DOJ backs unlicensed use of copyrighted works to train LLMs, easing a key policy risk for AI companies
The US government sides with OpenAI on a central copyright question for generative AI, filing a 20-page brief in the New York Times lawsuit that backs unlicensed use of copyrighted works for training large language models. For investors and founders, that lowers one of the biggest policy overhangs on model builders, even if it does not settle the case and leaves downstream licensing economics very much alive.
The filing matters because it signals where Washington wants the industry to go at a moment when AI infrastructure, application startups, and incumbent platforms are all pricing in continued access to enormous training corpora. The brief ties that position directly to US competitiveness, citing President Donald Trump’s January 2025 executive order on removing barriers to American leadership in artificial intelligence, which is a meaningful policy cue for a market already rewarding scale in foundation models.
In the Southern District of New York case brought by the New York Times against OpenAI, the administration argued the US has a strong interest in maintaining a robust and competitive AI industry and retaining global leadership. That is not a judicial ruling, and the authors of the brief do not control the court. Still, when the federal government formally aligns itself with a model developer on fair use, it changes the temperature around legal risk for the broader startup ecosystem.
The argument goes to the heart of how current systems are built. Models behind products such as ChatGPT, Claude and Gemini are trained on vast bodies of published material, including copyrighted books, articles and other media, often without permission. Publishers and creators have been trying to establish that such use is unlawful, while AI companies have framed training as transformative and therefore protected under fair use.
That distinction is now where much of the money sits. If courts broadly accept training as fair use, large model companies preserve the economics that made this cycle possible. If courts narrow that theory, value could shift toward licensed data suppliers, rights management intermediaries and smaller, domain-specific models trained on cleaner datasets. My read is that this filing makes the first outcome more plausible, but not costless, because regulators and courts can still distinguish between training on content and acquiring that content through unlawful means.
The Weaponization of Fair Use: Copyright as Industrial Policy
The US Department of Justice’s decision to intervene in the New York Times v. OpenAI lawsuit fundamentally redefines the generative AI regulatory landscape. By explicitly filing an amicus brief that links the "fair use" doctrine to national security and global AI leadership, Washington has signaled that it views frontier model development not just as a commercial enterprise, but as a sovereign imperative.
For enterprise decision-makers and investors, this intervention transforms a complex intellectual property dispute into a matter of industrial policy, creating a stark divergence from the EU's transparency-first approach under the AI Act and the Digital Services Act (DSA).
Margin Relief and the Oligopoly Threat
The DOJ’s stance provides a massive, albeit unofficial, subsidy to foundation model builders. If the Southern District of New York adopts the government's framing—that ingesting massive amounts of copyrighted data to train Large Language Models (LLMs) is inherently "transformative"—it preserves the underlying unit economics that fueled the generative AI boom.
Mandatory, broad-based licensing regimes would act as a devastating tax on model training. The DOJ explicitly argued in its filing that requiring licensing fees would "hamper competition" and "disproportionately benefit legacy media outlets," effectively creating an oligopoly where only trillion-dollar hyperscalers could afford the entry barriers. By lowering the threat of retroactive, multi-billion-dollar licensing settlements, this brief reduces the legal risk premium that venture capital must price into AI infrastructure investments.
The Pivot to Provable Data Provenance
However, as demonstrated by the recent $1.5 billion Anthropic settlement, the legal system is aggressively bifurcating the act of training from the method of acquisition.
While the act of a machine "reading" a copyrighted article might be protected as fair use, using a pirated shadow library or bypassing a publisher's paywall remains legally toxic. This distinction creates a new, highly lucrative infrastructure layer: Provable Data Provenance.
Enterprise Risk Mitigation: Corporate boards deploying AI agents for high-stakes legal, financial, or healthcare workflows will not tolerate ambiguity. They require models trained on auditable, legally pristine datasets to ensure their proprietary workflows are not contaminated by copyright claims.
The RLPF Premium: As the public internet becomes a legal minefield (and increasingly polluted by AI-generated synthetic data), the value shifts to proprietary, domain-specific data and Reinforcement Learning from Professional Feedback (RLPF). Startups whose primary moat is the secure capture of expert enterprise workflows—entirely insulated from public copyright claims—will command massive premiums.
The Global Regulatory Arbitrage
This DOJ filing sets the stage for a deeply fragmented global AI market. While the US effectively subsidizes domestic model builders by shielding them from copyright maximalism, jurisdictions like the EU and UK are empowering publishers with strict opt-out rights and transparency mandates.
For startups operating out of hubs like Bangalore, Singapore, or London, this requires a dual-track strategy. Models trained aggressively in the US regulatory haven might face severe deployment friction, heavy compliance audits, or outright bans when exported to regions with stricter data sovereignty frameworks. Ultimately, the next phase of the AI arms race will not just be fought over GPU clusters, but over who controls clean, legally unencumbered data supply chains across conflicting global jurisdictions.
What It Means
The immediate implication is not that copyright risk disappears, but that the market may start separating two issues that had often been bundled together. One is whether using copyrighted material to train a model can qualify as fair use. The other is whether the data was obtained legally in the first place. That split already showed up in the Anthropic case, where the company was ordered to pay a $1.5 billion copyright settlement to a group of writers, yet was not penalized for the act of AI training itself.
Judge William Alsup’s reasoning in that case is important because it gave the industry a framework investors understand. He compared training on books to a human reading in order to write something different, while faulting Anthropic for using illegal shadow libraries to pirate books. In practical terms, that suggests the market may reward companies that can document data provenance even if they continue to resist broad licensing regimes for training.
There is also a geopolitical layer here that many narrower legal stories miss. The administration’s brief explicitly connects fair use flexibility to national competitiveness and global standard setting. That pushes this beyond a dispute between one publisher and one AI company. It turns copyright doctrine into industrial policy, which is how venture capital and cloud platforms are increasingly treating AI infrastructure bets.
$1.5 billion: The size of the approved Anthropic copyright settlement, a case that punished the use of illegal shadow libraries while stopping short of condemning AI training itself.
For startup founders, the lesson is not simply that scraping wins. The better lesson is that defensible data supply chains may become as important as model quality. In this cycle, the winners are unlikely to be only the companies with the biggest GPU clusters. They may also be the ones that can prove where their training inputs came from, negotiate selective licensing where it adds strategic value, and survive enterprise due diligence on copyright exposure.
Background
The New York Times case has become one of the most closely watched tests of whether publishers can force a new economic settlement on AI developers. The paper has argued that OpenAI used its copyrighted work without authorization. OpenAI, like other model companies, has relied on the idea that training is transformative rather than substitutive, even as media groups push to show that chatbots can reproduce or diminish the value of original reporting.
This is not happening in isolation. Over the last two years, media companies, authors and rights holders have pursued two tracks at once: litigation against model developers and licensing deals with selected AI companies. That mixed strategy suggests the market already assumes a binary outcome is unlikely. Some content will be licensed because it is timely, premium or useful for retrieval and grounding. Some training practices may still be protected as fair use. The commercial middle ground is where many AI data businesses may be built.
The administration’s filing lands after another development that sharpened the legal boundary. In July 2026, the Anthropic settlement was approved, and the case was widely read as favorable to AI developers on the narrow question of training, even while imposing a large penalty related to pirated books. That was a striking number. It also reinforced a message that startup counsel have been repeating privately: courts may punish how data is sourced before they punish the act of learning from it.
Seen in a broader pattern, this looks like a maturation phase for generative AI. Early in the cycle, capital largely priced legal uncertainty as an abstract discount on the category. Now the market is beginning to differentiate. OpenAI remains central because of its scale and the visibility of its litigation, but the same logic touches companies building coding assistants, legal copilots, research tools and vertical search. If the legal system converges on a permissive standard for training, distribution, product trust and compute access become even more decisive competitive moats.
What Analysts Say
No formal analyst note is needed to see why public market investors would welcome this direction. The brief supports the operating assumption behind a large share of AI spending: that frontier model development can continue without a universal requirement to preclear every copyrighted input. That is favorable for cloud vendors, enterprise software groups embedding generative AI, and startups whose margins would collapse under blanket licensing costs.
Still, there is a bear case. Government support for OpenAI may harden opposition from publishers, creators and courts wary of letting industrial policy override copyright protections. If judges conclude that outputs too closely substitute for originals, or if evidence shows deliberate evasion in data collection, the legal calculus could change quickly. There is also reputational risk for enterprise buyers that do not want to build on models tied up in long-running copyright disputes.
The bigger trend line is that policy is slowly moving from moral panic to market design. Last year’s executive order framed AI leadership as a national objective. This week’s brief operationalizes that objective inside active litigation. Pair that with the Anthropic ruling and the continuing pressure on publishers to choose between suing and licensing, and the direction of travel becomes clearer: AI training may remain broadly defensible, while the market builds a more explicit price for premium content and cleaner data supply.
What to watch next is straightforward: the Southern District of New York’s handling of the Times case, any judicial reaction to the administration’s fair use framing, and whether more publishers strike commercial deals rather than extend litigation. Those are the triggers that will tell founders and investors whether this policy signal becomes durable doctrine or just a useful brief in a still unsettled fight.
Frequently asked questions
Why is the US government siding with OpenAI on copyrighted training data?
The US government argues that using copyrighted works to train large language models can qualify as fair use under existing copyright law. In the New York Times lawsuit, that position supports OpenAI on a core legal question, though it does not decide the final outcome of the case.
Does the government filing mean OpenAI has won the New York Times lawsuit?
No. The filing is a legal brief that supports OpenAI's position, but the court still has to weigh the facts and make its own ruling. It lowers policy uncertainty for AI companies, yet it does not end the litigation or eliminate potential liability.
What does this mean for AI companies that train models on copyrighted material?
It suggests the federal government is open to a broad fair use argument for training AI models on copyrighted content without a license. That could reduce one of the biggest regulatory overhangs for model developers, especially startups and public market AI names.
Will publishers still be able to demand licensing deals from AI companies?
Yes. Even if courts accept fair use for model training, publishers can still pursue licensing deals for business reasons, access, attribution, or premium content partnerships. The legal debate over training data does not automatically settle downstream economics.
Why does this matter to investors and founders in generative AI?
This matters because copyright litigation has been a major risk factor for AI model builders and their backers. A government stance favoring fair use can improve sentiment around AI infrastructure, model companies, and the long-term cost of training data.
Could this US position shape future AI copyright policy in America?
Yes. A formal government brief can influence how courts, regulators, and lawmakers frame the issue of copyrighted training data. While it is not the final word, it signals that Washington may lean toward protecting AI development under current copyright principles.
