NYT v. Brave explained: can a search index be sold to AI firms?

When an AI assistant answers a question about today's news, the words often did not come from its memory. They came from a search index, looked up a second before you saw the answer. On 2 November 2026 the New York Times, and separately three Alden-owned papers, sued the company that sells one of those indexes: Brave. The question for the courts is whether a pipe between a crawler and a chatbot is just "search", or something that needs a licence.

The plain-English version

Imagine a huge public library with a clever librarian who has read every book and written a card for each one. Anyone can walk in and ask, "Which book talks about X?" and the librarian points to the shelf. Nobody minds; that is how libraries work, and the authors benefit because readers go and find their books.

Now imagine the librarian starts a second business. She photocopies the useful pages of every book, bundles them neatly, and sells the bundles to robots that write answers for people. The robots' customers never visit the shelf. The authors, who wanted readers to come to them, see none of the traffic and none of the money.

That is, in simplified form, the claim against Brave. Brave is a web browser and search engine company. It built its own index of the web, which the complaints describe as more than 30 billion pages, according to TechCrunch. It sells access to that index through a paid service, the Search API. The publishers say the buyers include Anthropic, Perplexity, Mistral, xAI and Cohere. Those companies are named in the complaints but are not being sued.

Brave sees it differently. In its separate 2025 case against News Corp it has argued that indexing is necessary for any search engine to operate, and that its snippets are short excerpts (at most 300 characters, per its filing, as The Register reports).

Two more complaints sit alongside the main one. First, the papers say Brave's crawler never gives its name, so a website cannot say "not you". Second, they say a browser setting called "Paywall blocker" lets readers get around subscriptions.

What should you take away? The old fight was "did an AI learn from my writing?" This fight is about the middle layer: the companies that gather the web and hand it to AI at the moment of a question. Everything here is an allegation. No court has ruled, and Brave had not publicly responded in the coverage reviewed for this piece.

How it actually works

To see why an index matters to AI, follow a single question through a modern assistant.

1. You ask: "What did the Fed decide today?" A language model's weights were frozen months ago, so it cannot know.
2. The assistant searches. It sends a query to a search provider through an API.
3. The provider consults its index. The index is a pre-built copy of the web, organised so lookups take milliseconds rather than hours.
4. Content comes back. Older APIs returned links and a snippet. Newer ones return extracted page text, ranked and trimmed to fit a model's context window.
5. The model writes the answer from those passages and cites them.

Step 4 is the legal pressure point. Brave's API page calls its main product LLM Context, an endpoint that "returns ranked, extracted page content (text, tables, code, JSON)" for grounding, and lets customers set output from roughly 1k to 33k tokens. That is not a link. It is the article, or a big chunk of it, delivered to software.

Why own an index at all?

Building an index means crawling billions of pages continually. Per The Register, only a few companies do it at scale: Google, Microsoft (Bing), Baidu and Yandex are the largest; smaller engines such as DuckDuckGo rely on an index provider like Microsoft. Brave is one of the few independent alternatives, which is exactly why AI companies might want it: a source that is not Google or Bing.

The crawler problem

Websites control crawlers with a plain text file, robots.txt, naming each bot ("User-agent: X") and what it may fetch. This only works if the bot announces a name. Brave's crawler help page says it does not advertise a differentiated user agent, to avoid discriminating against sites that only allow Google; if Googlebot cannot crawl a page, Brave's bot will not either. It also says it does not use robots.txt to stop indexing; the supported way out is a noindex directive, which takes effect only after Brave re-fetches the page.

Think of a visitor who refuses to wear a name badge so that venues with a "Google only" guest list will still let them in. Brave's reason is competition. The publishers' reading is that a nameless visitor cannot be refused.

The three accusations in one table

Sources: The Next Web and PPC Land. These are the plaintiffs' allegations.

The deep dive

Procedural facts

Both suits were filed on 2 November 2026 in the Southern District of New York: the Times case is 26-cv-9549 and the Daily News/Tribune/Sentinel case is 26-cv-9554, per PPC Land and heise. Each has five counts: direct, contributory and vicarious copyright infringement over the index, Ask Brave and the API; DMCA circumvention over the Paywall blocker; and unjust enrichment under New York law. The relief sought includes statutory and actual damages, disgorgement of profits (including ad revenue), a permanent injunction and a jury trial. No dollar figure is stated.

The Times says it sent warning letters on 11 August 2025 and 13 April 2026 that Brave allegedly ignored, which the complaint can use to argue willfulness. The Daily News complaint describes no letters (PPC Land).

The evidence the complaints cite

• Verbatim output. PPC Land reports that in one Times test the API returned 1,291 of 1,295 words of a story, about 99.7%. This was the plaintiffs' own test; the filings do not show how often ordinary use produces such output.
• Marketing language. The complaints quote Brave describing the API as used "in training and grounding several of the Web's biggest large language models". The Next Web adds that a Brave blog post reportedly says Anthropic used the API to train Claude 2; that is a report of a document I have not read directly, so treat it as unverified.
• Endpoints. DEV Community, summarising the complaints, lists News Search, Web Search, LLM Context and Answers, with Answers described as the service behind Ask Brave. Pricing on Brave's page: Search $5 per 1,000 requests; Answers $4 per 1,000 requests plus $5 per million tokens.
• Scale. The Next Web reports "more than two million registered articles" allegedly affected in the three-paper case.

A back-of-envelope illustration, mine rather than any filing's: US copyright law allows statutory damages of up to $150,000 per work for willful infringement (17 U.S.C. §504(c)), a per-infringement figure News Corp also seeks in its countersuit, per the AP report. Two million works at that ceiling would be $3 \times 10^{11}$ dollars, which is far beyond what courts award in practice and beyond Brave's reported revenue. The point is only that statutory damages create leverage; real exposure depends on registration, willfulness findings and fair use.

Legal theory: three different uses in one pipeline

The key factor is the market effect. News Corp's July 2026 counterclaim makes this explicit: it argues that indexing for search differs from parsing content for AI pipelines and, per the AP report, that Brave distributes and sells "versions of articles" from the Wall Street Journal and New York Post to AI companies. The publishers' implied argument is that a licensing market now exists. Press Gazette's tracker lists deals including News Corp with Meta at up to $50m a year for at least three years, and reported Amazon–Times payments of $20–25m a year. If a licensing market exists, selling the same text without a licence substitutes for it.

Brave's counter: it sued News Corp first in March 2025 (N.D. Cal., case 3:25-cv-02503) seeking declarations that indexing is fair use and that using Brave with AI chatbots through APIs is not infringement. It argues the opposite rule would entrench Google and Bing and says it has under 1% of search share (Insurance Journal summary, chatgptiseatingtheworld).

Grounding vs. training

Training changes weights; grounding changes the prompt. Brave's own terms say that storing results or using them for training requires a plan with storage rights (Brave API page), which implies the default, grounding-only use does not include training. The complaints' training allegations rest largely on Brave's marketing. PPC Land notes the complaints also cite Meta's Llama 3 paper, which describes the model querying Brave's search when answering rather than training on the index, and that the filings do not establish which named companies use the API now or which endpoints they use. That cuts against the training part. This is why the training/grounding line may decide the case: it separates "learning from text" from "retrieving and re-presenting text".

The DMCA and ad counts

The Paywall blocker count is a different animal. §1201 does not ask whether the use is fair; it asks whether a technical access control was circumvented. The complaint cites a Reddit user who read the Times without logging in (The Next Web). The ad count requires no copying at all, so it could survive even if the copyright claims fail.

The comparison class

This is the latest in a long list. Press Gazette's tracker shows suits against OpenAI/Microsoft, Perplexity and, on 21 October 2026, a Canadian publisher group against Anthropic. What is new is the target: an infrastructure supplier rather than a model developer, and a counter-suit chain (Brave v. News Corp, then News Corp's counterclaim, now the Times).

Why it matters

Everyday users. If retrieval providers need licences, AI answers about news may get cheaper to license for big publishers and thinner for everyone else. The reverse also holds: if indexing for AI is fair use, assistants keep pulling from the whole open web, and publishers see fewer clicks. The complaints cite a study that AI crawlers generate 95.7% less referral traffic than Google search (via The Next Web; I did not review the study).

Developers. Many products treat a search API as a utility. This case says it is also a source of legal risk. Practical hedges: record which provider and fetch time each citation came from, keep the provider behind one interface, and run retrieval evaluations against a second provider so a switch is a day, not a quarter. These are the same suggestions made in the DEV Community analysis, and they cost little.

Companies. The named customers are not defendants, but discovery could reach contracts and logs. Any firm that buys content-bearing APIs should read what its supplier warrants about sourcing.

The field. The web is held up by an implicit bargain: let crawlers in, get traffic back. Search engines honoured the second half. If AI intermediaries stop honouring it, publishers will block more aggressively, and the independent-index route, one of the few ways to avoid dependence on Google or Microsoft, gets more expensive. That is the tension Brave itself points to.

What to be skeptical of

• These are allegations. Nothing has been tested. Brave had not commented in the coverage I found, and I could not read the complaints directly; claims here come from press and trade summaries.
• Selected outputs. The 1,291-of-1,295 result was the plaintiffs' own test. It shows capability, not typical behaviour.
• Numbers are shaky. PPC Land notes the "more than $100 million" revenue figure is a third-party estimate of Brave's total 2025 revenue, not tied to the API, and the filings give inconsistent user counts (close to 130 million versus over 100 million).
• Training claims are thin. The Claude 2 reference comes via a reported blog post; neither Anthropic nor Brave has been quoted on it in what I read. Anthropic is a named customer, and I note that because this site covers Anthropic regularly.
• The crawler argument cuts both ways. Brave's reason for no distinct user agent is competitive, and Brave offers noindex. Whether a nameless crawler is unlawful, rather than merely unpopular, is untested.
• Plaintiffs are not neutral. The Times has also sued OpenAI and Perplexity, and publishers are negotiating licences while litigating. Litigation is also leverage.
• Conflicting dates. Press Gazette dates Brave's suit against News Corp to February 2026; the docket (filed 12 March 2025) and other outlets say March 2025. I used the docket.
• Precedent is unsettled. heise notes the Times' older OpenAI case is still pending, and that the US government has taken OpenAI's side on training. Nothing here forecasts a result.

What to watch next

1. Brave's answer or motion to dismiss in S.D.N.Y. (26-cv-9549 and 26-cv-9554): does it lean on fair use, or attack the DMCA and unjust-enrichment counts?
2. Brave v. News Corp in N.D. Cal.: a ruling on indexing as fair use would be the first direct signal for the index-to-API pipeline.
3. Customer reactions. Whether Anthropic, Perplexity, Mistral, xAI or Cohere comment, change providers or add indemnity language.
4. Crawler changes. Whether Brave announces a named user agent or honours robots.txt for AI-bound access.
5. Licensing moves. More deals like those Press Gazette tracks, and whether any index owner offers a publisher-paid or opt-out API tier.
6. Related coverage. Our earlier piece on Shared AI's opt-out token for robots.txt covers the voluntary approach; this case tests what happens when the crawler will not play along. For Anthropic's financial and legal context see the S-1 explainer.

How this was made: researched and written by an AI model (Claude) from the primary sources listed above, then checked claim-by-claim against those sources in a separate AI fact-check pass. Spotted an error? Email [email protected] and we correct it publicly. Our process.