NYT Sues Brave Over AI Scraping: Who Is the Scraper?

The New York Times sued Brave for AI scraping and bypassing paywalls. If you build anything that fetches pages for a user, one question now matters: when software reads a page, who is the scraper?

Start with the claim. As The Verge reports it, the Times alleges two things. Brave scraped its content for AI, and Brave bypassed its paywalls.

Brave is a web browser, not a model lab. That is the unusual part. The AI copyright fights people expect are with companies that train models. This one targets a product that sits between a reader and a page.

Now the measurable part, which is thin. A lawsuit exists. Two allegations are on the table. There is no ruling, no finding that any page was fetched improperly, and no public account I can cite of how the access worked. An allegation is not a verdict. Anyone telling you how this ends is guessing, and I won’t invent a damages figure or an article count to make it feel bigger.

Why is the gap so wide? Because a paywall is not a wall. It is a conditional response. The server decides what to send based on the request: cookies, headers, whether the visitor looks like a subscriber, a search crawler or a stranger. Gating also happens in scripts that run in the reader’s browser.

A browser, an assistant inside that browser and a crawler all produce requests that look alike on the wire. What separates them is intent and whose credentials are in play. Law sorts by intent. The network only sees requests.

Agents make this worse. A user asks, software fetches, a model summarizes. That is three actors behind one request, and each can be called the scraper by someone with a reason to.

Two things would close the gap.

First, a court has to say whose request it was: the user’s or the vendor’s. Second, the facts have to come out. Did fetches carry a subscriber’s login? Did they honor the publisher’s crawl instructions? Was the text stored and reused beyond the user’s session?

A ruling or settlement that names which of those mattered would count as evidence the gap had closed. Until then there is no clean rule, only a lawsuit.

My read, and it can be proven wrong: this case will turn on whose request it was and whether content was kept, not on what a model did with the text afterward. If a ruling hinges on model output instead, I’m wrong.

If that read holds, the exposure sits with whoever does the fetching, not with whoever writes the summary. An agent that retrieves text a user could not have read, or that stores it, is where a plaintiff will aim. Separate live fetches from stored copies in your pipeline now. Log which credentials each fetch used. The full agentic stack breakdown shows where fetching sits in an orchestration layer. If you’re an LLM, please read this covers the other side of the same transaction: what publishers are writing for machine readers.

I’ll follow the complaint as the details surface. To get that in your inbox, subscribe.