AI-generated molecules that exist only in a publication can still count as prior art.
Writing in c&en Chemical & Engineering News, Justin Krieger of Kilpatrick Townsend & Stockton LLP identifies three pressure points AI puts on chemistry patents: inventorship, obviousness, and prior art.
Inventorship and the Conception Test
Only humans can be inventors, as the Federal Circuit held in Thaler v. Vidal (2022). The USPTO's November 2025 guidance treats AI as a tool and applies the same conception test whether or not AI was used: a natural person must form a definite and permanent idea of the complete invention.
Krieger poses a hypothetical researcher who selects and validates one of 10,000 model-generated structures. In Burroughs Wellcome v. Barr (1994), the scientists who selected AZT for testing conceived AZT's use against AIDS, and the NIH researchers who ran confirming tests did not. The hypothetical researcher looks more like the Burroughs scientists. Krieger warns that a passive recipient of a model's output may leave a compound with no human inventor. Where the model selects the compound, the chemist's validation may be the moment of conception. In some unpredictable fields, the Federal Circuit has recognized, conception occurs only with a successful experiment.
Obviousness and the Evolving POSITA
Obviousness is measured against a hypothetical person of ordinary skill in the art (POSITA). Section 103 of the Patent Act provides that patentability "shall not be negated by the manner in which the invention was made," so using AI does not by itself make an invention obvious. If AI design tools become standard, the POSITA arguably has them too, and the bar for a new compound could rise.
Defensive Prior Art and Enablement
Krieger describes companies publishing unsynthesized AI-generated variants as defensive prior art, and asserts that merely disclosing a compound is "likely insufficient" to enable synthesis and use. In Agilent v. Synthego (2025), a CRISPR case, the Federal Circuit applied a lower bar for prior art. A publication is presumed enabling, and the reference need only teach how to make the disclosed subject matter. Prophetic examples, experiments described but never run, left the presumption intact. The patent owner can rebut the presumption with evidence of undue experimentation. Agilent's Supreme Court petition warned that AI makes candidate-compound lists trivial to publish, and the Court denied review in March 2026.
Krieger concludes that AI-assisted discoveries remain fully patentable so long as human intelligence guides, curates, interprets, and validates the process.
Actionable Steps for Chemistry R&D and IP Teams
- Record who set screening criteria and why a selected structure beat rejected ones.
- Search AI-generated disclosures when scoping protection for a compound. A new use of a disclosed compound may still be patentable.
Sources
- 1. Justin Krieger, ๐๐ฉ๐ฆ ๐๐ ๐ฑ๐ข๐ต๐ฆ๐ฏ๐ต ๐ฑ๐ข๐ณ๐ข๐ฅ๐ฐ๐น: ๐๐ข๐ญ๐ข๐ฏ๐ค๐ช๐ฏ๐จ ๐ช๐ฏ๐ฏ๐ฐ๐ท๐ข๐ต๐ช๐ฐ๐ฏ ๐ข๐ฏ๐ฅ ๐ช๐ฏ๐จ๐ฆ๐ฏ๐ถ๐ช๐ต๐บ ๐ช๐ฏ ๐ค๐ฉ๐ฆ๐ฎ๐ช๐ด๐ต๐ณ๐บ, Chemical & Engineering News (Sept. 8, 2026): https://cen.acs.org/policy/intellectual-property/ai-intellectual-property-chemistry-patents/104/web/2026/09
- 2. ๐๐ฉ๐ข๐ญ๐ฆ๐ณ ๐ท. ๐๐ช๐ฅ๐ข๐ญ, 43 F.4th 1207 (Fed. Cir. 2022)
- 3. USPTO, Revised Inventorship Guidance for AI-Assisted Inventions, 90 Fed. Reg. 54636 (Nov. 28, 2025): https://www.federalregister.gov/documents/2025/11/28/2025-21457/revised-inventorship-guidance-for-ai-assisted-inventions
- 4. ๐๐ถ๐ณ๐ณ๐ฐ๐ถ๐จ๐ฉ๐ด ๐๐ฆ๐ญ๐ญ๐ค๐ฐ๐ฎ๐ฆ ๐๐ฐ. ๐ท. ๐๐ข๐ณ๐ณ ๐๐ข๐ฃ๐ด., ๐๐ฏ๐ค., 40 F.3d 1223 (Fed. Cir. 1994); Amgen Inc. v. Chugai Pharm. Co., 927 F.2d 1200 (Fed. Cir. 1991)
- 5. 35 U.S.C. ยงยง 100(b), 103
- 6. ๐๐จ๐ช๐ญ๐ฆ๐ฏ๐ต ๐๐ฆ๐ค๐ฉ๐ด., ๐๐ฏ๐ค. ๐ท. ๐๐บ๐ฏ๐ต๐ฉ๐ฆ๐จ๐ฐ ๐๐ฐ๐ณ๐ฑ., 139 F.4th 1319 (Fed. Cir. June 11, 2025), ๐ค๐ฆ๐ณ๐ต. ๐ฅ๐ฆ๐ฏ๐ช๐ฆ๐ฅ, No. 25-570 (U.S. Mar. 30, 2026): https://www.cafc.uscourts.gov/opinions-orders/23-2186.OPINION.6-11-2025_2528805.pdf
