The AI Trademark Paradox: Why the World's Most Valuable AI Brands Can't Get Protection

Executive Summary

On November 24, 2025, a federal judge blocked OpenAI — the company behind the most widely recognized AI product on the planet — from using the word "Cameo" in its Sora video tool. The temporary restraining order came after Cameo, a celebrity video platform with a fraction of OpenAI's resources, proved that consumers could not distinguish between the two product lines. OpenAI renamed the feature to "Characters."

This was not an isolated incident. It was the latest data point in a pattern that now defines the AI trademark landscape: the best-resourced technology companies in the world are systematically failing at basic brand protection.

The scale of the problem extends far beyond any single company. GenAI-related trademark filings surged approximately 1,000% between 2021 and 2023, according to analysis by Sagacious IP. The United States and China together account for nearly two-thirds of all GenAI filings globally. Yet the industry's most recognizable names — GPT, ChatGPT, Gemini, Grok, Copilot — have been refused, withdrawn, or remain unregistered.

The paradox: an industry that attracted $211 billion in venture capital in 2025 alone cannot reliably protect the names under which it sells its products. The consequences are playing out in courtrooms, trademark offices, and boardrooms worldwide. This report examines why, and what practitioners should do about it.

---

The Filing Surge: 1,000% and Counting

The Scale of the Boom

Generative AI has triggered one of the fastest trademark filing surges in modern IP history. Research by Sagacious IP, drawing on the Markify global trademark database, found that GenAI-related trademark filings grew approximately 1,000% from 2021 to 2023 — a period that spans from the earliest commercial GenAI products through the post-ChatGPT explosion.

The filing activity is concentrated along two dimensions: classification and geography.

Classes 42 and 9 together account for approximately 55% of all GenAI-related filings. This concentration reflects the nature of the products: Class 42 covers AI-as-a-service platforms and software development services, while Class 9 covers the downloadable software and applications through which users interact with AI models.

Geographically, the United States and China together account for nearly two-thirds of all GenAI filings. The top filers globally include Amazon, Microsoft, and OpenAI from the US, alongside Chinese entities Luosiyang Technology Co. and Dongguan Genai Technology Co. Emerging filing activity is growing in India, South Korea, Indonesia, and Australia.

Source: Sagacious IP / IPWatchdog (February 2025), based on Markify database through August 2024. The 1,000% figure represents aggregate growth from a low 2021 base; absolute filing counts are not publicly available.

The Broader Context

The GenAI surge sits within a broader technology filing trend. According to WIPO's World Intellectual Property Indicators 2025, Research and Technology is the number-one sector for non-resident trademark filings globally, accounting for 19.3% of all cross-border filings in 2024. Class 9 alone represents 11.3% of all non-resident filing class counts — the single largest Nice Class worldwide.

The AI filing surge is paralleled by explosive growth in GenAI patents. WIPO's GenAI Patent Landscape Report found that GenAI patent families grew from 733 in 2014 to over 14,000 in 2023 — roughly an 1,800% increase, with over 25% of all GenAI patents filed in the single year of 2023 alone. China-based inventors filed 38,210 GenAI patent families during 2014–2023, roughly six times the US total of 6,276.

The trademark and patent surges are driven by the same underlying force — over $211 billion in AI venture capital in 2025 alone, according to Crunchbase — but they are distinct phenomena with different dynamics. A patent surge reflects R&D investment. A trademark surge reflects commercial launch activity, brand positioning, and defensive filing. The two can diverge sharply, and in AI, the trademark side is encountering obstacles that the patent side is not.

---

The Naming Crisis: Why AI Brands Fail the Distinctiveness Test

The Descriptiveness Problem

The structural challenge facing AI trademarks is not complexity or cost. It is naming culture: AI products are overwhelmingly named after what they do — and descriptive names are the hardest category of marks to register.

The Abercrombie spectrum, the foundational framework of US trademark law, classifies marks from weakest (generic) to strongest (fanciful):

The AI industry's naming culture sits overwhelmingly at the descriptive end. "ChatGPT" describes a chat interface powered by a GPT model. "Copilot" describes an AI assistant. "Gemini" was already in use by a data analytics company. "GPT" describes an entire category of technology. Engineering teams name products to be understood; trademark law rewards names that are meaningless. The result is a structural mismatch: the most recognizable AI brands in the world sit at the weakest point on the distinctiveness spectrum.

The GPT Generic Ruling

The clearest illustration is OpenAI's attempt to trademark "GPT." On February 6, 2024, the USPTO issued a final refusal, finding the term generic — it describes "generative pre-trained transformer," a category of technology used across the entire AI industry. The examining attorney cited over 100 pages of evidence showing widespread third-party use. Even OpenAI's consumer survey was rejected for methodological flaws (it surveyed software developers rather than the general public). In May 2025, the USPTO reiterated that GPT is "likely generic" and "not protectable under any circumstance."

The ChatGPT case followed a similar trajectory. The USPTO refused registration under Section 2(e)(1), finding the name merely descriptive of "a chat interface powered by a Generative Pre-trained Transformer." In March 2026, the Trademark Trial and Appeal Board affirmed the descriptiveness refusal — but accepted OpenAI's alternative Section 2(f) claim of acquired distinctiveness, allowing the application to proceed to publication. ChatGPT remains unregistered as a word mark; OpenAI holds only a logo registration.

The Success Stories

Not every AI company made the same naming mistake. The marks that have succeeded share a common trait: they occupy the suggestive, arbitrary, or fanciful end of the spectrum.

• Claude (Anthropic): An arbitrary human name with no descriptive relationship to AI software. Filed February 2023 — one month before launch. Registered.
• Midjourney: A suggestive, evocative name. Filed September 2022. Registered.
• Anthropic: A fanciful coined word derived from "anthropos." Registered May 2025.
• GPT-3 and GPT-4 (OpenAI): Specific product identifiers rather than category descriptors. Both registered.

The pattern is clear: coined names and human names work. Technical descriptors do not. This lesson has significant implications for the wave of AI companies currently choosing names for their products.

---

The AI Trademark Scorecard

The following analysis consolidates the trademark prosecution status of major AI products as of March 2026, organized by the type of challenge each company faces.

The Genericization Risk

These companies chose names that directly describe what the product does. OpenAI's "GPT" describes an entire technology category. "ChatGPT" describes a chat interface powered by that technology. Both sit at the descriptive or generic end of the trademark spectrum — the hardest territory to protect.

The Prior Art Collision

Microsoft's Copilot illustrates how prior registrations can block even the largest companies. The company withdrew its own trademark application after discovering registrations by Trimble Maps (1997), Copilot.com LLC (2015), and Copilot Corporation (2017). Internal sources described the branding strategy as chaotic, with an employee stating: "There is a delusion on our marketing side where literally everything has been renamed to have Copilot in it."

xAI's experience with "Grok" is the starkest cautionary tale. The USPTO cited six existing trademark registrations and 22 pending applications with confusingly similar names — including Groq (a semiconductor company, registered since 2017). xAI attempted to add "XAI" as a prefix and filed logo marks to differentiate. The USPTO was unconvinced, issuing further refusals in June 2025. xAI continues to operate Grok without federal trademark protection.

Google's Gemini dispute exemplifies the consequences of inadequate clearance. The USPTO refused both Gemini trademark applications in May and August 2024, citing likelihood of confusion with Gemini Data Inc.'s prior registrations. Despite the refusal, Google continued using the name. Gemini Data sued, alleging Google had "actual knowledge" of their marks and had attempted to acquire their rights through "an anonymous strawman buyer."

OpenAI's Sora faces a parallel collision. OverDrive Inc. registered "Sora" in December 2020 for its educational reading app. When OpenAI previewed its Sora video tool in February 2024, OverDrive's attorneys contacted OpenAI about the conflict. OpenAI launched anyway. OverDrive filed an opposition in October 2025 and a separate infringement lawsuit, citing evidence of misdirected support emails from confused librarians.

The io hardware dispute produced the most significant legal precedent. After OpenAI acquired Jony Ive's IO Products for $6.5 billion in May 2025, IYO Inc. — a startup whose name "only differ[s] by one letter and [is] pronounced identically" — won a temporary restraining order. The Ninth Circuit affirmed the TRO in December 2025, establishing that even pre-market announcements can trigger trademark liability. OpenAI abandoned the "io" brand entirely.

The Winners

The common thread: these companies chose names that are distinctive under trademark law, not merely descriptive of the underlying technology. Anthropic's approach stands out — filing "Claude" a full month before the product launched, ensuring the mark was on file before any third party could react. OpenAI, by contrast, filed "ChatGPT" 27 days after launching the world's fastest-growing consumer application.

---

The NFT Parallel: What a Previous Filing Surge Can Teach Us

The GenAI trademark surge is not the first time a new technology triggered a rush of trademark filings. The NFT boom of 2021–2022 produced a strikingly similar pattern — and a cautionary outcome.

NFT trademark filings surged from 27 in 2020 to nearly 5,800 in 2022 before declining sharply in 2023 — with only 2,152 applications filed through September, a pace roughly 50% below the prior year when annualized. The collapse tracked the broader NFT market downturn: once speculative interest evaporated, the filing incentive disappeared with it.

Sources: CryptoFlies, InsideBitcoins, Meki P Law. Note: NFT filing counts use different identification methodology (keyword matching in goods/services descriptions) than the GenAI filing data, and direct numerical comparisons should be made with caution. The 2023 figure covers January–September only.

Where the Parallel Breaks Down

The AI trademark trajectory shares the NFT surge's speed but not its structural weakness. Three differences matter:

Real products, real revenue. The NFT filing boom was driven overwhelmingly by speculative filings from small entities and individuals attempting to claim naming territory in a market with little underlying commercial activity. The GenAI filing boom is driven by operating companies — OpenAI, Google, Microsoft, Amazon — with real products, real users, and real revenue. AI is being embedded into enterprise workflows, not traded as speculative assets.

Sustained R&D investment. GenAI patent families grew from 733 in 2014 to over 14,000 in 2023 (WIPO). Over $211 billion in VC flowed into AI companies in 2025 alone (Crunchbase). The investment base is broad and institutional, not retail-speculative.

Regulatory legitimacy. WIPO's Nice Classification 2026 formally added "AI as a Service (AIaaS)" to Class 42, recognizing AI as a standalone technological pillar. No comparable regulatory legitimization occurred for NFTs.

Where the Warning Applies

The NFT experience does highlight two risks that are relevant to AI:

Filing quality. Many NFT trademark filings were abandoned or refused because they lacked genuine use in commerce, were filed in inappropriate classes, or described products that never materialized. Early signs suggest similar quality issues in AI filings — the high refusal rates among major companies documented in this report imply that smaller filers are likely faring worse.

Naming congestion. The NFT boom created a crowded, fragmented naming landscape that made clearance searches significantly more difficult. The same dynamic is underway in AI: the sheer volume of filings containing terms like "AI," "GPT," "neural," and "intelligent" is making clearance in Classes 9 and 42 increasingly complex.

The verdict: AI trademarks are unlikely to follow the NFT boom-bust pattern in aggregate, but the naming and quality problems that plagued NFT filings are present and intensifying.

---

The Regulatory Landscape

The filing surge and its attendant failures have not gone unnoticed by trademark offices and regulators. From WIPO's classification updates to the USPTO's deployment of AI-powered examination tools to the EU's novel regulatory framework, institutions are adapting — though whether fast enough remains an open question.

Nice Classification 2026: AI Gets Its Own Category

The 13th edition of the Nice Classification, effective January 1, 2026, formally recognizes AI as a distinct service category. Three new entries were added to Class 42:

1. AI as a Service (AIaaS) — cloud-delivered AI infrastructure and platform access
2. Consultancy in AI — advisory services distinct from general IT consulting
3. Research in AI technology — for firms engaged in AI R&D

Previously, AI services were captured under broad umbrella terms such as "computer programming" or "software as a service." The update provides specificity that simplifies classification — and makes clearance searches more targeted. The AIaaS market is projected to grow from $20.26 billion in 2025 to $91.20 billion by 2030.

USPTO: AI-Powered Examination

The USPTO itself is deploying AI to manage the trademark surge. In March 2026, the office launched Class ACT (Classification Agentic Codification Tool), an AI agent that automates pre-processing of trademark applications — assigning international classes, design search codes, and pseudo marks. The tool reduces classification preparation time from five months to five minutes. Human review remains mandatory before finalization.

The office has also integrated AI-based specimen checks to detect fraudulent submissions, smart filters for identifying duplicate or suspicious applications, and guidance requiring practitioners to ensure AI-generated content meets accuracy and good-faith standards.

EU AI Act: The Trademark Trigger

The EU AI Act (Regulation 2024/1689) created an underappreciated risk for brand owners. Under Article 25, any entity that places its name or trademark on a high-risk AI system becomes the legal "provider" of that system — assuming full regulatory responsibility including conformity assessment, risk management, and post-market surveillance obligations. The original developer is released from provider status.

For brand owners who white-label or rebrand third-party AI systems, this provision transforms a trademark decision into a regulatory compliance decision. Affixing a brand to a high-risk AI system is no longer just a marketing choice — it is a legal commitment.

FTC Operation AI Comply

The Federal Trade Commission launched "Operation AI Comply" in September 2024, targeting companies making unsubstantiated AI claims. As of early 2026, the FTC has brought at least 12 enforcement actions, including:

For brand owners, the implications extend beyond advertising law. Trademark applications that incorporate "AI" in goods and services descriptions face increasing scrutiny — both from examining attorneys skeptical of descriptiveness and from regulators skeptical of substantiation.

---

Strategic Implications

The preceding analysis points to a set of concrete actions for practitioners advising AI companies or managing AI-adjacent brand portfolios.

Name Distinctively

The single most important lesson from this analysis: choose coined, arbitrary, or suggestive names for AI products. The entire descriptive end of the spectrum — names containing "GPT," "AI," "Neural," "Intelligent," or any term that describes what the product does — faces severe registrability challenges.

Anthropic's "Claude" is the model. A human name, arbitrary in relation to AI software, filed before the product launched, and now registered. Compare this to OpenAI's experience: 28 filings, four registrations, and its two most important brand names (GPT, ChatGPT) either refused or still unregistered after three years.

The .ai domain explosion — approximately 600,000 registrations and an estimated 300% growth in 2024, according to registrar data from NameSilo — illustrates the crowding problem. When every AI company competes for names in the same narrow linguistic space, collisions are inevitable.

Clear Thoroughly, Clear Early

Every major AI company in this report was sued or refused for inadequate clearance. The pattern is consistent:

• Google continued using "Gemini" after two USPTO refusals and a pending lawsuit
• OpenAI launched "Sora" despite OverDrive's prior registration in the same class
• xAI filed "Grok" into a space occupied by six existing registrations
• OpenAI acquired a company named "IO" without resolving a single-letter conflict with "IYO"

The IYO v. IO ruling established a new precedent: pre-market announcements alone can create trademark liability. The Ninth Circuit held that a product launch video constituted sufficient infringing activity to support an injunction — before a single unit was sold.

File Before You Launch

Anthropic filed "Claude" on February 10, 2023 — one month before the product launched on March 14. OpenAI filed "ChatGPT" on December 27, 2022 — 27 days after the November 30 launch. The difference in prosecution outcomes is instructive: early filing establishes priority, gives the mark time to clear examination, and reduces the risk of third-party filings during the gap.

Audit for Nice 2026

The formal addition of AIaaS, AI Consultancy, and AI Research to Class 42 creates both opportunity and obligation. Companies with existing portfolios should review whether their current classifications adequately cover AI-specific services. New filers should take advantage of the more precise descriptions now available rather than relying on catch-all terms.

Watch the EU AI Act

For any company operating in the EU market, Article 25 of the AI Act means that branding decisions carry regulatory weight. White-labeling or rebranding a third-party AI system triggers provider obligations. Legal teams should coordinate trademark strategy with regulatory compliance before making branding commitments.

What to Watch: 2026–2028

The agentic AI naming rush. As AI agents proliferate — Gartner reported a 1,445% surge in multi-agent system inquiries from Q1 2024 to Q2 2025 — trademark filings for agent-specific products will accelerate. The AI agent market is projected to grow from $7.84 billion in 2025 to $52.62 billion by 2030. Founders are already treating bot names as brand assets.

The clearance bottleneck. A 1,000% filing surge means the noise-to-signal ratio in clearance searches has exploded. Every new AI product launch now requires navigating a thicket of speculative filings, abandoned applications, and prior art in the most crowded Nice Classes on the planet.

AI-generated brand selection. As autonomous AI agents begin making purchasing decisions — selecting products, vendors, and services without human intervention — the foundational assumption of trademark law faces its deepest challenge. Trademark doctrine assumes a human consumer who can be confused. When an AI agent selects a brand algorithmically based on price, ratings, and availability — potentially disregarding brand recognition entirely — who exactly is confused?

This is the next frontier of trademark doctrine, and it arrives at a moment when the industry building that technology cannot protect its own brand names. By 2028, the most valuable AI companies in the world will either own their names or be licensing them from someone who filed first. That window is closing. It will not reopen.

---

Methodology and Sources

This report draws on publicly available data from trademark offices, court filings, regulatory publications, and industry analyses.

Trademark filing data:

• Sagacious IP / IPWatchdog: "The Trademark Race in Generative AI" (February 2025) — GenAI filing surge data from the Markify global trademark database through August 2024.
• WIPO World Intellectual Property Indicators 2025 — Global filing volumes, class distributions, and sector analysis for calendar year 2024.
• WIPO GenAI Patent Landscape Report (2024) — GenAI patent family data for contextual comparison.

Company trademark data:

• Originality.AI: OpenAI Trademarks List (updated August 2025) — OpenAI's USPTO filing portfolio. Filing counts and statuses should be verified against USPTO TSDR for the most current data.
• ArentFox Schiff: Google Gemini Analysis — Gemini trademark refusals and litigation.
• Gerben IP: Grok Trademark Refusals — xAI filing history.
• Brighter Naming: Who Owns Copilot? — Microsoft Copilot withdrawal analysis.
• Trademarkia: Anthropic PBC — Anthropic filing portfolio.

Court decisions and disputes:

• CNBC: Cameo v. OpenAI TRO and TechCrunch: Preliminary Injunction
• IPWatchdog: IYO v. IO/OpenAI Ninth Circuit
• Accelerate IP: OverDrive v. OpenAI
• Stones Law: ChatGPT TTAB Decision (March 2026)
• TechCrunch: GPT Trademark Refusal

Regulatory sources:

• WIPO Nice Classification 13th Edition
• USPTO: Class ACT Announcement (March 2026)
• EU AI Act Article 25
• FTC: Operation AI Comply

Market context:

• Crunchbase: Global AI Funding 2025 — $211 billion in AI VC investment.
• Stanford HAI: 2025 AI Index Report — AI patent and market data.

Analysis based on Signa's global trademark intelligence platform, which includes over 147 million records across 200+ jurisdictions including USPTO, EUIPO, WIPO, CNIPA, and national offices worldwide.