Episode Details

Copyright Office Guidance Denies Protection for AI Written Code

A startup preserves copyright and patent rights in code an AI coding assistant wrote largely on its own only by documenting the human editing, architecture, and conception behind it, a standard the Copyright Office and the United States Patent and Trademark Office both tightened in 2025. The Copyright Office’s January 29, 2025 report concluded that a prompt alone does not create a copyrightable work, though meaningful human selection, arrangement, and editing of the output can. The USPTO’s revised guidance, effective November 28, 2025, requires that a named inventor be able to describe the claimed invention with particularity, a standard AI cannot satisfy and a human cannot satisfy merely by accepting what the AI proposed. Neither agency, and now neither the Supreme Court, will treat an engineer’s prompt as the missing ingredient.

The background is two decisions a decade apart in spirit but identical in structure. The Federal Circuit held in Thaler v. Vidal, 43 F.4th 1207 (Fed. Cir. 2022), that an inventor under the Patent Act must be a natural person because Congress wrote the statute using words like herself and himself. The D.C. Circuit reached the copyright version in Thaler v. Perlmutter, 130 F.4th 1039 (D.C. Cir. 2025), holding that the Copyright Act protects only works of human creation, a holding the Supreme Court left in place when it denied certiorari on March 2, 2026. Both cases involved a machine that operated with no human editor at all, which made them easy. What the agencies had to work out since is the harder case every startup actually faces: not an autonomous machine, but an engineer typing prompts, accepting some suggestions, rewriting others, and shipping a pull request by Friday.

The Copyright Office’s own illustration involved a bespectacled cat reading the Sunday newspaper, not a codebase, but the reasoning translates directly. A prompt controls the goal, not the expression. An engineer who prompts for a caching layer, gets one back, and ships it unmodified has authored the prompt, not the cache. An engineer who prompts for the same layer, rewrites the eviction logic, restructures the module boundaries, and rejects three of the assistant’s five suggested functions has authored something: not the whole file, but the parts that carry a human’s original selection and arrangement. Registration applicants have had to disclose and disclaim the machine generated portions since the Office’s March 16, 2023 guidance, and the disclaimed portion gets no protection at all. What survives is thinner than most founders assume: closer to the weak, thin copyright the law gives a compilation than to the full copyright in a novel.

The patent side moved on a parallel track and arrived somewhere stranger. The USPTO’s February 13, 2024 guidance told examiners to apply the joint inventorship factors from Pannu v. Iolab Corp. whenever a human worked with an AI system, weighing the AI’s contribution against the human’s as though the AI were a silent co-inventor. The Office reversed itself on November 28, 2025, holding that Pannu applies only between two or more natural persons and that running an idea through an AI system does not change the human to human inventorship analysis at all. What controls instead is the older, blunter conception standard: an inventor must be able to describe the complete, operative invention with particularity, in their own words, before recognizing what an AI system proposed stops counting as inventing it. The guidance nudges engineering teams back toward a habit the industry mostly dropped after the 2013 shift to first inventor to file: contemporaneous notes, written by the human, showing the invention took shape in that human’s mind and not merely in a chat window.

A third pressure layers on top of both agency standards. The Ninth Circuit heard argument on February 11, 2026 in Doe v. GitHub, Inc., over w