Patents and IP for AI companies

LAST REVIEWED 2026-08 · SOURCED FROM 1 SESSION, MAY 2026 · CASE LAW MOVES — CHECK DATES

TL;DR

Training-method innovations that improve how a model operates are patent-eligible; feeding a known model new data or pointing it at a new vertical is not. Frame everything as a technological advantage (less compute, less storage, lower error rate), never a business advantage (saves the user time). File a provisional (~$6K in the AI space) before any collaboration or public pitch. AI cannot be an inventor, and free AI tools’ terms of use may claim your output.

What’s eligible after the recent case law

The line drawn in the session: training a machine-learning model is eligible when it integrates a mathematical concept into a practical application that improves how the model itself operates. Merely running new input data through a known model, or applying an existing model to a new industry, is not eligible.

A decision handed down weeks before the session sharpened the drafting lesson: results-oriented claims — describing what is achieved — were invalidated even though the “how” lived in the specification. Eligibility is a claims question. The fix: broad independent claims layered over increasingly structural dependent claims, so something survives.

Regulatory wind matters too. The patent office’s current leadership has signaled a pro-AI-patentability shift — the session quoted the director’s warning that categorically excluding AI innovation would jeopardize American leadership in the technology — and discretionary denials of the inter-partes-review challenges nicknamed “patent death squads” are rising, which favors patent holders.

Frame the advantage as technological

The examiner distinction that decides cases:

Framing	Reception
“Saves the operator hours”	Business advantage — not persuasive
“Reduces processing time, storage, or error rate”	Technological advantage — persuasive

Same invention, different sentence. If your edge is embedded human expertise, frame it as reducing error rates, not saving time.

The traps specific to AI companies

• AI cannot be an inventor. Improper inventorship invalidates a patent — track which contributions are AI-generated as you build.
• Free tools’ terms of use. The blunt version from the room: if you’re using something for free, assume the provider claims ownership of everything it generates — and your inputs may train their models, jeopardizing trade secrets.
• Before any collaboration: file a patent application first as a marker, and put ownership of anything generated together in the contract.
• Provisionals protect only what’s filed. Don’t publicly present later inventions that aren’t in the provisional.

“A stronger provisional patent application is a better provisional patent application.”

— Patent attorney, IP boutique · session, May 2026

And on NDAs, from a founder in the room: an NDA is only as good as your pockets are deep to enforce it.

What patents are actually for at this stage

Startups rarely litigate. The early value is investor validation and exclusivity signaling. Creative enforcement paths exist later — licensing to an infringer’s biggest competitor, or litigation-funding entities — but VCs weigh execution over patent moats, especially for software.

Numbers from the room

Figure	Value (AI practice area, 2026)
Full non-provisional	~$12K ±10–20%
Typical provisional	~$6K
Rush “cover-the-pitch” provisional	Much less

Sources

One session (May 2026) with two patent attorneys at an IP boutique — one with ~28 years in software/AI patents, one in life sciences — covering Section 101 eligibility, recent case law, and filing strategy.