Russian State Media Used AI to Write News Scripts. No AI Detector Would Have Caught the Finished Broadcast.

Quick answer

Anthropic's September 2026 threat intelligence report disclosed that four individual accounts used Claude as an editorial desk to help produce content that ultimately aired on Russian state media, including Sputnik Moldova, RIA Novosti, Sputnik en Español, Sputnik Africa, and RT's English-language newsroom. Anthropic caught this by comparing the accounts' own Claude activity against what those outlets actually published, a comparison only the AI company itself could make. A public AI-text detector scanning the finished, translated, professionally-edited broadcast copy afterward would very likely not have flagged it. That gap is the real lesson here: detection works differently depending on whether you can see the raw draft or only the polished result.

What Anthropic actually disclosed

Anthropic publishes a periodic threat intelligence report on misuse of its Claude models; this was its fourth, following reports in March 2025, August 2025, and November 2025. The September 2026 edition covered activity disrupted between December 2025 and August 2026 and was widely covered by mainstream outlets including Reuters and the BBC, both of which independently confirmed the report's existence and its headline findings around cyber operations, weapons research, and influence campaigns.

One case study, labeled GTG-24015 in Anthropic's internal tracking, singled out state-media use. Anthropic wrote that it "identified and removed four accounts in which individual actors used Claude as an editorial and news production desk to distribute content via Russian state-media outlets." The company assessed with high confidence that outputs from those sessions were shared with, and ultimately published or broadcast through, Russian state-owned and state-funded media. Distribution ranged from a Telegram post that picked up roughly 2,000 views to material that made it onto Russian television airwaves in the form of tickers, screen captions, voiceover scripts, and short headlines.

How the pipeline actually worked

A former Sputnik Moldova editor-in-chief used Claude to turn Romanian and Moldovan news coverage, polling data, and opposition social media posts into Russian-language articles, which were then published on Sputnik Moldova's Telegram channel and on RIA Novosti and amplified across a network of aligned outlets. Anthropic says the same operator used this pipeline to amplify fabricated, defamatory claims about Moldovan president Maia Sandu ahead of the country's September 2025 parliamentary election.

A separate contractor pulled source material directly from Russian military-linked Telegram channels and used Claude to write Spanish-language articles for Sputnik en Español, working under the editorial oversight of an actual Sputnik Mundo presenter and producer. A third case involved an employee at a Russian state-owned outlet who used Claude to draft material destined for live broadcast, pulling from Russian newswires and government sources, that Anthropic confirmed reached Russian airwaves in at least one instance.

Anthropic's own framing of the significance is worth quoting directly: "Claude was integrated into an already running, professionally edited pipeline as the sub-editor layer, taking a single staffer's output well beyond what they could produce unaided." The company was also careful to disclose the limits of what it could confirm, stating plainly that it is "not able to determine what share of the outlets' total output passed through the pipelines that involved Claude."

Why an AI-text detector wouldn't have flagged the finished version

Here's the part that matters for anyone who works with AI-text detection, including our own users. The text that eventually aired or published was not the raw output of a single Claude session. By the time it reached an audience, it had typically been translated into a different language, folded in with real polling data and real news events, edited by a working newsroom, and in at least one case reviewed under the "editorial watch" of a named human producer before broadcast.

That sequence, translation plus heavy human editing plus blending with genuine source material, is close to a worst-case scenario for any tool that estimates AI authorship from the statistical properties of finished prose. Detectors built on signals like perplexity and burstiness are measuring how predictable and how varied the surviving sentence structure is; both of those signals degrade substantially once a human editor rewrites, trims, and localizes a draft, and they degrade further across a language boundary. A newsroom's own house style, applied on top of an AI first draft, is exactly the kind of transformation that pushes a detector's confidence toward "inconclusive" rather than "likely AI." None of this means detection tools are useless. It means they are reading a different signal than "was a large language model involved at some point," and a professionally produced state-media segment is designed, whether intentionally or not, to strip out the traces a detector looks for.

What actually caught it

Anthropic didn't catch this by running the published RIA Novosti article or the aired RT broadcast through a downstream AI-text scanner. The company's own account of the case describes matching "individual Claude-produced output against published content" — meaning Anthropic had the original prompts and completions from its own platform and could directly compare them, sentence by sentence, against what later appeared in public. That is a fundamentally different kind of evidence than anything a probabilistic text classifier can produce from cold text alone. It requires being the platform the content was generated on, with visibility into the account's actual usage, not just the finished result.

That distinction lines up with something we've written about before: a detection score reflects statistical likelihood, not proof, and it works best on text close to its original, unedited state. Once a document survives editing, translation, and organizational review, treating a detector's read on the final version as decisive misunderstands what it can measure.

What this means if you work with wire copy, press material, or unverified sourcing

If your job involves verifying suspicious content, the practical lesson isn't "detectors don't work." It's that the tool is best matched to text that hasn't yet been through a heavy editorial pass. A leaked early draft, a raw press release before a comms team rewrites it, or a tip submitted directly to a newsroom are all reasonable candidates for a first-pass AI-text check. A finished broadcast script that already passed through a professional editor in a second language is a much harder target, and a clean or ambiguous score on that kind of text shouldn't be read as evidence the content is genuinely independent reporting.

For newsroom fact-checkers, comms teams vetting inbound pitches, and researchers tracking coordinated messaging, running unedited source material through TheChecker.AI's free detector is still a useful early signal, one data point alongside provenance checks, source triangulation, and the kind of account-level pattern analysis that actually broke this story. It should never be the only check on a byline you don't recognize or a "grassroots" post that reads a little too polished, but as an honest, cheap first filter on raw text, it earns its place in the same toolkit journalists already use for source verification. Our own reporting on AI transparency legislation covers why disclosure rules haven't closed this gap either, and our explainer on why a single detector score is never proof applies just as directly to a suspicious wire story as it does to a suspected student essay.

FAQ

Did Anthropic say Claude wrote the propaganda itself? Anthropic described Claude as functioning at the "sub-editor" layer inside pipelines that were already professionally run by humans at Sputnik Moldova, RIA Novosti, Sputnik en Español, Sputnik Africa, and RT. Human editors, translators, and in one case an on-air producer remained part of the process from source material to final broadcast.

Would running the published Sputnik or RT content through an AI-text detector prove AI was involved? No. Once text has been translated, professionally edited, and blended with real reporting, the statistical signals a detector relies on are substantially weakened. Anthropic's own detection method here relied on comparing account-level Claude activity against published output, not on scanning the finished text cold.

Is this the same kind of AI-content problem TheChecker.AI addresses for schools or hiring? It's a related but distinct problem. Our detector estimates whether a piece of writing carries the statistical fingerprint of LLM generation, which is most reliable on text close to its original form. Catching a coordinated state-media pipeline required platform-level account monitoring that only the AI company running the model could perform, a capability outside what any downstream text detector, including ours, is built to do.