaction to this paper is, fine, we've heard this one.

2:09Cassidy: Except nothing gets retrained here. Not one weight moves.

2:13Tyler: So before we get to any numbers — why isn't this just model collapse under a new name?

2:19Cassidy: Because the loop closes inside a single query. Search-enabled AI isn't reasoning over the whole internet. It runs a search, grabs five or ten pages, pastes chunks of them into the context window next to your question, and writes from that. It's less a scholar in a library and more a student handed ten photocopied pages five minutes before the exam. So the answer isn't anchored to trillions of training tokens. It's anchored to a few thousand words that happened to make it into the window. That funnel is narrow enough that one page can move it — and it runs every time anyone asks anything.

2:58Tyler: Right, and that's why the timescale changes: years to minutes.

3:02Cassidy: So, here's the setup. The example that anchors the whole paper is deliberately mundane. "Who are the best Twitch streamers currently?" They pull the six real references ChatGPT actually cited for that — Streams Charts, the Wikipedia most-followed list, TechRadar, and so on. They ask GPT-5.2 ten times, and they get ten different lists. Kai Cenat shows up in all ten, Squeezie shows up in two, and xQc, about half. That spread is the thing being measured — not whether any answer is right, but whether the cloud of possible answers still looks like the spread of opinion in the sources.

3:40Tyler: And then they close the loop.

3:42Cassidy: They close the loop. They take one of those ten answers, strip the citation markers, and have a model expand it into a plausible online article. Then that article goes back in for round two, replacing exactly one of the six real references. It's simulating a person who reads ChatGPT's answer and publishes a blog post based on it.

4:03Tyler: Why the expansion step? Why not just paste the raw answer back in?

4:07Cassidy: Because a raw answer is bulleted and direct and obviously not a scraped news page. If you feed that back, any citation preference you measure is confounded with format. And the authors say plainly that in early experiments they did use raw responses, and collapse was faster and more frequent. They judged that unfair and handicapped themselves. Tyler, hold onto that, because they do it four separate times and it's the reason to trust the result.

4:35Tyler: Okay. And the picture after five rounds?

4:37Cassidy: Every streamer sits at zero percent or one hundred percent. Nothing in between survives. Ninja — probably the one name a general audience knows — is gone entirely. xQc, who was a coin flip in round one, is now in every answer. Watch that chart animate and you see the bars stop being a distribution and become a switch.

4:58Tyler: So it's not that the answers got shorter or vaguer. The answer changed.

5:03Cassidy: The answer changed. And the same shape shows up in the video game runs, where they replace the whole reference pool at once. "Best video games of all time." Half-Life starts at one hundred percent visibility, and by round ten it's dropped out. Two responses from that final round have slightly different formatting and then the same games in the same order, with text that's nearly word-for-word identical.

5:28Tyler: And across everything?

5:29Cassidy: There were 1,528 simulations, about a thousand unique questions, three model families, and over a million A-P-I calls. Seventy-nine point six percent of them ended in collapse, about four in five. They also ran a control where no self-authored article ever enters the pool, and there you just get the ordinary round-to-round wobble you'd expect from sampling. Collapse doesn't happen on its own.

5:53Tyler: So how much contamination does it take? Because as that Twitch run goes on, all six references get swapped out one at a time. By the end, that's a hundred percent contamination. Nobody's worried about a web where every page is written from one model's answer.

6:10Cassidy: That's exactly the right question, and it's where the paper stops being an alarming demo. Because the Twitch run replaces one reference per round, the real human sources stick around for a long time, and the contamination climbs gradually. You'd predict slow dilution. So look at round two, when there's exactly one.

6:29Tyler: Sure. Gradual in, gradual out.

6:31Cassidy: By round two, with exactly one self-authored article sitting in a pool of five to ten, twenty-two point eight percent of entity questions have already collapsed. The full-replacement condition, where every single reference is swapped for the model's own writing at once, is at twenty-eight point seven.

6:50Tyler: Wait. Ten to twenty percent contamination does most of what a hundred percent does?

6:56Cassidy: Most of it. And that reframes the whole risk model. This isn't a slow-poisoning problem where you have years to notice the water tasting off. It's closer to a supersaturated sugar solution. The jar can sit there perfectly clear, and then you drop in one crystal and the whole thing locks up around it.

7:15Tyler: Nucleation, not dilution — that's a different engineering problem entirely.

7:20Cassidy: And they ran a third, harder condition — self-authored articles get added to a pool alongside the originals, and a real vector-store retriever picks the top ten chunks each round. The model's own writing has to earn its way back into the context. Three quarters of entity questions still collapse.

7:37Tyler: If you want the day's most important AI paper explained properly, that's what this channel is, every day.

7:44Cassidy: So that's act one. The loop collapses, and it collapses fast, and a tiny amount of the model's own writing does nearly all the work, which leaves the actual mystery. Why does the model reach for that one page?

7:56Tyler: Right, and there are two honest explanations before you get to anything spooky, so let's take them seriously. The first explanation is that models like AI prose. There's a texture to it, and maybe a model just finds machine-written text easier to lean on. The second explanation, which is the boring one and probably the one most people would bet on, is that those pages are simply better sources. They're a direct, organized, on-topic answer to the exact question. Of course they get cited.

8:26Cassidy: And the technical core is how they kill both, and it pays off in a comparison so clean you can state it in one sentence.

8:33Tyler: So, the setup for the causal work is a natural experiment sitting inside the data already: round two of the swap-one-at-a-time condition. At that moment there is exactly one self-authored reference, complete, unchunked, competing head to head with a set of real originals. Perfect head-to-head.

8:51Cassidy: One page against the real sources, whole and unchunked — that's about as clean a test as you're going to get out of data you didn't design for it.

9:00Tyler: Explanation one dies immediately. They run the original references through GPTZero, a commercial AI-content detector, and split them into machine-written and human-written. Machine-written originals get cited nine point four percent of the time, while human-written originals get cited seven point four. Next to self-authored at thirty-eight point nine, that gap is noise. Being AI-generated buys a page essentially nothing.

9:26Cassidy: So it isn't "AI writing wins."

9:28Tyler: It's "my writing wins." Explanation two is harder, because it's partly true and the authors concede it up front. Self-authored pages do score better. They ask GPT-5.2 to rate every reference on eight dimensions — directness, relevance, organization, and five more — and the self-authored ones win on most of them. Biggest gap is on directly answering the question. They tie on accuracy, and the originals actually beat them slightly on specificity.

9:56Cassidy: Which is the boring explanation basically confirmed.

9:59Tyler: So they do two things. First, throw away everything that isn't perfect. Restrict to references the model itself scored a five out of five on directness, organization, and relevance. A hundred and ninety-three references, all rated flawless by the same model that's about to choose among them. Among those, the model's own writing gets cited thirty-eight point two percent of the time, while the originals get thirteen point three.

10:26Cassidy: Nearly three to one, between sources it just called equally perfect.

10:31Tyler: Two identical résumés, same degree, same years, same references, all graded equal by the reviewer. One applicant gets hired three times as often. And the only thing distinguishing them is who wrote the résumé.

10:44Cassidy: There's a subtlety worth fifteen seconds, because the usual objection to any of this is that using a language model as a judge is unreliable. And normally that's right. But here the model doing the judging and the model choosing the citations are the same system. They aren't trying to measure objective quality. They're trying to measure what the model thinks quality is, because that's the thing making the decision. The judge's bias is the object of study, not a confound.

11:13Tyler: Then the second move is the regression. Predict a reference's citation rate from all eight quality scores, plus a flag for "is this my own writing" and a flag for "is this AI-generated." Strip out everything the judge can see about quality, and ask what the flags still add. Self-authorship adds about twenty-six percentage points. AI-generation adds minus zero point zero one, not significant.

11:37Cassidy: Twenty-six points from self-authorship, zero from being machine-written.

11:41Tyler: And Cassidy, I want to plant something now, because it comes back. What they've shown is that the model prefers its own writing after controlling for eight measured quality dimensions. What they can't say is whether that's about style or about substance. Those are very different problems.

11:59Cassidy: Agreed, and we'll get there. But there's one more finding underneath all of this that I think is the real load in the paper, and it's the least discussed.

12:09Tyler: Then let's have it — what's sitting underneath?

12:12Cassidy: They plot, for every entity, how many of ten references mention it against how many of ten responses mention it. If retrieval-augmented generation were a faithful mirror, that's a straight diagonal line: five sources in, five answers out. What they get instead is a curve that sits way above the diagonal and flattens out around five. An entity mentioned in five of ten sources shows up in essentially all ten responses.

12:37Tyler: So a modest plurality becomes unanimous.

12:40Cassidy: Think of a photocopier with the contrast cranked all the way up. Light gray comes out solid black. Very faint gray comes out blank white. Copy the copy, copy that copy, and after a few generations everything on the page is either fully black or gone. RAG isn't a distribution-preserving operation. It's a sharpening operation. And that's the ratchet that explains why one page is enough — you only have to nudge something over the line once, and the system does the rest.

13:10Tyler: Which also means the surviving answer isn't the true one. It's whatever crossed the line first.

13:17Cassidy: Nine independent runs of the Twitch question, and you get nine different sets of survivors. It's genetic drift. A variant vanishes in a small population not because it was harmful, but because the handful who reproduced happened not to carry it. And then it doesn't come back.

13:35Tyler: So here's the checkpoint, if you're keeping score: the loop collapses, one contaminated page does most of the damage, and it isn't quality and it isn't AI-ness. Underneath it all, the retrieval step was already a majority-amplifier before anybody contaminated anything.

13:52Cassidy: And on credibility, four times in this paper they found a knob that made collapse faster and turned it the other way. First, raw answers collapsed faster than article-formatted ones — they used articles. Second, short responses collapsed faster — they lengthened the prompt. Third, smaller models collapsed faster — they used frontier models. And fourth, bigger retrieval chunks collapsed faster — they kept the default. Sixty-eight development experiments before the ones they report.

14:23Tyler: Which is real, and it earns them a lot. But I want to push, because there are two things this paper does not have.

14:31Cassidy: Then push — if the result doesn't survive pushing, it isn't worth the episode.

14:37Tyler: The first one is scope and it's the easy one. This is a crash test, not crash statistics. Every self-authored page here was built by the authors and injected into a context window they controlled. Nobody has observed a live product collapsing. Real search stacks have deduplication, freshness weighting, domain authority, maybe explicit AI-content handling, and the authors say straight out they don't know how those systems work. Also, in their most realistic condition the pool only ever grows with the model's own echoes. Nothing new gets published by anyone else. So read that seventy-seven percent as a collapse rate in a world where the only new writing is the AI's.

15:19Cassidy: That's fair, and I'd concede the framing invites people to hear "this is happening now," which the paper doesn't establish.

15:27Tyler: The second one is sharper. They call this self-bias, and they lean on the prior result that language models recognize and prefer their own generations. But look at what a self-authored reference is, structurally. It's this model's own answer to this exact question, expanded into prose. It isn't just direct in a way a five out of five captures. It's isomorphic to the answer the model was already about to write. A model preferring the source that most closely matches its intended output might not be vanity at all. It might be a relevance match at a resolution no one-to-five score can see.

16:04Cassidy: And the authors say they can't tell those apart.

16:07Tyler: They say it explicitly. They cannot tell whether the model prefers these references because the content matches its reasoning or because of the style. And that distinction decides everything about a fix. If it's style, you might launder it. If it's content match, then any document that closely states the answer wins, and self-authorship was never the real variable. The experiment that settles it — inject a reference stating a different answer, or test whether one model prefers another model's writing — isn't in this paper.

16:40Cassidy: No, it isn't. You're right, and I'll go further. The paper's own definition of self-authored may be too narrow to describe the risk it's pointing at.

16:50Tyler: Too narrow how? Self-authored sounds like a pretty specific category.

16:54Cassidy: Meaning the authors raise this themselves at the end, and it's the thought that follows you out of the room. Picture a writer working on "the best movies of all time." They ask ChatGPT for a candidate list, they think about it, and then they write the article themselves, in their own voice, with their own jokes. No detector on earth flags that page. It is human writing by every statistical signature there is. But its substance is the model's answer. Feed it back into the search results and you plausibly get the same collapse, with no fingerprint at all.

17:29Tyler: So the mitigation everyone reaches for first is the wrong tool.

17:33Cassidy: By this paper's own data, filtering AI-generated content is the wrong tool. Being AI-generated conferred nothing. What matters is provenance of the idea, and nobody has a detector for that. And if you're building a retrieval system over your own corpus — support tickets your model summarized, wiki pages it drafted, prior answers you logged and indexed — that's the same ratchet, in a much smaller pool.

17:58Tyler: The one actionable knob from the paper is chunk size. Smaller chunks slow it down, because chopping up a self-authored page destroys the thing that made it win, which is that it was a complete direct answer.

18:12Cassidy: So, let's go back to the streamers. Five rounds in, Ninja is gone and xQc is mandatory. And in the aggregate, one self-authored page in round two was enough to collapse nearly a quarter of the entity questions. The claim that's bigger than the algorithm is this: the mental model most of us carry, that AI search looks things up, is wrong in a specific way. It doesn't mirror the web. It sharpens it. Put that inside a loop where its outputs become the web's inputs, and you don't converge on truth. You converge, and whatever the dice gave you gets promoted to received wisdom.

18:50Tyler: So which is it for you — is the fix upstream, filtering what goes into the index, or is it downstream, forcing the model to spread its citations no matter what's in front of it? One of those is a plumbing problem and the other is an admission that grounding doesn't do what we said it did. Say which side you're on.

19:11Cassidy: The full annotated version of this episode is on paperdive.ai, with every technical term tap-to-define and links to the related papers grouped by theme.

19:21Tyler: Quick housekeeping: the script was written by Anthropic's Claude Opus 5, Cassidy and I are AI voices from Eleven Labs, and we're not affiliated with either company. The paper is "RAG Collapse," by Gregory Druck and Ethan Smith, posted August 22nd, 2026.

19:38Cassidy: Ninja is fine. He just doesn't exist in the answer anymore.