September 2, 2026 · 16 min read · By Logo China Pro

Can You Trademark an AI-Generated Logo in China? The 90% Rejection Rate Explained

A founder in Portland opens Midjourney, types "minimalist phoenix logo for a tea brand," and gets back a clean, modern mark in nine seconds. She loves it. She files it with CNIPA through a Chinese agency. Six months later, the rejection notice arrives.

She is not alone.

Across 2025, CNIPA examined 9.8 million trademark applications and rejected roughly 35% at the preliminary examination stage -- over 3.4 million marks sent back (Source: CNIPA 2025 IP Work Progress (gov.cn)). For foreign applicants, the rejection rate is even higher. For foreign applicants filing raw AI-generated logos, the failure rate approaches 90% when all five risk vectors documented below are combined.

This guide breaks down exactly why -- with verified case law, real court decisions, and the specific provisions of China's 2026 Trademark Law revision that make the window closing tighter than ever.

The Two-Layer Problem: Copyright vs. Trademark

Most foreign founders confuse copyright and trademark. They should not. The distinction is the single most important concept in this entire article.

Copyright protects the expression in a creative work -- the specific lines, shapes, and colors of a logo as an artistic piece. Under both Chinese and American law, copyright requires a human author. A machine cannot hold a pen. A machine cannot be an author.

Trademark protects the commercial function of a mark -- its ability to tell consumers "this product comes from Company X." Trademark law does not ask who drew the logo. It asks whether consumers can identify a single source behind the mark.

Here is where it gets dangerous: even if a logo could function as a trademark, the lack of copyright protection underneath creates a structural vulnerability that most founders do not see until it is too late.

A logo without copyright is a logo that any competitor can copy and use in a different product category -- and the original owner has no legal basis to stop them. Trademark only blocks use on related goods or services. Copyright blocks copying, period.

This distinction drives everything that follows.

Five Reasons AI Logos Fail at CNIPA

1. No Human Authorship Means No Copyright Shield

On November 27, 2023, the Beijing Internet Court delivered China's first judgment on AI-generated image copyright (Li v. Liu). The plaintiff had used Stable Diffusion to generate a character image, then published it on Xiaohongshu. The defendant copied the image and removed the watermark.

The court ruled in favor of the plaintiff -- but not because AI output is automatically copyrightable. The court found that the plaintiff had invested substantial intellectual labor through designing specific prompts, setting individualized parameters, iterating through multiple generations, and making selective choices among the outputs. This chain of human decisions transformed the AI output into a "work of fine art" under Chinese copyright law. Damages: RMB 500.

The critical takeaway

The copyright belonged to the human who directed the process, not to the AI tool. And the court explicitly noted that the human contribution had to be substantial -- not a single prompt typed in two seconds.

For the overwhelming majority of founders who type a short prompt into Midjourney or DALL-E and accept the first appealing result, that threshold of "substantial intellectual labor" is never met. The output remains in a legal gray zone where no one owns the copyright -- including the person who paid for the subscription.

Across the Pacific, the U.S. Copyright Office has reached the same conclusion. In Thaler v. Perlmutter (130 F.4th 1039, D.C. Cir. 2025), the D.C. Circuit affirmed that a machine cannot be an author under the Copyright Act. The Supreme Court denied review on March 2, 2026, leaving that holding intact. The Copyright Office's January 2025 report, Copyright and Artificial Intelligence, Part 2, applies the same principle: a human author may obtain protection for original expression they contribute to an AI-assisted work, but protection extends only to the human-authored elements.

Prompts alone do not establish authorship. The Copyright Office stated this plainly: "With generally available technology, prompts alone are unlikely to establish authorship of the output."

2. CNIPA's Distinctiveness Test Crushes Generic AI Output

China's trademark system operates on a first-to-file basis with substantive examination. Every application is checked against over 50 million existing registrations (53.03 million valid marks as of end-2025). The examiner asks: is this mark distinctive enough to identify a single commercial source?

AI image generators are trained on massive datasets of existing logos and design patterns. They optimize for results that look "acceptable" to the widest audience. The inevitable consequence: AI-generated logos converge on the same visual language -- soft gradients, geometric swooshes, trendy sans-serif typefaces, abstract animal silhouettes.

When CNIPA examiners see a mark that looks like it could belong to any of a dozen companies in the same class, they reject it for lack of distinctiveness.

The numbers behind it

One analysis of AI-generated trademark submissions found that out of 100 schemes generated by AI tools, only 3 passed examination after human designers made substantive modifications (Source: CSDN IP Analysis). The 35% base rejection rate for all trademarks becomes far worse when the mark was generated by a system designed to produce "average" rather than "distinctive."

For a practical walkthrough of how CNIPA's examination works, the China trademark search guide covers the exact framework examiners use.

3. Prior Art Collisions You Never Saw Coming

AI models are trained on millions of existing images -- including millions of registered trademarks. When a model generates a logo, it may reproduce elements that are confusingly similar to marks already on CNIPA's register.

This is not theoretical. In February 2024, the Guangzhou Internet Court ruled that an AI platform was liable for generating images highly similar to the Ultraman character, finding infringement of both the reproduction right and the adaptation right. The court's logic was clear: the AI platform's output resembled protected intellectual property, and the platform bore responsibility regardless of whether the similarity was intentional.

What this means for trademark applicants

For trademark purposes, the risk is identical. If an AI-generated logo bears resemblance to a registered mark in the same or related class, CNIPA will reject the application based on prior art conflict. The applicant's defense -- "the AI made it, I did not copy anything" -- is not a valid legal argument. Trademark infringement is assessed by consumer confusion in the marketplace, not by the method of creation.

The U.S. Copyright Office's January 2025 report on AI and copyright put it bluntly: "The legal test is consumer confusion in commerce -- authorship of the mark is irrelevant." (Source: U.S. Copyright Office -- Copyright and AI, Part 2)

4. The Same Prompt Produces the Same Logo for Different Users

AI image generators are deterministic to a significant degree. Similar prompts produce similar outputs. A founder who prompts "minimalist coffee shop logo, brown and cream" may receive a design that looks very much like the logo another founder received from the same tool a month earlier.

This creates two distinct problems:

First, the first company to file and use the mark in commerce will typically have priority in China's first-to-file system. The second company, despite creating their logo independently, faces an uphill battle.

Second, and more insidiously, if both logos end up on CNIPA's register in related subclasses, neither party has a clean position. Both marks become vulnerable to cancellation proceedings, opposition, or mutual coexistence disputes.

The problem compounds in crowded classes. Class 25 (clothing), Class 35 (advertising and business management), Class 9 (electronics and software), and Class 42 (technology services) are already saturated with millions of existing marks. An AI-generated mark in any of these classes faces an extraordinarily high probability of collision with prior art.

5. The 2026 Trademark Law Amendment Tightens the Screws

On June 26, 2026, China's Standing Committee of the National People's Congress approved a comprehensive revision of the Trademark Law -- the first complete overhaul since the law took effect in 1983. The new law expands from 8 chapters and 73 articles to 9 chapters and 87 articles, and takes effect on January 1, 2027.

(Source:World IP Review -- China's Trademark Law Amendment)

Key changes that directly affect AI-generated marks:

The structural direction is unmistakable: China's trademark system is shifting from "registration-oriented" to "use-and-good-faith-oriented." Marks that lack genuine commercial backing, human creative investment, or clear distinctiveness face an increasingly hostile examination environment.

What the Courts Actually Said

Beijing Internet Court: The "Substantial Intellectual Labor" Standard

The November 2023 decision (Li v. Liu) established China's working framework for AI-generated content copyright. The court's reasoning hinged on a specific factual finding: the plaintiff had not merely typed a prompt and accepted the output. Instead, the plaintiff:

• Designed specific, detailed prompts involving compositional choices
• Set individualized parameters (aspect ratio, style weights, seed numbers)
• Iterated through multiple generations
• Made selective choices among outputs based on aesthetic judgment
• Performed post-generation modifications

This chain of decisions constituted "substantial intellectual labor" that transformed the AI output into a copyrightable work.

The implication for trademark

A founder who types "modern logo for my startup" and files the raw output will almost certainly fail to meet this standard. The human contribution must be documented, substantive, and traceable.

U.S. Copyright Office: Prompts Are Instructions, Not Authorship

The American position is even more restrictive. Here is a comparison of the three landmark decisions:

The Copyright Office's position as of mid-2026: over 7,000 applications containing disclosed and disclaimed AI-generated material have been registered, each with explicit limitations excluding the raw AI output.

The Platform Rights Trap

Here is the gap that catches most foreign founders: having commercial permission from an AI platform does not mean owning the copyright to the output.

Midjourney grants commercial rights to paid subscribers. DALL-E and ChatGPT assign all rights in the output to the user under OpenAI's terms. Adobe Firefly offers IP indemnification for paid subscribers. On paper, these terms look reassuring.

But "commercial rights" under a platform's terms of service is a contractual license -- not a property right. It means the platform will not sue you for using the output commercially. It does not mean:

The trap in one sentence

"If your studio simply prompted a visual and published it, you own absolutely nothing. Anyone can legally take that asset and use it without your permission." (Source: U.S. Copyright Office -- Copyright and AI, Part 2)

For trademark purposes, this creates a specific vulnerability: if your logo has no copyright protection underneath, a competitor in an unrelated industry can copy the exact design and use it -- and trademark law only blocks use on related goods or services. The gap between copyright coverage and trademark coverage is where raw AI logos fall through.

How to Make an AI Logo Actually Trademarkable

None of the above means AI tools are useless for logo design. It means the raw output needs to pass through a transformation process before it becomes a legally defensible trademark.

The modification must be substantial enough that the final mark is no longer recognizably derived from the AI output. This is the same standard the Beijing Internet Court applied -- not cosmetic adjustments, but genuine creative transformation.

A comprehensive China trademark search before filing covers identical, phonetic, and translated equivalents -- which is the exact framework CNIPA examiners use.

The Real Cost of Getting It Wrong

Consider the actual cost structure of a failed AI logo filing:

For a foreign brand entering China, the stakes are higher. A rejected mark leaves the brand exposed to squatters who can file the same or similar mark immediately. China's first-to-file system does not reward prior use -- it rewards the first valid application.

The trademark squatters problem in China is already severe. An AI-generated mark that gets rejected creates a public record that squatters can monitor and exploit.

Decision Matrix: 5 Common Scenarios

4 Common Myths About AI Logos and China Trademarks

Myth 1: "The AI platform gave me commercial rights, so I own the logo"

It does not. Commercial rights under a platform's terms of service is a contractual license -- not a property right. The platform promises not to sue you. It does not give you exclusive ownership, copyright protection, or the ability to stop competitors from using the same image. For trademark purposes at CNIPA, "the platform said I could use it" is not a defense against rejection or infringement claims.

Myth 2: "AI-generated logos are unique, so they won't conflict with existing marks"

They are not. AI models are trained on existing images -- including millions of registered trademarks. The same prompt produces similar outputs for different users. An AI-generated mark in Class 25 (clothing) or Class 35 (business management) faces an extraordinarily high probability of collision with prior art, because those classes are already saturated with visually similar marks.

Myth 3: "I just need to add a few tweaks to the AI output and it becomes mine"

Cosmetic adjustments do not meet the "substantial intellectual labor" standard established by the Beijing Internet Court. The court required a documented chain of creative decisions -- detailed prompts, individualized parameters, multiple iterations, selective choices, and post-generation modifications. Changing the color from blue to green is not "substantial intellectual labor."

Myth 4: "If it gets rejected, I can just appeal"

Appeal is possible, but it costs additional time and money -- and the underlying problems do not go away. If the mark lacks distinctiveness because it looks like every other AI-generated logo in the class, an appeal will not fix that. If it conflicts with prior art, an appeal will not remove the conflict. The office action process is designed for marks that are fundamentally registrable but need clarification -- not for marks with structural legal deficiencies.

Frequently Asked Questions

Can I trademark an AI-generated logo in China at all?

Technically, yes -- if it passes CNIPA's substantive examination. But a raw, unmodified AI output faces near-certain rejection due to lack of distinctiveness, prior art collisions, and the absence of copyright protection. The mark must undergo substantial human modification before filing.

How does China's approach to AI-generated content compare to the US?

Both jurisdictions require human authorship for copyright protection. China's Beijing Internet Court established the "substantial intellectual labor" standard in November 2023. The U.S. Copyright Office's position is arguably more restrictive -- prompts alone never establish authorship, and even 600+ iterations were deemed insufficient in Theatre D'opera Spatial. The practical effect is the same: raw AI output is unprotected in both countries.

What happens if I file a raw AI logo and CNIPA rejects it?

The rejection becomes public record. You lose the filing fee (RMB 270/class), agency fees, and 6-12 months of waiting time. More critically, the public record alerts competitors and trademark squatters to your brand interest. In China's first-to-file system, someone can file the same or similar mark immediately after your rejection.

Does the 2026 Trademark Law amendment specifically target AI-generated marks?

Not directly -- the amendment does not mention AI. But several provisions disproportionately affect AI-generated marks: Article 19 targets bad-faith filing (which includes mass AI-generated applications), Article 54 introduces direct financial penalties, and the shortened opposition period (3 to 2 months) gives opponents faster tools to challenge weak marks.

How much does it cost to properly trademark a logo in China in 2026?

CNIPA's e-filing fee is RMB 270 per class (~$40). With agency fees, the total typical range is $380-790 for one class. See the detailed breakdown in the China trademark registration cost guide.

What is the "substantial intellectual labor" standard?

Established by the Beijing Internet Court in November 2023 (Li v. Liu), this standard requires a documented chain of human creative decisions -- not a single prompt. The plaintiff succeeded because they designed detailed prompts, set individualized parameters, iterated through multiple generations, made selective choices, and performed post-generation modifications. For trademark purposes, a professional designer's transformation of an AI starting point meets this standard. A founder's single Midjourney prompt does not.

Will CNIPA know my logo was AI-generated?

Not directly -- CNIPA does not currently screen for AI origin. But the examination process will catch the symptoms: lack of distinctiveness, similarity to prior art, generic visual patterns. And if your mark is challenged later, the absence of copyright protection (because no human authorship exists) becomes a structural vulnerability that opponents can exploit.

Related guides: How to Register a Trademark in China | China Trademark Basics for Foreign Brands | China Trademark Search: CNIPA Guide | Trademark Squatters in China

Sources

• CNIPA 2025 IP Work Progress Press Conference (gov.cn)
• World IP Review -- China's Trademark Law Amendment
• Supreme People's Procuratorate -- AIGC Copyright Judicial Consensus
• U.S. Copyright Office -- Copyright and AI, Part 2 (Jan 2025)
• U.S. Copyright Office -- Theatre D'opera Spatial Review Decision

Got questions about AI-generated logos and China trademarks? Leave them in the comments below.