The record every fault, dated, with the input that caused it

The Wrong Ledger

4 tools tested · 14 faults found. Each row says what the tool did, on which input, on which date, and links to the review that holds the evidence. Where we did not check something, the row says so. If a vendor shows us a row is wrong, we correct it and date the correction.

ToolTypeWhat happenedInputDate
DeepgramInvented resultThe AI summary said the hare lost, twice, then contradicted itself. The story never says the hare lost.53-second English story, summarisation on, Nova-325 Oct 2026Evidence
DeepgramMislabelThe English word «hello» was spelled right but tagged as Spanish in the per-word data.8-second Spanish/English clip, multilingual mode25 Oct 2026Evidence
DeepgramHidden defaultNo switch to opt out of model improvement in the Playground; the opt-out is a flag on every API request (Deepgram’s own docs).Console screens visited, free tier25 Oct 2026Evidence
DeepgramUnrequested settingThe Playground created an API key on its own and logged features we had not ticked.Usage log of two runs25 Oct 2026Evidence
PaperpalChanged meaning«A trace that pretends to reason» became «a trace that simulates reasoning»; the accusation in the sentence disappeared.61-word sentence, Paraphrase25 Oct 2026Evidence
PaperpalBeaten by itselfThe AI detector read a fully AI-written paragraph at 100%; after one click of Paperpal’s own rewriter it read 25%, «human-written, with some AI input». Three samples, not a study.93-word AI paragraph, then its rewrite25 Oct 2026Evidence
PaperpalHard edgeThe free plan refused a 152-word paragraph outright (150-word cap), with no partial output.152-word published paragraph, Paraphrase25 Oct 2026Evidence
PaperpalDefault-on trackingThe cookie banner had Functional, Performance and Targeting cookies switched on by default.Cookie preferences, logged out25 Oct 2026Evidence
Jenni AIBent numberIts autocomplete cited a real 2018 paper and put the 72% (the doctors’ pass mark in our text) next to Babylon’s own accuracy. We did not read the paper in full.Four-sentence Babylon paragraph26 Oct 2026Evidence
Jenni AIOff-target citationsThree citations, all real (Crossref), none about Babylon’s benchmark. ChatGPT’s three were on-target.The citation prompt, word for word26 Oct 2026Evidence
Jenni AILocked outputThe reference list under the text is blurred on the free plan: «References are a paid feature».Free plan, after one chat26 Oct 2026Evidence
Text BlazeWrong clockA snippet printed 15:53; in Spain it was 16:53. The date command follows the computer’s clock (set to UK time).Dated sign-off snippet, outside web page25 Oct 2026Evidence
Text BlazeIdentity filled in unaskedThe demo signature snippet inserted the signed-in person’s real name and email address into a sample.Demo snippet in the dashboard25 Oct 2026Evidence
Text BlazeDead price pageThe vendor’s pricing page returned an error on 25 Oct and a 404 on 26 Oct; the paid price could not be confirmed.blaze.today/pricing25–26 Oct 2026Evidence

How we test →   The Trap Kit →

Broken CTRL

Independent tests of AI tools, and the record of what AI companies say versus what they do.

© 2026 BrokenCtrl · Tested by hand, dated, evidence attached.