DOJ's OpenAI Fair-Use Brief Rewrites the Brand-Protection Calculus - From New York to Nice Class 36

Copyright Steps Back; Trademark Must Step Up

On 1 September 2026, the US Department of Justice filed a Statement of Interest in the Southern District of New York that reordered the IP landscape. The brief, covering every case in the consolidated In re OpenAI, Inc. Copyright Infringement Litigation (MDL No. 25-md-3143), declared that training large language models on copyrighted text should generally be treated as fair use. It was, according to Reuters, the first time the US government had taken a formal position on whether AI training on copyrighted material constitutes infringement.

The filing is not a ruling. Judge Sidney H. Stein still decides. But the signal is blunt: brands that have relied on copyright as a first line of defence against AI systems ingesting their content, pricing data, customer communications or proprietary research need to think again.

For trademark professionals, the more telling detail sits in a procedural footnote to the same case. When the New York Times sought to amend its complaint in June 2026, it quietly dropped two claims: trademark dilution and contributory copyright infringement. That withdrawal is easy to miss amid the noise about fair use, but it matters. It suggests that dilution arguments in AI-training contexts are proving harder to sustain when the defendant can argue transformative use at the training stage.

The Gap That Trademark Must Now Fill

The legal analysis splits neatly. Copyright protects original expression and, in most jurisdictions, requires human authorship. It does not reliably prevent AI-generated impersonation of a brand's identity, voice or visual style. Trademark law fills that gap: source-identification, consumer-confusion and goodwill protections address the commercial harms that AI-generated replicas actually cause.

Brand impersonation has already become a technology-driven problem. AI has lowered the cost dramatically, allowing bad actors to generate convincing phishing pages, clone an executive's voice, or spin up dozens of fake social profiles without specialist skill. For fintech and financial-services brands, the heartland of Nice Class 36, attackers increasingly combine ads, messaging apps and phishing sites into coordinated funnels that move victims from initial exposure to financial harm.

The filing data tells a worrying story. AIBD analysis of IPO trademark data shows only 1,674 UK trademark applications in Class 36 in Q3 2026, a fall of 29.8% against the prior period (AIBD analysis of IPO (TMD) data, as of September 2026). Fewer registrations mean fewer registered rights to enforce; with the US government now signalling that copyright will not catch AI training, the coverage gap for financial-services brands widens at exactly the moment they need it least.

Three Regimes, One Problem

What makes the DOJ filing consequential outside the United States is what it confirms about regulatory divergence. Organisations operating internationally now face three distinct regimes that will coexist for the foreseeable future. The US fair-use doctrine is broad and fact-specific. The EU's rights-reservation framework under the Digital Single Market Directive gives rights holders an opt-out from text-and-data mining, but enforcement is uneven. The UK operates a narrower text-and-data mining exception without a commercial research carve-out comparable to the EU's, making it, on paper, more protective for rights holders, but practically harder to police at scale.

The DOJ's filing does not bind any of those regimes. It will, however, shape the negotiating environment, the licensing table, and the confidence with which AI developers press ahead with training on commercially sensitive datasets, including the kind of financial and market data that Class 36 service brands generate daily.

The EUIPO has its own message for brand owners. Its 2026 Guidelines, which entered into force on 1 July, tightened specification requirements and confirmed that broad or vague wording in trade mark applications will no longer pass examination. Precision is now mandatory on both sides of the Atlantic, for different reasons.

What This Means for Class 36 Brand Owners

Financial-services brands, including insurers, asset managers, payment platforms and mortgage brokers, sit in one of the more heavily impersonated service classes. The AI impersonation risk is operational as well as legal: a spoofed brand in a phishing funnel is a fraud risk to customers, not just a dilution risk on a register. Separating those two harms, and building a response that addresses both, requires a registered trade mark as the enforcement foundation.

The picture is this: copyright's reach in AI contexts is being cut back by the most significant government filing the space has yet seen. The UK's own text-and-data mining regime offers some shelter, but it cannot substitute for the active brand-monitoring and enforcement posture that registered rights enable. UK filing data suggests that Class 36 brand owners have been reducing, not increasing, their registration activity at the precise moment the threat environment has intensified.

Copyright and trademark law are diverging in function: copyright may not catch the training stage, trademark may not catch the output stage, and the gap between them is where brand damage actually lives.

What to Do This Week

Get a proper audit of your Class 36 specification wording before the next renewal falls due. The EUIPO's tightened 2026 Guidelines and the UKIPO's own examination standards both reward precision. If your existing registration covers "financial services" without more, that breadth may fail to support an opposition or infringement action at the moment you need it most. Speak to your trade mark attorney, and cross-check your current portfolio against the UK register at TrademarkDashboard before the quarter closes.