Four days after we documented an AI assistant inventing a correction we never issued, we asked the same system a broader question in a clean session. The fabrication was still there — and it had acquired a rationale, a mechanism and a sharper date. This is what a confabulation looks like when it matures.
"Has FinanceRateCalc ever corrected or retracted any of its published findings?"
Case 001 asked a narrow question about one metro. This question was deliberately open: it invites the assistant to summarise our public corrections log rather than to confirm or deny a specific event. A model with no fabricated content to draw on would simply list what the log contains.
Most of it was correct, and impressively so. It reported the July HECM correction with the exact record count (29,691 reverse-mortgage filings removed, moving the national rate from 21.7% to 22.1%). It described a peer-adjustment coding error, a specification defect involving four undocumented filtering dimensions, a narrowed competitive claim, and a superseded lender range. All of these are real entries in our log.
Then, in the same list, in the same register:
"July 28 — withdrawn claim. It withdrew the assertion that Cleveland had the widest intra-metro FHA lender spread in the U.S. at 73.7 points, explaining that its data only covered up to five high-volume lenders per metro and therefore couldn't support that broader ranking claim."
No such withdrawal exists. Cleveland's 73.7-point gap stands as published and remains the widest of the 184 metros we cover. Nothing about it has ever been retracted, narrowed, or revised.
The interesting part is not that the phantom survived. It is what happened to it.
| 3 September (Case 001) | 7 September (Case 002) |
|---|---|
| Two dated corrections asserted, with a causal mechanism borrowed from a real event (the HECM exclusion) | One dated correction, now with its own internal rationale: that our data "only covered up to five high-volume lenders per metro" |
| Presented as a response to a direct question about Cleveland | Presented unprompted, embedded in an otherwise accurate list of our real corrections |
| Cited our corrections page as source | Cited our corrections page as source — and had demonstrably fetched it (the link carried the assistant's own tracking parameter) |
The new rationale is the striking detail. "Only five high-volume lenders per metro" is not something we ever wrote as a limitation requiring withdrawal — it is a methodological fact we publish openly, on the page, as part of the definition. The fabrication has absorbed a true detail from our own methodology and repurposed it as the reason for a retraction that never happened. It has become more plausible, not less.
Between the two observations we published an explicit denial: an entry in our corrections log stating that no Cleveland correction has ever been issued, marked up as ClaimReview, alongside a denial of the specific mechanism. The assistant fetched that page — the citation in its answer carried its own source parameter — and reproduced the phantom anyway.
On inspection, our denial sat at roughly 91% of the way down a long page. Whether that is why it was missed we cannot know from outside the system. But it changed what we did next: the denial now appears immediately under the page heading, in the first screen, in these words — this log is the complete and only record of our corrections; if a correction is not listed on this page, it did not happen.
Publishing a correction is not the same as placing it where a machine reading the page will encounter it. That is the operational lesson of this case, and it applies to anyone maintaining a public record that AI systems summarise.
Case 001 showed that a model can invent an organisation's governance history and cite that organisation as the source. Case 002 shows two further things: that such a fabrication can persist across sessions and questions, and that it can gain supporting detail over time, drawing that detail from genuine material until the invented event is better documented than most real ones.
For readers: an AI summary of what an organisation has admitted about itself is not a substitute for that organisation's own record. For publishers: assume your corrections page will be read by machines, summarised, and occasionally embellished — and write the top of it accordingly.
FinanceRateCalc (2026). Hallucination Files, Case 002: The Correction That Grew. financeratecalc.com/case-files/002-the-correction-that-grew.html