Advanced Research Projects Agency for Health AI policy

For applicants

• Some ISOs bar generative AI text. HEARING says proposers shall not generate solution summary and proposal text using generative AI. STOMP asks proposers not to generate solution summary text or proposal text using generative AI, and AIR asks them not to generate proposal text using generative AI.[1][2][3]
• For STOMP, ARPA-H's FAQ added that proposers are expected to cite where generative AI was used to generate figures.[4]

For reviewers

• ARPA-H may use AI to help organize, summarize, and surface key information from proposals. It says it will disclose this in the ISO and that human experts remain responsible for all technical evaluations and final decisions.[5]
• In the ADVOCATE and Delphi ISOs, ARPA-H disclosed that large language models may be used to assist reviewers and will not substitute for human judgement. ADVOCATE says they may be used to extract data elements, summarize text, and generate text. For IGoR, ARPA-H said it would pilot secure LLM tools to assist with the initial review of submitted materials, and that human experts remain responsible for all evaluation and decisions.[6][7][8]
• In the ADVOCATE, IGoR, and Delphi ISOs, ARPA-H says that when LLMs are used, all data goes only through secure API calls with web searches disabled, data sent for inference is never used to train public models, and all access, retention, and logging are tightly restricted and fully auditable.[7][8][9]

Sources

1. [1]

   HEARING ISO (ARPA-H-SOL-26-154), Appendix B: Proposal Format, Instructions, and Evaluation (Amendment 03)

   Advanced Research Projects Agency for Health, July 24, 2026

   “Proposers shall not generate solution summary and proposal text using generative AI.”

2. [2]

   STOMP ISO (ARPA-H-SOL-26-152), Amendment 3

   Advanced Research Projects Agency for Health, June 2, 2026

   “Please do not generate solution summary text using generative AI.”

   “Please do not generate proposal text using generative AI. Proposers are encouraged to submit proposals shorter than the page limit.”

3. [3]

   AIR ISO (ARPA-H-SOL-26-146), Amendment 03

   Advanced Research Projects Agency for Health, March 3, 2026

   “Please do not rehash the ISO material without providing explicit solutions for the challenges presented. Please do not generate proposal text using generative AI.”

4. [4]

   STOMP FAQs (archived copy; the live page now returns Access denied)

   Advanced Research Projects Agency for Health, May 11, 2026

   “As stated in the ISO, proposers should not generate solution summary or proposal text using generative AI. Proposers are expected to cite where generative AI has been used to generate figures.”

5. [5]

   ARPA-H 101: Understanding proposal evaluation criteria

   Advanced Research Projects Agency for Health, June 22, 2026

   “ARPA-H may also use artificial intelligence (AI) to help organize, summarize, and surface key information from proposals. The use of AI will be disclosed in the ISO. Human experts remain responsible for all technical evaluations and final decisions.”

6. [6]

   Use of AI tools in IGoR proposal reviews

   Advanced Research Projects Agency for Health

   “IGoR will pilot the use of secure large language model (LLM) tools to assist with the initial review of submitted materials.”

   “LLM tools will help organize, summarize, and surface key information from solution summaries. Human experts will remain responsible for all evaluation and decisions.”

7. [7]

   ADVOCATE ISO (ARPA-H-SOL-26-142), Amendment 02

   Advanced Research Projects Agency for Health, February 23, 2026

   “LLMs may be used to extract data elements, summarize text, and generate text to assist reviewers. LLMs will not substitute for human judgement.”

   “all data will be transmitted exclusively through secure API calls, with external tool calls such as web searches disabled. Vendor contracts and technical controls enforce strict security measures: data sent for inference is never used to train public models, and all access, retention, and logging are tightly restricted and fully auditable.”

8. [8]

   Delphi ISO (ARPA-H-SOL-26-153)

   Advanced Research Projects Agency for Health, April 1, 2026

   “Vendor contracts and technical controls enforce strict security measures: data sent for inference is never used to train public models, and all access, retention, and logging are tightly restricted and fully auditable.”

   “Any use of LLMs will be used to assist reviewers and will not substitute for human judgement. There will be human review of original submissions.”

   “all data will be transmitted exclusively through secure API calls, with external tool calls such as web searches disabled. Vendor contracts and technical controls enforce strict security measures: data sent for inference is never used to train public models, and all access, retention, and logging are tightly restricted and fully auditable.”

9. [9]

   IGoR ISO (ARPA-H-SOL-26-155), Amendment 02

   Advanced Research Projects Agency for Health, July 8, 2026

   “When large language models (LLMs) are used (to extract data elements, summarize text, and generate text), all data will be transmitted exclusively through secure API calls, with external tool calls such as web searches disabled.”

   “all data will be transmitted exclusively through secure API calls, with external tool calls such as web searches disabled. Vendor contracts and technical controls enforce strict security measures: data sent for inference is never used to train public models, and all access, retention, and logging are tightly restricted and fully auditable.”