ARPA-H Will Let LLMs Read Your IGoR Proposal: How the Federal AI-Review Firewall Just Cracked

When the Advanced Research Projects Agency for Health launched the Intelligent Generator of Research program (IGoR) on May 5, 2026, the press attention focused on the program's $3 million-per-year-or-thereabouts ambition to use AI to rewrite the biomedical research workflow. Solution summaries are due June 25, 2026; full proposals on August 6, 2026; performance is structured across three phases over five years, with awards going to roughly three multidisciplinary teams capable of addressing all four technical areas. What did not make most of the headlines is a single-page disclosure quietly published alongside the solicitation: ARPA-H, when it reviews IGoR proposals, intends to use secure large language model tools to assist with initial review of submitted materials. ARPA-H is the first federal grantmaking agency to disclose that it will let generative AI systems read incoming proposals as part of the formal review process — and the disclosure marks the most consequential federal proposal-review policy departure in two years.

The context matters. In June 2023, NIH issued a guide notice (NOT-OD-23-149) banning peer reviewers from uploading any portion of confidential grant applications or critiques into "online generative AI tools" — ChatGPT and its cousins. The agency's stated rationale was confidentiality: the moment proposal text enters a commercial LLM's input pipeline, the agency cannot guarantee that proprietary scientific ideas and applicant identities remain protected. NSF followed a similar posture, declining to authorize AI use in peer review and forming an internal working group in late 2024 to evaluate whether and how generative AI might safely participate in merit review. Through 2025 and into early 2026, NSF's working group floated proposals for "guardrails" without committing to any deployment. The Department of Energy's Office of Science adopted NIH-style prohibitions in late 2025. The American science-funding establishment, in other words, settled on a default position of "no AI in peer review" and then began the slow process of negotiating exceptions.

ARPA-H just broke ranks.

What ARPA-H Actually Said

The disclosure is short. ARPA-H stated that "IGoR will pilot the use of secure large language model (LLM) tools to assist with the initial review of submitted materials." The agency specified that the tools will help "organize, summarize, and surface key information from solution summaries" — the brief 5-to-10-page documents proposers submit by June 25 before being invited to write full proposals. ARPA-H emphasized that "human experts will remain responsible for all evaluation and decisions" and that review panels will include specialists in contracting, regulation, engineering, product design, informatics, and biomedical research. The agency framed the pilot as part of a broader federal learning exercise: "Beyond advancing an AI-enabled biomedical research ecosystem, IGoR is helping ARPA-H and the broader U.S. government understand how to use AI responsibly at scale."

The framing is careful. ARPA-H did not say AI will score proposals, did not say AI will rank applicants, and did not say AI will replace human reviewers. The disclosed function is closer to a high-end document-triage and summarization assistant: an LLM reads a solution summary, produces structured outputs that human reviewers can use to navigate the document more efficiently, and surfaces specific claims or proposal features that warrant deeper human attention. That posture is closer to how legal-discovery review tools have worked for a decade than to how AI systems pick stock trades or write code.

But the careful framing should not obscure what has actually changed. For the first time in the federal grant-making system, a U.S. agency has officially disclosed that confidential proposal text will be processed by a generative AI system as part of the formal review pipeline. The word "secure" is doing important work in the disclosure — it implies a hosted or air-gapped deployment of the model rather than a public cloud API call to ChatGPT or Claude — but the policy precedent is set: confidentiality concerns are now solvable with the right deployment architecture, and the categorical NIH-style ban is, at the federal level, no longer the default.

Why ARPA-H Could Move First

ARPA-H is an unusual creature in the federal funding landscape. Established in 2022 within HHS as an analog to DARPA for biomedical research, the agency uses Other Transaction agreements rather than standard federal grants for most of its programs, including IGoR. OT agreements live outside the Federal Acquisition Regulation and outside the standard Uniform Guidance framework that governs traditional grants. ARPA-H also operates with a "Program Manager" model borrowed from DARPA — empowered technical leaders who design programs, write solicitations, and manage performer portfolios with substantially more latitude than program officers at NIH or NSF have.

That structural latitude is why ARPA-H can deploy an AI-assisted review pilot months or years before NIH or NSF would clear comparable internal review. ARPA-H does not need an external advisory board to sign off on a peer-review process change. The agency does not need to consult institutional review boards or notify a congressional committee. The Program Manager for IGoR decided, with ARPA-H Director Alicia Jackson's authorization, to pilot the tool inside a single program with a defined applicant pool and a transparent disclosure. The decision is reversible, scoped, and easier to defend if something goes wrong than a sweeping policy change at NIH would be.

The choice of IGoR as the pilot venue is also strategic. IGoR is a program about AI-enabled research. Using AI to triage AI-research proposals is thematically coherent in a way that using AI to triage cardiology grant applications at NHLBI would not be. The applicant population for IGoR is technologically sophisticated and self-selected for openness to AI deployment in research workflows. If something embarrassing happens — an AI summary mischaracterizes a key claim, a confidentiality concern surfaces, a bias issue emerges — the reputational damage is contained to a single program rather than spilling across the agency's full portfolio.

What This Means for Other Federal Programs

The IGoR disclosure sets a precedent that NIH, NSF, DOE, and other agencies will now have to navigate. The path forward is not necessarily uniform adoption. NIH's institutional culture around confidentiality and the strength of the peer-review system's reputation make a fast pivot unlikely. NSF's working group will probably take twelve to twenty-four months to publish formal guidance. DOE Office of Science will likely move slower than ARPA-H but faster than NIH.

But the categorical ban is dead. Once one federal agency has publicly said "secure LLMs can read proposals," every other agency operates inside a world where the question is "under what conditions and with what controls," not "yes or no." Internal AI-deployment plans at NIH and NSF will accelerate. RFP language at other agencies will start including AI-review disclosures, both to align with where federal policy is heading and to head off the reputational risk of being caught using AI tools without telling applicants.

What Applicants Should Do Differently

For teams writing IGoR solution summaries by June 25, the AI-assisted review pipeline implies several practical adjustments. First, structure matters more than it did in a pure-human review process. LLMs extract information more reliably from clearly headed sections, bulleted lists, and tables than from dense narrative prose. A solution summary that reads beautifully but buries the technical approach in long paragraphs may produce a less effective AI-generated summary than one that is more clinically organized — and the AI summary is what human reviewers will see first.

Second, technical claims should be made explicit. AI summarization is good at extracting clearly stated claims and weaker at inferring claims from context. If the proposal's central innovation is "ten times faster validation of biological knowledge than traditional methods," that sentence should appear, verbatim or near-verbatim, in the executive summary, the technical approach section, and the impact section.

Third, applicants should expect that human reviewers will spend less time on first-pass document navigation and more time on substantive evaluation of claims surfaced by the AI tool. The premium on having those substantive claims defensible — supported by preliminary data, citations, or clear technical reasoning — is higher than it would be in a purely human review pipeline, because reviewers will reach the substantive-evaluation stage faster and with more concentrated attention.

For the broader federal-grants applicant community, the IGoR disclosure is a signal flare. The agencies that fund your research are going to start using AI tools to read your proposals. Some agencies will disclose this; some, at least initially, will not. The ARPA-H pilot is the public-facing edge of a transition that will reshape how federal proposal review functions over the next three years — and it began on a single disclosure page next to a solicitation almost no one is reading.