We invented 50 startup names. AI thought 38% of them were tech.

Across 250 industry guesses on 50 invented words, the modal LLM answer was "tech" — 78% of the time for nonsense, 40% of the time for mid-priming words, 13-30% even when the name has a literal non-tech root. AI does not survey the space of possible industries. It defaults to tech and only deviates when the name shouts something else. That has consequences if you're building anything else.

The headline numbers

Counting only labels that are purely tech-cluster — Tech, Software, Cloud Computing, etc. — and excluding compound categories like "Financial Technology" or "Lighting Technology":

The pattern is structural: the more semantic content a name carries, the less the tech default dominates — but only the strongly-descriptive bucket actually breaks the default reliably. The mid-tier is unpredictable.

Pure nonsense → confident tech

For the three words below, all five LLM samples agreed: tech. We made up the words. There is no company named "druv" or "skab" or "nyrr." The LLMs picked tech because tech is what they reach for when given an unfamiliar string and asked about a "startup."

None of these have a tech morpheme. None of them sound tech. The signal is the absence of any other signal — and the absence reads as tech.

Sometimes the semantic prior wins

For these four mid-tier words, the morpheme carried a strong-enough non-tech signal that the LLMs caught it and stayed off the default. This is the optimistic case: careful word choice does sometimes route the LLM correctly.

"Ferment" → food. "Helix" → biotech. "Vellum" → digital media (the LLMs translated "calfskin paper" into a modern document/media analog). "Lumina" → lighting specifically — not generic tech. When the morpheme is recognized, the LLMs are actually quite specific.

And sometimes the prior fails completely

But this is where the danger lives. Three more mid-tier words. Each has a real-word root with a clear semantic field. The LLMs ignored all of it and reached for tech anyway.

"Obviate" — a real Latin-root verb meaning "to avoid." Should prime prevention, automation, problem-solving. Five for five tech. "Crysta" — sounds like crystal (material, structure, mineral). Five for five tech. "Plurra" — phonetic relative of "plural" (data, language, money). Five for five tech. The same kind of prior that saved "ferment" and "vellum" failed here. We cannot tell from the morpheme alone which side of the line a name will fall on.

What this means for your name

Three things follow.

First, if you're building anything that isn't a software startup, your name is fighting a strong default. The LLMs lean tech for any name without an obvious non-tech morpheme. Consumer brands, services, physical products, vertical industries — all start with a wrong-category presumption that takes ongoing brand-work to override.

Second, "vibe" naming is unreliable. Choosing a name with the "feel" of finance or biology or media doesn't reliably transfer to the LLM's read. "Ferment" worked. "Obviate" didn't — even though both have real Latin roots with clear semantic fields. From the outside, you can't predict which side your name will land on.

Third, your AI-era brand has to do the override work. The product description, the landing-page copy, the schema markup, the OG card — these are what eventually move the LLM's category guess. The name itself isn't enough. Plan for the override budget.

The action

For a founder choosing a name today: probe your shortlist before you commit, and check what category the LLMs actually default to. If it's tech and you're building tech, fine. If it's tech and you're building anything else, you'll spend the rest of your runway fighting that first impression. If it's the right non-tech category, you got lucky — your morpheme carried.

That measurement is exactly what namedesk does — for any name, in about ten seconds.