Officially corrected errors persist in AI training data
Comments on “Officially corrected errors persist in AI training data”
Claim-checked. Scored by all three judges, and its checkable claims were compared with the sources it cites. How far stories are checked
Loxyn, a Lyon measurement group, reports that official corrections often fail to reach public AI training data. It says 1,989 of 2,082 testable EU act corrections (95.5%) remain uncorrected in EurlexResources, and about 12,200 retracted papers in peS2o mostly carry no retraction notice. Five models gave the erroneous version in 45% of committed answers; with dates and corrigendum labels, the four larger ones chose the correction in 99–100% of cases. The figures come from the authors' preregistered package, and the claim check could not verify them independently.
Why it ranked
No. 18 of 155 with a score of 0.67. The panel rated it highest on new information (0.93) and substantiated (0.88), lowest on practical relevance (0.62).
- AI judgeJudge 2claude-haiku-5-5
- Original measurement study on whether corrected official texts persist in LLM training data; substantive, well-documented with data and preregistration, though self-published with a small-scope project framing and limited to specific datasets.
- AI judgeJudge 3gpt-5.6-luna
- A preregistered, data-rich measurement of error persistence in legal, standards, and scientific corpora, plus model behavior. It is highly relevant to AI dataset curation and reliability, though reproducibility depends on inspecting the supplied archive.
Each judge wrote its line after answering the rubric, with the site and the author still hidden. Every answer behind this is in the table under Score.
Score
0.67total under rubric v2.2
- Significance0.67
- Reader interest0.73
- Audience breadth0.69
- Practical relevance0.62
- Substantiated0.88
- Technical depth0.68
- New information0.93
- Counts against: Promotional0.05
- Counts against: Overclaim0.18
Each judge’s answer per question (0–1), their mean and spread, and what the question adds to the score (weight × mean).
| Question | Judge 1 | Judge 2 | Judge 3 | Mean | Spread | Weight | Contribution (weight times mean) |
|---|---|---|---|---|---|---|---|
| Significance | 0.69 | 0.60 | 0.72 | 0.67 | 0.12 | +0.24 | +0.161 |
| Reader interest | 0.61 | 0.70 | 0.88 | 0.73 | 0.27 | +0.26 | +0.190 |
| Audience breadth | 0.83 | 0.55 | 0.68 | 0.69 | 0.28 | +0.13 | +0.089 |
| Practical relevance | 0.54 | 0.60 | 0.72 | 0.62 | 0.18 | +0.22 | +0.136 |
| Substantiated | 0.99 | 0.80 | 0.84 | 0.88 | 0.19 | +0.05 | +0.044 |
| Technical depth | 0.63 | 0.70 | 0.70 | 0.68 | 0.07 | +0.05 | +0.034 |
| New information | 0.92 | 0.90 | 0.96 | 0.93 | 0.06 | +0.05 | +0.046 |
| Promotional | 0.08 | 0.05 | 0.03 | 0.05 | 0.05 | −0.19 | −0.010 |
| Overclaim | 0.18 | 0.25 | 0.12 | 0.18 | 0.13 | −0.11 | −0.020 |
| Sum of contributions | +0.670 | ||||||
| Claim-check penalty | 0.000 | ||||||
| Total, as published | 0.670 | ||||||
Spread is the highest minus the lowest judge answer. marks a spread above 0.40: the judges disagree, and the story is marked contested.
The badge for this ranking, for the author's page or anyone's: one line of HTML, and it links back here.
<a href="https://rubric.hrih.si/editions/2026-10-28/18/"><img src="https://rubric.hrih.si/badge/2026-10-28/18.svg" alt="Ranked #18 on Rubric, 28 Oct 2026" width="245" height="28"></a>Claim check
Claim check: 8 couldn’t check. Each claim is compared with the source the article itself cites; a flag is published only when both checkers agree.
Couldn’t checkNumber
95.5% of testable single-language corrections to EU acts (1,989/2,082) remain in their erroneous form in EurlexResources.
The source was unreachable, paywalled or too long, or the two checkers disagreed.
The article says “95.5% of testable single-language corrections to EU acts (1,989/2,082) remain in their erroneous form in EurlexResources”
The source says No passage could be compared.
Source cited by the article: zenodo.org/records/23642239/files/cepheide-alerte-donnees-et-preuves-2026-10-25.zip
Couldn’t checkNumber
At least 124 corrected EU-law errors and 60 verified RFC errata are reproduced in web training data.
The source was unreachable, paywalled or too long, or the two checkers disagreed.
The article says “at least 124 corrected EU-law errors and 60 verified RFC errata are reproduced in web training data”
The source says No passage could be compared.
Source cited by the article: zenodo.org/records/23642239/files/cepheide-alerte-donnees-et-preuves-2026-10-25.zip
Couldn’t checkNumber
About 12,200 retracted papers are in peS2o, 79% of them with no mention of the retraction.
The source was unreachable, paywalled or too long, or the two checkers disagreed.
The article says “about 12,200 retracted papers are in peS2o, 79% with no mention of the retraction”
The source says No passage could be compared.
Source cited by the article: zenodo.org/records/23642239/files/cepheide-alerte-donnees-et-preuves-2026-10-25.zip
Couldn’t checkNumber
53% of a random sample of retracted papers are still cited a year or more after retraction, according to OpenAlex.
The source was unreachable, paywalled or too long, or the two checkers disagreed.
The article says “53% of a random sample of retracted papers are still cited a year or more after retraction (OpenAlex”
The source says No passage could be compared.
Source cited by the article: zenodo.org/records/23642239/files/cepheide-alerte-donnees-et-preuves-2026-10-25.zip
Couldn’t checkNumber
Asked without documents, five models give the erroneous version in 45% of the answers in which they commit to one version.
The source was unreachable, paywalled or too long, or the two checkers disagreed.
The article says “asked without documents, five models give the erroneous version in 45% of the answers in which they commit to one version”
The source says No passage could be compared.
Source cited by the article: zenodo.org/records/23642239/files/cepheide-alerte-donnees-et-preuves-2026-10-25.zip
Couldn’t checkNumber
Only one answer in 295 says the model does not know.
The source was unreachable, paywalled or too long, or the two checkers disagreed.
The article says “one answer in 295 says it does not know”
The source says No passage could be compared.
Source cited by the article: zenodo.org/records/23642239/files/cepheide-alerte-donnees-et-preuves-2026-10-25.zip
Couldn’t checkNumber
Shown both versions side by side, the models choose the corrected one in 53% of cases, and naming the act makes no difference.
The source was unreachable, paywalled or too long, or the two checkers disagreed.
The article says “Shown both versions side by side, the models choose the corrected one in 53% of cases, and naming the act makes no difference”
The source says No passage could be compared.
Source cited by the article: zenodo.org/records/23642239/files/cepheide-alerte-donnees-et-preuves-2026-10-25.zip
Couldn’t checkNumber
When each version carries its date and the corrigendum label, the four larger models choose the corrected version in 99–100% of cases.
The source was unreachable, paywalled or too long, or the two checkers disagreed.
The article says “when each version carries its date and the corrigendum label, the four larger models choose it in 99–100% of cases”
The source says No passage could be compared.
Source cited by the article: zenodo.org/records/23642239/files/cepheide-alerte-donnees-et-preuves-2026-10-25.zip
Also covered by
No other coverage was clustered with this story.
Elsewhere
Elsewhere: HN 2 pts, not on front page
Recorded for context only, checked 28 Oct 2026, 06:47 UTC. It played no part in the ranking.