AI Named Your Startup. Now the USPTO Just Rejected It.

In 2023, OpenAI filed a trademark application for the name ChatGPT. It was rejected. They also filed for GPT. Also rejected. The examining attorney's reasoning, paraphrased: the marks are merely descriptive, they describe what the product is rather than identifying who provides it.

This is not an isolated case. The United States Patent and Trademark Office (USPTO) has been quietly rejecting AI-generated startup names at a rate that intellectual-property lawyers are starting to write nervous papers about. The pattern is consistent enough that you can predict, before filing, whether your AI-generated name will clear, and most of the time, it won't.

Here's why, and what to do about it.

The Descriptiveness Problem

A trademark protects an identifier that distinguishes one company from another. The USPTO refuses registration when a mark merely describes the product, the service, or a characteristic of it. The legal phrase is "merely descriptive", and it's the single most common ground for refusal.

Words and phrases that get rejected for descriptiveness include:

• Cold Beer for a beverage company.
• FastTax for tax preparation software.
• ChatGPT for, well, a chatbot that uses generative pre-trained transformers.

The USPTO's logic is straightforward: if competitors need the same word to describe what they sell, you can't lock that word up. Cold and Beer belong to everyone who sells beer. Chat and GPT belong to everyone who builds chatbots on transformers.

This is where AI generators run into a structural problem.

Why AI Generators Produce Descriptive Names

Most AI name generators work by combining or transforming words from the brief you give them. If you ask for names for an organic juice bar, the model will draw from a probability distribution centred on words like fresh, pure, green, organic, juice, vita, blend. The names it generates, FreshBlend, PureGreen, VitaJuice, OrganicFlow, are, by construction, descriptive.

The model isn't doing anything wrong. It's doing exactly what large language models do: producing the most probable next tokens given the prompt. The problem is that the tokens most probable to describe your category are also the exact tokens most likely to be refused by the USPTO.

The cleverest-feeling generated name, the one that captures your product perfectly in one neat compound, is usually the most descriptive, and therefore the most likely to be rejected.

A few patterns AI generators reliably produce that the USPTO reliably rejects:

• Two category words mashed together. PayBridge, FintechFlow, DataPulse, EdTechGenius. These are descriptive by definition.
• Category word + intensifier. SuperLegal, MaxResume, ProTax. The intensifier is non-distinctive; the rest is descriptive.
• Category word + tech suffix. Legalify, Resumeify, Taxify. Suffix doesn't save a descriptive root.
• Category word + AI / Tech / Lab / Cloud. LegalAI, ResumeAI, TaxAI. These are now near-automatic refusals.
• Adjective that describes a product attribute. PureLegal, FastTax, EasyLegal. Adjectives describing the product attributes are textbook descriptive.

If your generated shortlist is mostly in those buckets, your shortlist is mostly trademarkable-on-paper-but-rejectable-in-practice.

The Search-Coverage Problem

There's a second, subtler issue with AI-generated names: even when they aren't descriptive, the AI usually didn't check the trademark register before suggesting them.

Modern AI generators may include a domain-availability check (most do) and a social-handle check (fewer do, Nymly is one of them). But almost none check the USPTO TESS database. That means:

• A name you love and the domain is available can still be trademarked by someone else.
• The AI doesn't know about pending applications that haven't yet been published.
• Common-law trademarks, rights established by use without registration, are completely invisible to most generators.

What you assume is a fresh, clear name often has an existing mark sitting one phonetic vowel away in your International Class. When you file, the examining attorney finds it, cites a likelihood-of-confusion refusal, and your application is dead.

The Cases That Made Lawyers Pay Attention

A few representative refusals from the last two years (specifics paraphrased for narrative; the patterns are real):

• A skincare startup filed an AI-generated name combining Glow + Lab. Refused, descriptive, and confusingly similar to three existing Glow-prefixed cosmetics marks.
• A code-assistant tool filed CodeMate AI. Refused, Code is descriptive of the service, AI is descriptive of the technology, and Mate is a non-distinctive companion modifier. The whole mark adds up to "a chatbot for coders".
• A real-estate analytics startup filed PropPulse. Refused, Prop (property) and Pulse (analytics signal) are both descriptive. The combination doesn't add inherent distinctiveness.

In every case, the founder said the same thing afterwards: "I assumed if the domain was free and Google didn't show another company, I was fine." That's not how trademark law works, and it's the single most expensive mistake in the early stages of a brand.

What Actually Clears

The flip side is also true: the names that do clear the USPTO at a high rate tend to share characteristics.

• Arbitrary names. Real words used in a context unrelated to their meaning. Apple for computers. Tesla for cars. Stripe for payments. Slack for software.
• Coined (fanciful) names. Made-up words. Kodak. Spotify. Verizon. Nymly. These are the cleanest possible trademarks because they have no inherent meaning to anyone but you.
• Suggestive names. Hint at a quality without describing it. Google (suggests vastness via googol). Netflix (suggests "net" + "flicks" without describing streaming). Lovable (suggests something endearing without describing AI app generation).

If you look at modern tech naming, Cursor, Linear, Arc, Notion, Lovable, every name is either arbitrary or coined or suggestive. None are descriptive. That's not coincidence. The founders who run the trademark math first reliably end up in this category.

The Seven-Point Clearance Checklist

Before you file, before you fall in love with the domain you just bought, run this seven-point check. It catches roughly ninety percent of the rejection patterns above.

1. Read your name out loud and ask: "Does this say what I do?"

If yes, it's probably descriptive. Try harder.

2. Search USPTO TESS for the exact mark

Go to tmsearch.uspto.gov. Search your exact name. If anything comes up in Class 9 (software), Class 35 (business services), Class 42 (technology services), or whichever class is yours, stop. Investigate.

3. Search for phonetic equivalents

The USPTO refuses on sound-alike, not just exact match. Quora and Korra would conflict. Run searches for likely misspellings and homophones.

4. Search the goods-and-services description

If a competitor described their product the way you'd describe yours, they may have a mark that conflicts even if it doesn't share letters with yours. Search the actual product descriptions.

5. Check the EUIPO and WIPO databases

If you intend to operate outside the US, do the same searches at the European Union Intellectual Property Office and the World Intellectual Property Organisation. A name that clears in the US can still cost you in Europe.

6. Common-law search

A careful Google for the exact name plus your industry, plus the words trademark, brand, company. Common-law rights exist without registration, and they can block your filing.

7. Pay a trademark attorney two hours

This is the step that pays for itself ten times over. A trademark attorney will catch what you missed, including indicators like recent filings that haven't published, parallel proceedings, and unusual examining-attorney patterns in your class. Two hours at $400/hr beats a rebrand at $40,000.

Where AI Naming Tools Still Earn Their Keep

It's tempting to read the above and conclude that AI name generators are net-negative. They aren't. The right way to use them is as a creativity engine for the brainstorm step, not as a finished-names engine for the launch step. The tools differ a lot in how they handle this step, we compare the main ones in our best AI business name generators roundup.

A good workflow:

1. Use an AI generator (we recommend Nymly, naturally) to produce 100-300 candidate names.
2. Filter the output yourself, looking specifically for the arbitrary, coined, and suggestive candidates, not the descriptive compound-word ones.
3. Check domain and social availability, Nymly does this in the same pass.
4. Run the seven-point checklist on your survivors.
5. File the survivor, or hire a lawyer to file it for you.

Our 2026 AI engine upgrade was largely about tilting our generator away from descriptive compound names and toward the arbitrary/coined territory the USPTO actually clears. The model isn't a substitute for the seven-point check, but it does give you a longer list of clearable candidates to start from.

The Underlying Lesson

Trademark law was written before generative AI existed, and yet, almost spookily, it rewards exactly the kind of distinctive, non-descriptive naming that the best human brand strategists already favoured. The USPTO and the modern indie-hacker naming aesthetic (single-word, hard-consonant, open-vowel) are pulling in the same direction.

If you're naming a startup in 2026, the AI is an amplifier of your taste, not a substitute for it. The faster you accept that, the cheaper your first round of legal fees will be.

And before any of this, start with the pre-launch naming runbook. The order of operations is the entire game.