The New York Times shipped text no editor read, and told its union at the bargaining table

Search was a sensible place to put this if the point was for nobody to notice. It is the product Times employees themselves single out as trailing everything else the paper ships ^[18], and it is now the first place the news operation has published text that no journalist or editor passed ^[3]. The AI work that came before it sat elsewhere: Wirecutter Finder draws summaries and tips from the review site's own recommendations ^[15], and newsroom staff have had internal tools for their own work for some time, behind the reporting rather than in front of the reader ^[16]. What moved is the boundary between machine text the newsroom uses and machine text the reader is served ^[4].

The narrow scope helps, and then stops helping. Because the summariser reads only Times journalism and does not reach the open web ^[6], it cannot invent a fact about the world. It can still misstate what the reporting said ^[7]. That is the failure the paper's editing layers exist to catch before publication, and Semafor's argument is that the institution's standing rests on that catching rather than on never being wrong, since it already makes mistakes and appends corrections ^[19]. A summary composed when the query arrives has nowhere to put that step ^[3].

So the Guild's two AI clauses are not equal weight. The proposed pool splitting 22.5% of the paper's AI licensing deals among unionised staff ^[10] is a price, and management can pay a price while keeping what it has built. The human-oversight clause is a stop order: it would make the search tool impractical to run ^[11]. That turns the table into an argument about withdrawing a live feature rather than approving a planned one ^[2]. Jim Luttrell, the unit chair and a senior staff editor, says the current policies are meaningless and that management is violating them with the summarisation tool ^[12], and grounds his case for contract language in the record of deployed AI summaries making embarrassing mistakes ^[13]. The company's position is that this is an experimental feature aimed at a better search experience ^[5]. Both statements can hold at once, which is the useful part: an experiment is precisely what an unenforceable policy cannot reach.

The rest of the trade press should read the cost side. The Washington Post's trial of an AI-generated podcast feature produced errors and fictional quotes, Semafor reported in December ^[21]. BuzzFeed cut another third of its staff after betting on the technology ^[22]. Pew found strong signs of AI authorship in around 10% of the webpages it sampled ^[23], so the text is arriving on the open web whatever the Times does; the variable is whose masthead sits above it. Semafor also flagged the obvious incentive for someone to spend a day trying to trip the summariser into saying something stupid, with the paper's political opponents hunting for ways to discredit its journalism ^[20].

None of this is an ambitious use of the technology, as Semafor noted: a few sentences over the paper's own copy ^[27]. The exposure is not proportional to the ambition. A wrong summary of a correct story is still a correction with the masthead on it.