Original research

State of Software Patentability (2026)

Short answer: AI-assisted software is patentable when human conception and prior-art clearance are strong — but most first-pass scores fail on §103 or documentation, not because AI was used.

Aggregate signals from Patent PreCheck scoring and USPTO corpus analysis. Published 2026-06-22. Directional statistics from anonymized free-tier analyze runs and corpus audit snapshots. Not legal advice or grant-rate predictions.

Key findings

42%
§103 non-obviousness

of AI-assisted software modules score below our non-obviousness band on first pass — prior-art overlap is the dominant weak pillar.

28%
§101 Alice eligibility

fail eligibility on first score when the submission is primarily CRUD, dashboard, or generic API glue without a technical hook.

35%
Human conception

show weak human-conception documentation signals when builders paste agent output without design notes.

1 in 4
Prior-art proximity

free scores surface a granted-patent or application match above 0.75 semantic similarity in the top-5 neighbors.

Corpus behind the scores

Patent PreCheck retrieval spans 1.05M+ active documents across 40+ implemented workers (patents, papers, code, standards). Daily scholarly backfill + weekly USPTO sprint.

Recommendations for builders

  1. Score the inventive module, not README marketing copy.
  2. Document rejected AI suggestions and benchmark-driven design choices.
  3. Run prior-art search before public launch — GitHub repos are §102 prior art.
  4. Use MCP/CI precheck in Cursor and GitHub Actions to catch weak pillars pre-filing.

Hub: AI-code patentability guides · Scoring methodology

Last updated: 2026-06-22

Run a free score on your code →