The Aporia
The Aporia
This is every mistake The Aporia knows it has made: what a reader would have seen, when it started, when it was noticed, and how long it took to fix.
It is part of About because a site that asks you to weigh up other people's reliability has to publish its own record, including the parts that are unflattering, and especially the interval between finding a problem and fixing it.
| Started | Noticed | Fixed | Live for | What was wrong | Kind |
|---|---|---|---|---|---|
| 2026-09-12 | 2026-11-06 | 2026-11-06 | 55 daysfixed the day it was noticed |
The front page labelled stories "leans left" or "leans right" from scores that were usually wrong.
What you saw: Story cards showed a leaning taken from per-article scores. Checked by hand against 120 articles, an article scored left really leaned left 23% of the time, and one scored right 32%. The errors ran one way: there were about 2.7 left scores for every right one when the real split was close to even, so the news looked more left-leaning than it was.
Why: The model tended to give an article the direction of its subject rather than of its sympathies: a New York Post story about a left-wing podcaster attacking JD Vance was scored left.
Fix: Stories now show which outlets covered them, placed by each outlet's average over 60 days, never one article's score.
Check it:
/about/methodology#outlet-positions
|
Method |
| 2026-09-11 | 2026-09-13 | 2026-09-14 | 3 days1 day after noticing |
Some story titles came out partly in Chinese.
What you saw: 11 of 5,468 story titles mixed Chinese text into an English title, two of them on the front page.
Why: The language model that writes titles sometimes drifted into Chinese, and nothing checked the title before it was stored.
Fix: A title in another script is now refused and written again. The 11 were rewritten.
|
Copy |
| 2026-09-05 | 2026-09-06 | 2026-09-06 | 1 dayfixed the day it was noticed |
The news stopped updating, and nothing on the site said so.
What you saw: From 5 September 2026 the news pages showed nothing newer than 04:21 UTC that morning, while looking like a normal, current front page.
Why: Each update was being cut off by a three-hour time limit before it reached the step that builds the stories readers see. Articles kept being collected, so from the inside everything looked busy.
Fix: The time limit was removed. Since November the admin console and an outside health check also warn when the published news is more than 12 hours old.
|
Access |
| 2026-09-04 | 2026-09-06 | 2026-09-06 | 2 daysfixed the day it was noticed |
The donation page got its own arithmetic wrong.
What you saw: The support page said "The goal is CA$500 a month, which is CA$2,000 across a year." CA$500 a month is CA$6,000 a year; the two goals were separate settings, and the sentence treated one as the other.
Why: A sentence restating two figures that were already shown above it, and joining them with "which is".
Fix: The sentence was removed. Both goals are still shown, as they were, with no claim about how they relate.
Check it:
/support
|
Copy |
| 2026-09-01 | 2026-11-11 | still wrong | 73 days |
Some CBS video articles were stored with another video's text.
What you saw: About 1 in 6 stored CBS video articles carry text that shares no word with their headline, so their summaries and analysis described a different video.
Why: On short video pages we stored the text of a neighbouring clip.
Fix: New CBS videos use the page's own description of the video. The summaries of the affected stored articles are hidden and their pages say the analysis came from the wrong text; the text itself has not been collected again yet.
|
Data |
| 2026-09-01 | 2026-11-11 | 2026-11-12 | 72 days1 day after noticing |
Some summaries described our copy of the article instead of the news.
What you saw: 87 article summaries said things like "the text does not contain the article", and 229 story summaries were written as a note to us rather than as a summary for readers.
Why: When the stored text was thin or wrong, the model said so in the summary instead of failing, and nothing caught it.
Fix: Those summaries are hidden, and a story summary of that kind is now refused and written again.
|
Copy |
| 2026-09-01 | 2026-10-02 | still wrong | 73 days |
Some summaries named people or numbers the article does not contain.
What you saw: About 1 in 10 article summaries included a name or a figure that does not appear anywhere in the article, and 184 summaries drifted into Chinese.
Why: Summaries are written by a language model, and nothing checked them against the article before they were stored.
Fix: Every new summary is now checked against the article's text and refused if it fails. The worst of the stored ones were redone on 2 October 2026; about 2,500 older ones have not been rechecked yet.
|
Copy |
| 2026-09-01 | 2026-09-10 | 2026-09-24 | 23 days14 days after noticing |
Subscription adverts were scored as political writing.
What you saw: 17 Reason / Volokh Conspiracy posts were labelled "leans strongly right" with high confidence. What we had stored for them was not the post but a 241-character advert for a Reason subscription, and the label rested on its words "Always independent".
Why: The page text we collected for those posts was the site's advert, and nothing refused text that was identical across many articles from one outlet.
Fix: Text repeated word for word across four or more articles from one outlet is now stored as empty, and those articles were analysed again.
|
Data |
| 2026-09-01 | 2026-09-10 | still wrong | 73 days |
Leaning scores counted other people's quotes as the outlet's own view.
What you saw: An article that quoted a politician could be marked as leaning that politician's way, because the quote was taken as the article's own wording. Articles scored since mid-September 2026 are checked; about 4,900 older ones still carry the old score.
Why: The check that a leaning rests on the article's own words did not tell the article's voice apart from the people it quotes.
Fix: Quoted and reported speech no longer counts as evidence of a leaning. Older scores are being redone a few dozen at a time.
Check it:
/about/methodology#leaning
|
Method |
| 2026-09-01 | 2026-09-06 | 2026-09-07 | 6 days1 day after noticing |
Unrelated articles were merged into one story.
What you saw: A story titled "Lindsay Clancy Murder Trial" held 67 articles, a quarter of them about other things, including 17 Gloria Steinem obituaries. The largest story mixed the Iran war, the Canada trade war and a proposal to rename Lake Ontario.
Why: For some NBC and Today video pages we stored the text of the previously played clip, so a story about one event carried another event's text. Grouping is transitive, so one wrong link joined two whole stories.
Fix: Video pages are now read correctly, and every story is checked for a weak point where two groups are only joined by a thread; those are split.
|
Data |
What gets recorded. Something this project got wrong that a reader could have seen and acted on: a measurement computed the wrong way, a page that stated something untrue, a signal published when it should have been withheld. Not every bug — a crash nobody saw and a slow query are not on this page. The test is whether somebody who trusted the site was misled by it.
Why three dates and not one. "A correction was published" says nothing about how long the wrong thing was live, or how long it sat known-broken before anyone fixed it. Those two intervals are the ones that say whether a project is actually careful, so they are columns rather than something a reader has to reconstruct. Where a start date is an estimate it is the honest estimate rather than the flattering one, and where a mistake was found by somebody else the Noticed column says "not recorded" rather than implying it was found in-house.
Kinds. Data — wrong or missing data; Method — a measurement computed the wrong way; Copy — a page that said something untrue; Access — something readers could not reach.
All times are UTC. Entries are never deleted or edited after the fact except to add the date a mistake was fixed.