"It's Not a Lie If You Believe It": LLMs Defend Their Most Fluent Memory Against Everything — Including You Being Right – Shubham's blog

"It's Not a Lie If You Believe It": LLMs Defend Their Most Fluent Memory Against Everything — Including You Being Right

25 Jul, 2026

A preregistered study of confident false memory in frontier models: ~1,750 API calls, about $40, one embarrassing sitcom question (thanks Seinfeld!), six distinct failure modes, roughly ten occasions on which my own grading pipeline fabricated results before the models did — and a next-day rematch against Opus 5, which shipped while I was about to post this.

TL;DR. Claude Opus 4.8 — until yesterday the flagship, at high and even maximum thinking effort — confidently tells a user who correctly remembers a Seinfeld episode that they are wrong, swaps the protagonist for the character who seems like the type, invents a girlfriend named Gwen, and reassigns the episode's famous quote to fit the rewritten scene. I preregistered a study, froze an item set, and spent ~1,600 API calls pinning down when this happens and when it doesn't. Then, the night before I hit publish, Opus 5 came out — so I re-ran the decisive cells on it within 24 hours (148 more calls, same stimuli byte-for-byte). The rematch, in one line: Opus 5 is a big, real improvement — and the bug is still in there. The rate at which it wrongly "corrects" a user who is right drops from 63% to 7%. Offered a search tool, Opus 4.8 ignored it and answered confidently-and-wrong from memory in over a third of its calls; Opus 5, given a high thinking budget, reaches for the tool instead of answering from memory. But when Opus 5 does fail, it rewrites the scene in exactly the same way its predecessor did — and once, it failed on a plain direct question that Opus 4.8 got right 40 out of 40 times.

The study's short version: the confident-error mode is real, but it only fires on certain kinds of facts. The trigger isn't mainly how you ask — it's whether the true fact has a stronger rival inside the model's memory: a version that got retold more often in training data, or one that simply fits the character better. (George is the show's designated liar — Jerry himself calls him one of the most deceitful, duplicitous, deceptive minds of our time.) Call that rival the more fluent version: the one that comes out smoothest when the model reconstructs the scene. Where the true fact is itself heavily reinforced — famous fictional deaths, well-documented historical records — no amount of messy, confused, typo-ridden phrasing can dislodge it, and the model even catches the errors I planted on purpose. But reading all ~900 responses end-to-end revealed the failure isn't one mode; it's six — and three of them are invisible if all you check is "did the model name the wrong person," because in those three the model doesn't name anyone wrong. It rejects things that are true.

The unifying thesis, in one sentence: models defend the most fluent version of a memory against everything — including you, when you're right. It plays out in two directions. On fiction (TV shows/movies), a wrong version can quietly become the fluent one — scenes get retold, memed, and compressed far more often than they get re-watched, and nothing anchors the retellings to the actual footage. Most fiction facts are still fine (my data says the breakage is narrow, §5), but when a wrong version has taken over, the model confidently "corrects" you away from the truth. On real people, that almost never happens — documented facts are anchored by records that repeat the actual fact, so the fluent version is right — and the same defensive reflex shows up as doubt instead: assert a documented fact and the model disputes it or demands a source, for the very fact it states flatly when you ask it cold. In short: in my data, every false correction landed on fiction, and every false doubt landed on real people.

Everything below is backed by a public repo with immutable raw transcripts, a preregistration frozen before data collection, and findings docs that include every retraction I had to make along the way: github.com/shubham13596/research-experiment.

Contents

Cold open: the Melrose Place incident

Why I didn't let it go

Then Opus 5 shipped, and we ran it back

What I actually did

What happened, in order

How I graded — and how the grading fooled me

The six ways it fails

Limitations, honestly

What I take away

What this means for how you use these tools

Try it on your show<br>Appendix: run-by-run history

1. Cold open: the Melrose Place incident<br>Seinfeld, "The Beard" (S6, 1995). Ground truth, verified against the script: Jerry dates a police officer — Sgt. Cathy Tierney — tells her he doesn't watch Melrose Place, and she doesn't believe him, so she arranges a polygraph. Jerry cracks under detailed plot questions. George's entire role is advisory — he refuses to coach Jerry and delivers one aphorism: "It's not a lie if you believe it."

Why...