Corrections

A product whose claim is that it tells you what someone actually said has one job, and a correction is what that job looks like on the days we fail at it.

What we correct

We correct, without waiting to be pushed:

• Wrong attribution. Words placed in the mouth of someone who did not say them. This is the most serious error we can make and it is corrected ahead of everything else.
• Wrong transcription. Words misheard by automatic speech recognition, where the error changes the meaning.
• Wrong date, source or context that changes how a statement reads.
• Unmarked transformation. Translated, interpreted, written or shortened material presented as though it were verbatim speech. A shortened quotation presented as a whole one is a misquotation.

Due prominence

A correction is published where the error was seen. If a misattributed quotation appeared on a leader's page, the correction appears on that page, not only in the log below. If it was carried in a partner feed or an API response, we tell the partner.

Corrections are dated and written in plain terms. We name what was wrong rather than describing it as a technical issue.

We append. We do not overwrite.

A correction is added to the record. It is never written over the top of the mistake, and a published entry is never deleted. If a correction is itself wrong, we publish a further entry saying so.

This is not a stylistic preference. If the only evidence of what we once published is the thing we have since edited, then nobody, including us, can audit our own corrections. A record that can be quietly revised is not a record.

Suppression is not deletion

Where we cannot stand behind material, we suppress it: it stops being served, and it stays in the archive, correctly labelled, with a reason code and a ledger entry recording who suppressed it, when, and why. If the position changes, it is reversible the same day.

We prefer this to deletion because deletion destroys the evidence of our own decision. A reader who asks “why is this no longer here” deserves an answer better than silence.

What we mark, so it never needs correcting

Most of what would otherwise become a correction is prevented by declaring it on the page in the first place.

• Speech delivered in another language and rendered into English by an interpreter is marked as interpretation. Those are the interpreter's words, not the speaker's, and the page says so.
• Translations we perform are marked as translations, and the verbatim original is preserved and shown.
• Statements submitted in writing rather than spoken aloud are marked as written, and are counted separately from our spoken line totals.
• Where an excerpt is shortened, the shortening is declared in the same breath as the excerpt.
• Where a speaker cannot be identified with certainty, we publish nothing rather than guess.

Ask us to correct something

Email complaints@otamediagroup.com. You do not have to make a formal complaint to get a correction, and most corrections never become complaints. Our complaints procedure sets out the timescales and the escalation route if you want the formal process.

If you are the person quoted, say so and we will prioritise it.

Corrections log

1. 16 October 2026

   What we published: The European Parliament is held out of the record while we re-align it. A single passage of ours sometimes runs two speakers together, so it can carry one member’s name above another member’s words. We found it in every sitting we hold, not in a handful, which is why we are holding the whole source rather than correcting a list of passages.

   What we found: Found by us, on 16 October, by reading samples of our own record before changing anything in it. Nobody reported it. The European Parliament publishes its debates correctly; the fault is in how we divided them into passages.

   What we did: From 21 October 2026 all 29,757 European Parliament passages are held: they are not searchable and they do not appear on anyone’s page. Nothing is deleted and nothing is edited - every passage stays in the corpus and the hold is one update to reverse. Where a person’s page was built from those passages, the page keeps its address and says what is held rather than disappearing. The source returns when we have rebuilt the division into passages from the Parliament’s own record of who was speaking.

2. 12 October 2026

   What we published: A line attributed to Frank Bruno MBE was wrong in one word that mattered. Our transcript read “I had the dream of being the World Champion”. What he said, at four seconds into his Off The Agenda interview of 23 April 2021, was “had a dream of being heavyweight champion”. He was the WBC heavyweight champion, and the two titles are not the same thing, so the error put a claim in his mouth that he did not make.

   What we found: Found by us, on 12 October, by listening to the recording rather than reading the transcript. Speech to text produced it; nobody typed it. The line had reached a finished share card and two share videos, under the words “Search what people actually said”.

   What we did: The line is withdrawn, not edited. The wrong words stay on the row so the correction can be audited, and the row returns when the publisher transcript of that episode replaces it. The card and the videos are withdrawn from use. Nothing else from that interview is affected. We have also listed every line on every share card beside the record it quotes, so each can be checked against the audio before it is published, and we are checking them.

3. 9 October 2026

   What we published: A person’s page now shows only the lines whose speaker we can stand behind. 28,104 lines were removed from 310 pages. They were not labelled with anyone’s name, but they sat under a page headed with a person’s name, and a page is an attribution even when the line beside it is blank. The same rule that stops us putting those lines on a quote card now governs the page.

   What we found: Found by us, on 9 October, checking whether a line with no name shown could still imply a speaker somewhere else on the site. Search was already safe and so were quote cards. The page was not, and nor was the single editorially chosen lead quote, which has been removed.

   What we did: 276 of the 310 pages keep 88,433 lines and say how many are held. 34 pages now hold everything they have; each keeps its address, answers normally and says so, rather than disappearing. Nothing was withdrawn from the record and no line was deleted: all 28,104 remain in the corpus and remain findable in search, which shows no name against them. The way back is voice matching, which identifies who speaks when; a line returns to a page when it can name its speaker.

4. 9 October 2026

   What we published: The United States Congressional Record is held in full, 60,463 lines, while we rebuild how we read it. A reading of 50 lines drawn at random found 19 that were not the words of the member named: mostly the printed text of bills, which the Record prints under whichever member was speaking nearby, and also a committee-meeting notice and a letter from outside signatories read into the Record.

   What we found: Found by us, on 9 October, by reading a random sample of the source rather than only the parts we had already repaired. The repairs made on 8 and 9 October hold: of 13 repaired lines in that sample, 12 read as clean floor speech.

   What we did: Held under one reason, reversible, with nothing deleted. The source returns after launch one kind at a time, and only when a fresh reading of 50 lines finds no wrong speaker at all. Congress is about 1.4% of the lines a reader can see, and we would rather hold a source than serve it at that error rate. Our United States figures are now zero until it returns.

5. 8 October 2026

   What we published: 26,212 lines from the United States Congressional Record were withdrawn because they were printed in the Record but not spoken by the member named. Eight kinds: statements inserted into the Extensions of Remarks section, which members submit in writing and never deliver aloud; the text of bills and amendments, which is drafting rather than speech; notices written by the clerk or spoken by the chair, such as “Mr. Daines submitted an amendment”, which is the Record describing a senator rather than quoting one; procedural headings; lists of names rather than words, which are the roll of a recorded vote or the senators who signed a motion; the text of funding tables and bill sections, such as a line of defence procurement figures; the words of whoever was in the chair, which the Record prints under the member speaking next; and fragments left over once the page furniture was removed, such as a page number or the tail of a severed sentence. Two cases were the plainest kind of error: an amendment submitted by Senator Thune was filed under Senator Tuberville’s name, and nine passages were one member’s words filed under another’s, because the Record’s own name label had been split across a page break.

   What we found: Found by us, on 8 October, auditing the United States record after a reader reported something unrelated. The ingest had treated every passage on a page as words spoken by the member whose contribution it sat under, which is true of floor debate and not true of the rest of the page.

   What we did: Each was withdrawn under its own reason so the eight kinds stay separable, and each is reversible. 9,206 further lines are real speech with the page header printed in front of them, or with another speaker running on after them; those were repaired rather than withdrawn, the words as first published are kept alongside, and every repaired line was read again from the corrected text. Repairing rather than withdrawing matters: the Senate whip really does say “I announce that the Senator from Illinois is necessarily absent”, and the Record then prints the vote under his name, so cutting the vote away keeps his sentence where withdrawing the passage would have destroyed it. Our United States holding fell from 86,675 lines to 60,463, a fall of 30.2%. A smaller true record is worth more than a larger one, and the public figures were re-measured as soon as the work finished. The total rose from the 25,047 first published on 8 October to 26,212 on 9 October, as each class was re-checked against the text that remained after the page furniture was cut rather than against the raw page, and a sample of every class was read before any of it was withdrawn. The reading is the part that matters. It kept lines a count would have discarded, and it also discarded a rule we were about to apply: a test for lists of names would have withdrawn a roll of fallen police officers that a senator may well have read aloud, so we replaced it with a narrower one.

6. 7 October 2026

   What we published: Nineteen lines from the Odd Lots podcast were shown as Zohran Mamdani’s own words. They carried no established basis for naming him as the speaker: the audio had been split into voices by machine, one of those voices had been guessed to be the episode’s subject, and that guess is not evidence that he said the words. His record page had already been withdrawn for exactly this reason on 3 October, so the page said nothing while the quotes beside it spoke in his name.

   What we found: Found by us, on 7 October, auditing our own display rules rather than that page. The rule that stops an unestablished line being presented as somebody’s words had been written for one narrow case and never generalised, so it was permitting 2,026,161 lines in principle, of which these nineteen had actually reached a reader through a cache of selected quotes.

   What we did: Withdrawn the same day. The rule now refuses any line whose speaker is a raw machine label and which carries no recorded basis, on every programme rather than a listed few, and it is applied when the quotes are read as well as when they are chosen, so a stale cache cannot serve what the rule refuses. Mamdani’s page serves no quotes and his withdrawal notice stands.

7. 4 October 2026

   What we published: On 3 October we withdrew 7,339 record pages whose lines no longer carried an established basis for naming a speaker. Each kept its URL and served a withdrawal notice instead of a name. Between 04:18 and 12:30 UTC on 4 October those pages were live again, showing names we had decided not to stand behind.

   What we found: Found by us, while checking how many of the withdrawn pages a new naming pass had restored. The answer was none, and the reason was that the withdrawal had been undone: our nightly job rebuilds the index of record pages by replacing it wholesale, and the withdrawal was stored as a flag ON that index, so the rebuild reset it. Nothing failed and nothing was reported. A derived cache was holding a decision.

   What we did: The withdrawals were re-applied the same day, and now live in a table the nightly rebuild cannot replace. The rebuild rehydrates the flag from that table and refuses to report success if fewer withdrawals are in force than are on record, so a lost decision stops the job rather than passing quietly. 7,273 pages are withdrawn as of 4 October; the small change from 7,339 is a re-measurement, not a reversal.

8. 22 September 2026

   What we published: We published an interview with Dr Joanna Abeyie MBE under the name “Dr. Joanna Abey MBE”. Her name is Abeyie, which is the spelling that carries her MBE and her business. The misspelling stood on all 56 lines we hold of her and on the title of the episode.

   What we found: Found by us while applying publication dates supplied from our own records: the name in the supplied file did not match the name in the corpus, and the corpus was the one that was wrong. The cause is that her name was typed once, at ingest, and nothing downstream had any way to know it was not how she spells it.

   What we did: Corrected on every line and on the episode title the same day. The original spelling is kept in full beside the corrected one, on every row, in a record that cannot be edited or deleted. This is not a cosmetic fix: a name held in a form nobody would type makes a person’s own record unreachable to the person most likely to look for it, and they would have no way of telling that we hold it at all. We are now checking the supplied names for every guest of this programme against the names the corpus holds.

9. 22 September 2026

   What we published: We published an interview with Sir Damon Buffini in which the transcription misspelled the names of his own firm, his former firm and the estate he grew up on. Permira appeared as "Pamira", "Pomera", "Pera", "Meira" and "Premier"; Schroder Ventures appeared as "Short Adventures", "its show adventures", "Showdentures" and "Schroeder Ventures"; Thurnby Lodge appeared as "Thurnby at Lodge"; and the council estate where he was raised appeared as a "council state" and a "councillor estate". Nineteen such errors across ten passages.

   What we found: Reported internally by Edwin Danso and corrected the same day. These are errors of NAMED ENTITY rather than of meaning, and that makes them worse rather than milder: a misspelled firm name is a claim about what a man said, on a page he is likely to read. The cause is automatic speech recognition, which transcribes an unfamiliar proper noun as whatever common words it most resembles, and nothing downstream had ever asked whether the names in a transcript were the names the speaker actually used.

   What we did: All ten passages were corrected and re-embedded so that a search for the correct name now reaches them. The original transcription of every one is kept in full alongside the corrected text, with the list of substitutions that produced it, in a record that cannot be edited or deleted: a correction here is appended, never written over the thing it corrects. We are now running the same check across all 26 guests of this programme, seeded with the names each of them actually uses rather than with names a machine has guessed at. Several further errors in the same interview are held and not corrected, including two people’s names, because we would be guessing at them and guessing is the fault we are repairing.

10. 8 September 2026

    What we published: We published 2,450 passages of United Nations Security Council speech attributed to 25 members of the House of Lords and House of Commons who did not say them. The largest was 781 passages attributed to Lord Fowler, dating from 1995 to 2000. Those words were spoken by Robert Fowler, Canada’s Permanent Representative to the United Nations, and one of them begins "Canada warmly welcomes the Secretary-General’s presentation". A further 200 passages across six members are UN speech whose speaker we have not yet confirmed either way.

    What we found: Found by us, not reported by a reader, while drawing the first cases for a public benchmark we are building on attribution accuracy. The very first case drawn was one of these passages. The cause is that our records identify a speaker by an internal key such as "uk-315", and that key says who a speaker was taken to be, never where the words came from. Every check we had asked the first question and none asked the second. The passages themselves were matched to a British politician on a surname alone, which is exactly the match our live search already refuses to make when a user asks for one. That rule had never been applied to the records as they were written.

    What we did: All 2,650 passages were removed from every search result and profile page on the day the error was found. The 2,450 matched on a surname are recorded under a reason code marking the attribution as false. The remaining 200 are held, not restored, until each is checked against the United Nations’ own record of who was speaking, which names the country: some of these are genuine, since British ministers do address the Council, and we will not return any of them on an assumption. We have also changed the search function that served them so that it refuses United Nations material for a named politician even after the records are corrected, which a companion function had always done and this one never had.

11. 3 September 2026

    What we published: We published 28 passages of podcast speech attributed to Andy Burnham, from an episode of The Rest Is Politics titled "533. Andy Burnham’s Big Gamble: Can He Beat Reform?". The episode is a discussion about him between the presenters; he does not speak in it. The words we attributed to him include advertising read by the presenters and one passage in which a presenter introduces himself by his own name. The passages were served from 14 May 2026 until the day of this correction.

    What we found: Found by us, not reported by a reader. While verifying the figures in an internal planning document, one line about this show could not be sourced, and checking it surfaced the misattribution. The cause is the same one recorded in the 27 August entries: an episode about a person can pass checks designed for an interview with them. This episode was attributed under an older naming convention that predates the safeguards added since.

    What we did: All 28 passages were removed from every search result and profile page on the day the error was found, recorded internally under a reason code marking the attribution as false. We are now re-checking every existing podcast attribution for UK and devolved politicians against the same fault before any further podcast material concerning them is published.

12. 27 August 2026

    What we published: We published 138 passages of podcast speech attributed to Amy Coney Barrett, across two episodes. One was an episode of rePROs Fight Back titled "SCOTUS with Amy Coney Barrett and the ACA and LGBTQ Rights, Oh My!" (54 passages). The other was an episode of Conversations with Consequences titled "Ep. 77 Senator Thom Tillis & Erika Bachiochi on Supreme Court Nominee" (84 passages). Justice Barrett does not appear in either. Both are discussions about her nomination between hosts and other guests, and the words we attributed to her were spoken by other people.

    What we found: Found by us during an internal audit, not reported by a reader. Our naming step checks whether a person is named aloud and whether one speaker dominates the recording. Both are true of an episode ABOUT someone, so a discussion about a person can be mistaken for an interview with them. The second episode was found only because we checked the first correction before publishing it: her profile still showed 84 lines after the first removal, and 84 is not zero.

    What we did: All 138 passages were removed from every search result and profile page on the day the error was found, and are recorded internally under a reason code marking the attribution as false rather than merely out of scope. We also found that five internal lookups could still return removed passages; that has been fixed and verified on every path. The naming step that produced this is paused until we have a reliable way to tell an episode about someone from an interview with them.

13. 27 August 2026

    What we published: We published 23 passages of podcast speech attributed to Tony Allen, from an episode of Revolutions Per Minute titled "Tribute to Tony Allen". The episode is a tribute broadcast after his death, in which the presenters discuss him and his music. He does not speak in it.

    What we found: Found by us in the same internal audit, on the same day, and by the same cause as the entry above. No reader reported this.

    What we did: All 23 passages were removed from every search result and profile page on the day the error was found, and their removal is verified on every internal lookup. The naming step that produced it is paused.

OTA Media Limited, registered in England and Wales no. 16472914. Editorial responsibility: Edwin Danso, Chief Executive and Editor.