The old bargain of web search was imperfect, but recognizable: publishers made their pages discoverable, search engines indexed them, and readers followed links to the original work. AI search complicates that exchange. When a search service produces a fluent answer from several sources, a reader may get what they need without visiting any of them. The publisher still bears the cost of reporting, editing, updating, and publishing. The platform increasingly controls the encounter with the audience.

That tension helps explain why publishers have pursued lawsuits, licensing deals, competition complaints, and technical restrictions on automated access. Their objections are not all the same. Some concern copies used to train a model. Others concern a service fetching an article to answer a particular question. Still others concern how much of the original work appears in an answer, whether the source is accurately identified, and whether participation can be refused without sacrificing ordinary search visibility.

This is an analysis of the dispute, not legal advice or a report of a new ruling. It draws on documented announcements, complaints, research, and legal materials, principally from 2024 and 2025. Dates are stated where they matter. A lawsuit’s allegations are not established facts; this article does not claim to establish the October 2026 procedural status of every case discussed.

The crucial distinction: training is not the same as search

Training involves processing large collections of material to develop a model’s statistical capabilities. At a different stage, an AI search product may retrieve documents when a user asks a question, select passages, and generate an answer informed by that material. This is often described as retrieval-augmented generation, or RAG. A product can combine both processes, but permission for one should not be assumed to cover the other.

A publisher might license an archive for training while reserving rights over current articles, or allow traditional search indexing while objecting to answer generation. The legal analysis can also differ. How content was acquired, what copies were made, what the system does with them, and what reaches the user can all matter. Saying that an AI company has read a page is therefore too imprecise to resolve a rights dispute.

The U.S. Copyright Office’s May 2025 pre-publication report treats data collection, training, retrieval-augmented generation, and outputs as distinct parts of its analysis. It also examines licensing and market effects. That structure is useful precisely because the public debate often collapses several activities into a single word: scraping. Copyright Office report.

What publishers are actually objecting to

Publishers’ concerns usually fall into four overlapping categories. The first is permission: was the work acquired and used on terms the rights holder accepted? The second is substitution: does the answer replace a meaningful reason to visit or subscribe? The third is attribution: does the service distinguish the original reporting from background material and identify sources accurately? The fourth is bargaining power: can a publisher reject a particular use without disappearing from the channels that help readers find it?

These are related but not interchangeable. A prominent citation can address attribution while doing little for revenue. A payment can address compensation while leaving an inaccurate summary unresolved. A crawler block can restrict access without answering whether earlier copies were lawfully made. A serious agreement has to specify which problem it is solving.

Nor are publishers a single constituency. A subscription newspaper, an advertising-supported local outlet, a reference publisher, and an affiliate-funded review site have different economics. For one, an AI answer may substitute for a paid explanation. For another, it may eliminate the visit that generates advertising income. For a third, the valuable asset may be a carefully maintained database rather than a single article.

The litigation has expanded beyond model training

The New York Times filed its copyright lawsuit against OpenAI and Microsoft in December 2023, alleging that its journalism had been used without permission and that the resulting products competed with it. That filing became an important reference point for the broader dispute. The complaint states the newspaper’s position; it is not itself a judicial determination of infringement. The Times’ account of its lawsuit.

OpenAI has disputed the newspaper’s claims and defended training as fair use. Its February 2024 motion to dismiss illustrates why readers should distinguish the arguments of a litigant from the conclusions of a court. Whether a particular acquisition, training process, or output is lawful requires analysis of the relevant facts and claims, not simply choosing a side’s description of the technology. OpenAI court filing.

By September 2025, the conflict also included lawsuits aimed more directly at answer engines and search summaries. Reuters reported that Encyclopaedia Britannica and Merriam-Webster sued Perplexity, alleging unauthorized use of their material and harm to traffic and revenue. The significance is the product being challenged: an answer service drawing on reference content, not merely a model’s training pipeline. The allegations still require proof. Reuters on the Britannica lawsuit.

Reuters also reported in September 2025 that Penske Media, whose publications include Rolling Stone, Billboard, and Variety, sued Google over AI Overviews. That challenge brought publisher objections into the core search relationship: the visibility of reporting, the use of that reporting in generated answers, and the terms under which publishers participate. It should not be read as a finding that Google’s product violated the law. Reuters on Penske’s complaint.

Why fair use is not a universal permission slip

In the United States, fair use is a fact-specific legal defense. The statutory factors include the purpose and character of the use, the nature of the copyrighted work, the amount and substantiality used, and the effect on the potential market. Commercial activity is not automatically disqualified, and calling a process transformative does not end the inquiry. 17 U.S.C. §107.

Copyright also distinguishes facts from protected expression. A newspaper does not own an event merely because it reported it. But that does not mean the article’s original wording, photographs, or other protectable material can be copied without limit. A rights analysis must identify what was used rather than treating all information on a page as either completely owned or completely free. 17 U.S.C. §102.

A June 23, 2025 order in Bartz v. Anthropic illustrates the importance of separating uses. The court treated the model-training use at issue as fair use while distinguishing the acquisition and retention of pirated books in a central library. That historical ruling was not a general authorization to acquire any work from any source, and it did not determine the legality of every AI search product. The court’s order.

For publishers and platforms, the practical lesson is to avoid sweeping claims. Neither AI training is always lawful nor every AI summary infringes is a reliable account of the law. A summary that gives a brief explanation with a useful link presents different facts from an answer reproducing substantial protected material and replacing the original. Contracts, access conditions, and the specific jurisdiction may introduce additional questions.

The traffic question: evidence of a problem, not a universal verdict

One of the clearest public measurements comes from Pew Research Center. In a July 22, 2025 analysis of browsing data from 900 U.S. adults during March 2025, Pew found that users clicked a traditional search-result link on 8% of visits to Google results pages containing an AI summary, compared with 15% of visits without one. Links inside the summaries were clicked on just 1% of visits to pages that had a summary. Pew’s findings and methodology.

Those findings support a narrow but important observation: in the studied sample, searches with AI summaries were associated with fewer clicks to external results. They do not establish that every publisher lost the same percentage of traffic, that every search category behaves alike, or that the summary alone caused the difference. Queries receiving summaries can differ from those without them. March 2025 behavior is also not a direct measurement of October 2026 behavior.

Google offered a different aggregate account in an August 6, 2025 post by search executive Liz Reid, saying total organic click volume from Search to websites was relatively stable year over year and that it was sending more quality clicks. The company argued that some outside reports relied on flawed methodologies or isolated examples. That is Google’s characterization, not an independently reconciled answer to Pew’s research. Google’s response.

The two accounts need not describe identical populations or measures. Total outbound clicks can remain stable while individual outlets or information-heavy query categories lose visits. A smaller number of more engaged visitors can be valuable, but not necessarily enough to replace lost advertising impressions or subscribers. Publishers need measurements of revenue, subscriptions, return visits, and the cost of serving automated traffic—not only a ranking report.

Licensing offers a negotiated path—but the details matter

Litigation is only one strategy. In May 2024, News Corp and OpenAI announced a multi-year partnership giving OpenAI access to current and archived content from specified publications and permission to display that content in response to user questions. The announcement shows that publishers and AI companies can negotiate arrangements spanning more than abstract model development. The partnership announcement.

The public announcement does not reveal every contractual term. Readers should not infer a standard price, unrestricted access to every asset, or permission that extends to unrelated publishers. Nor does a deal establish that unlicensed use elsewhere is unlawful: a commercial agreement and a copyright defense answer different questions.

A useful contract would distinguish training from retrieval and display; define whether archives, current articles, or both are included; specify excerpt limits and attribution; and establish how corrections, updates, and withdrawn material are handled. It should also address reporting and audit rights, retention, third-party access, duration, and termination. Without that specificity, a broad reference to AI rights can obscure what the parties have actually exchanged.

For smaller publishers, bargaining power is a separate concern. A national group with a deep archive and widely recognized brands can negotiate differently from a local outlet with a small staff. Any sustainable market must consider whether specialist and local reporting can participate on workable terms, rather than relying only on a few large catalogues.

Crawler controls make consent more concrete

Publishers also have technical tools. Robots.txt communicates crawling preferences to automated clients, while bot-management systems can enforce access restrictions at the network layer. These tools are useful but different: a declared preference is not a physical barrier, and an access block is not a court ruling about copyright.

In July 2025, Cloudflare announced a private beta of pay per crawl, designed to let website owners charge participating AI crawlers for requests. It also introduced controls aimed at differentiating content use for AI training. These announcements signal an attempt to move from indiscriminate access toward identifiable, governed transactions. They are evidence of tools being offered—not proof that every AI company participates or that every publisher can earn meaningful revenue from them. Pay-per-crawl announcement and training-control announcement.

Implementation still matters. Operators should verify crawler identities, inspect actual request logs, and understand which legitimate discovery channels a restriction affects. Blocking a named bot does not necessarily address requests routed through other services. Technical settings should support a clear rights policy rather than substitute for one.

Copyright and competition are different questions

A publisher may object not only to copying but also to the conditions imposed by a powerful distribution platform. In July 2025, Reuters reported an EU antitrust complaint from independent publishers over Google’s AI Overviews, including concerns about opting out. That was a competition complaint, not a copyright judgment. The distinction matters because a permitted use under one body of law does not automatically settle questions under another. Reuters on the EU complaint.

The policy question is whether publishers have a meaningful choice about answer generation while retaining access to ordinary discovery. That requires examining the actual product controls and the market in question. U.S. fair-use analysis cannot simply be transplanted into Europe, and an EU complaint should not be presented as establishing a worldwide rule.

Attribution must do more than decorate an answer

Citations are valuable when they make an answer verifiable. They are less useful when they point to a page that does not support the claim, blur the distinction between original reporting and commentary, or omit the source of a crucial finding. A generated answer can also lose qualifications, confuse dates, or mix an old article with a current question. Those errors can damage both readers and the publication being cited.

Better source treatment would preserve dates, identify who first reported a finding when that is known, link directly to supporting material, and make corrections accessible. Attribution does not by itself determine whether copying is permitted, but it is a basic requirement for a trustworthy information product. The rights dispute and the accuracy problem overlap without becoming the same problem.

What publishers and professionals should watch next

Publishers should start with an inventory of rights and uses: which material they own, what contributors have licensed, which agreements already exist, and what automated clients actually request. They should measure referral quality and commercial outcomes separately, preserve evidence of disputed reproductions, and obtain advice before treating a technical setting as a legal solution. The appropriate strategy may combine selective access, licensing, direct subscriptions, and carefully chosen enforcement.

Professionals using AI search have a simpler responsibility: follow the source. Verify the publication date, read the underlying document for consequential claims, and do not assume that a citation establishes permission to republish an article or image. For legal, financial, or public-policy decisions, a confident synthesized answer should be the beginning of checking, not the end.

The most consequential developments will be specific rather than rhetorical: rulings on defined uses, licensing terms that distinguish training from search, controls that offer a real choice, and transparent evidence about who receives traffic and revenue. A headline announcing that AI has won or publishers have won is unlikely to capture those differences.

The bottom line

AI search can make complex information easier to navigate. But useful answers still depend on people and institutions producing dependable source material. Publishers’ pushback is ultimately about whether that production remains economically and legally sustainable when the answer is delivered somewhere else. The durable bargain will need permission where required, defensible legal boundaries, accurate attribution, meaningful choice, and compensation models that do not make original reporting the unpaid input to someone else’s product.

Sources and scope

The links throughout this article lead to the relevant statutory provisions, a court order and party filing, the Copyright Office’s analysis, Pew’s research, company announcements, and Reuters reporting. Company statements and legal complaints establish what their authors said or alleged; they are not independent verification of all underlying claims. Historical litigation examples are dated, and no current docket outcome is asserted where it was not independently verified. No interviews, proprietary traffic measurements, or hands-on product tests were conducted for this analysis.