The Self-Improving AI Monitoring Act is a House measure circulated on August 31, 2026 that would amend section 22A of the National Institute of Standards and Technology Act (15 U.S.C. 278h–1) to direct the Director to assess trends in AI capabilities — naming in particular a system's capability to autonomously conduct AI research and development — and to prohibit the Director from entering memorandums of understanding for pre-deployment evaluation of frontier systems unless those memorandums carry specified disclosure and evaluation requirements (Self-Improving AI Monitoring Act — discussion draft bill text (August 31, 2026)).
Status and provenance
The text located is an unnumbered discussion draft. Its face reads "H. R. ll" with a blank committee referral and a blank original-signature line, so no bill number attaches to it, and its introduction line names only Representative George Whitesides (D-Calif.). Trade coverage on September 4, 2026 described the measure as introduced on August 31, 2026 and attributed it to Representatives Pat Harrigan (R-N.C.) and Whitesides jointly; the file is hosted on Harrigan's House subsite (Source: nextgov.com). No introduced counterpart carrying a bill number has been located. The short title appears in the bill text and in none of the coverage located.
Key provisions
Assessment of capability trends. A new paragraph (6) of section 22A(b) would direct the Director to assess "trends regarding the capabilities of artificial intelligence systems, including such a capability to autonomously facilitate or conduct research on or development of an artificial intelligence system, that might contribute to scientific research and development breakthroughs," drawing on the disclosures obtained under the new subsection (h) and on other information. The object is the trend across systems rather than any individual system.
Conditions on memorandums of understanding. A new subsection (h) would provide that the Director "may not enter into" a memorandum of understanding with the developer of a frontier AI system — frontier status being "as determined by the Director" — to conduct pre-deployment evaluation unless two conditions hold. First, the memorandum must require the developer, on request, to disclose information about the extent to which an AI system was used to facilitate or conduct research on or development of the frontier system, illustrated by three examples: a metric the developer already maintains in the ordinary course of business, an estimate of use without human review, and the methodology behind that estimate. Second, the evaluation must itself cover the frontier system's capability to autonomously facilitate or conduct AI research and development.
The obligation runs against the Director rather than against developers. A developer is reached only through a memorandum it chooses to enter, so the measure's practical scope depends on how many developers seek pre-deployment evaluation from NIST's Center for AI Standards and Innovation. Trade coverage described the measure as one that "would allow NIST to impose requirements on certain memorandums of understanding"; the text is drafted as a prohibition on the Director rather than as a grant of permission (Self-Improving AI Monitoring Act — discussion draft bill text (August 31, 2026)).
Relation to other federal instruments
Section 22A of the NIST Act is the section carrying the authorities exercised by the Center for AI Standards and Innovation, so the measure operates on an existing evaluation relationship rather than establishing a new regulator. It addresses the same object as the frontier-evaluation provisions of the Obernolte–Trahan discussion draft (Frontier Act / Great American AI Act (Obernolte–Trahan)) through a different lever, conditioning the agency's authority to contract rather than imposing duties on developers directly.
After the draft circulated, two developers published measures of the kind subsection (h) would have them disclose. OpenAI reported internal metrics of agent use in its research on September 6, 2026 (Research acceleration: The view inside OpenAI (OpenAI, September 2026)), and Anthropic on September 17, 2026 published a prototype R&D Automation Index rating how much of its AI R&D work Claude performs, with its methodology, and proposed that any frontier developer could publish comparable figures regularly (Measurements for understanding the pace of AI development inside frontier labs (Anthropic Institute, September 2026)). Neither disclosure refers to the draft.
Relationships
- related: NIST CAISI (Center for AI Standards and Innovation) — the body whose authorities the amended section carries.
- related: Recursive Self-Improvement (RSI); Automated Alignment Research — the capability the draft names.
- related: AI Pre-Release Vetting; Independent Verification Organizations (IVOs) — the evaluation regime the memorandums sit within.
- related: Frontier Act / Great American AI Act (Obernolte–Trahan) — the other live federal frontier-evaluation instrument.
- anchor-sources: Self-Improving AI Monitoring Act — discussion draft bill text (August 31, 2026).
- related: Research acceleration: The view inside OpenAI (OpenAI, September 2026), Measurements for understanding the pace of AI development inside frontier labs (Anthropic Institute, September 2026), Anthropic R&D Automation Index — voluntary developer disclosures of AI use in their own R&D.