The New York Times is testing AI summaries on readers

The New York Times has started showing readers text written by a machine. A new search page answers queries with excerpts, links to Times stories and AI-generated summaries of Times reporting. Max Tani, Semafor’s media editor, reported the test on Sunday.

The paper rolled it out quietly to a small subset of visitors over recent weeks. It is the news arm’s first test of AI-generated text that no Times journalist or editor has passed.

The Times hopes the tool will fix a search feature Semafor describes as troubled. Some of the paper’s own journalists are alarmed by it.

What the Times says

“We are always testing new ways for our users to discover and engage with Times journalism,” spokesperson Graham James told Semafor.

“This experimental feature is an example of how we are using the latest technology to create a better search experience.”

What the tool is not

The summaries draw on Times journalism and nothing else. This is not a chatbot answering questions about the world, and it does not reach across the open web.

That narrows the failure modes without removing them. A summary of accurate reporting can still misstate what the reporting said.

Wirecutter got one first

The Times has been running AI features in its non-news sections for months. Wirecutter, its review site, recently started testing a search tool called Wirecutter Finder. It points readers to AI-generated summaries and tips drawn from the site’s own recommendations.

Those experiments have stayed outside the newsroom. The search summaries do not.

The paper has also given its journalists internal AI tools for their own work for some time. Those sit behind the reporting rather than in front of the reader.

The union found out at the bargaining table

Times management told its union it would test AI summaries in search during a contract bargaining session in July. The media newsletter Breaker, written by Lachlan Cartwright, reported that on 7 August, describing a heated session. Semafor cited the report this week.

The announcement caused frustration inside the company’s editorial union. The union says the Times is breaking its own editorial standards on AI use.

What the union wants

The Times Guild has put a set of AI rules on the table. One would create an employee payment pool, splitting 22.5% of the paper’s AI licensing deals for training data among unionised staff.

Another would require human oversight of any AI use. That clause would make the new search tool impractical to run.

‘Meaningless’

Jim Luttrell, the Guild unit chair and a senior staff editor, said the paper’s current AI policies are “meaningless”.

“Management can change and even violate the company’s policy in any way they want, as they are doing with the summarization tool,” he told Semafor.

“Every time AI summaries have been deployed, they have made embarrassing mistakes. That’s why we need strong guardrails on AI in the contract, to protect readers’ trust in our journalism.”

Why the paper is doing it

Search and summarisation tools are now standard across websites, apps, inboxes and rival news organisations. Semafor described the Times as playing catch-up.

Several Times employees told Semafor the change is necessary. Search has long lagged the paper’s other publishing products, they said, which Semafor calls otherwise best in class.

The talks are still running. Management and the Guild returned to the bargaining table this month with the AI clauses unresolved.

The accuracy problem

A few sentences summarising the paper’s own reporting is not an ambitious use of the technology. It is a harder call for the Times than for most publishers, Semafor argued, because the institution runs on layers of fact-checking and editing.

The paper still makes mistakes and appends corrections. What it has built is a system for catching them first, and a reputation that rests on that system.

There is also an incentive for bad actors. Semafor raised the prospect of someone spending all day trying to trip the summariser into saying something stupid. The paper’s political opponents are looking for ways to discredit its journalism.

It has gone wrong elsewhere

The Washington Post trialled an AI-generated podcast feature that produced errors and fictional quotes, Semafor reported in December.

Other publishers have leant harder on the technology and paid for it. BuzzFeed cut another third of its staff after betting on AI.

AI text is spreading either way

Pew found strong signs of AI authorship in around 10% of webpages it sampled in July.

That is the environment the Times has spent the AI era positioning itself against.

Deepfakes, bots and hallucinations have made the basic question of what is real harder to answer. News organisations have responded by selling themselves as the places where a human checked the work.

Taking the human out, even in a few sentences on a search page, cuts against that pitch. Semafor made the point plainly, and it is the argument the union is making inside the building.

The money underneath it

Publishers are being pulled towards AI from several directions at once. OpenAI is funding 13 Axios local newsletters under a three-year deal.

Google, meanwhile, wants publishers to install a button that makes readers pick them inside its own products.

The Times has its own reasons to want readers staying longer. It fell short on digital subscriber growth in the second quarter.

What has not been said

The Times has not said which model generates the summaries, how many readers are seeing them, or how long the test runs.

It has not said whether the summaries carry a label, or what happens when one of them gets something wrong.

And it has not said whether the feature goes beyond the test group, or what would stop it if the union got its clause.