May funding radar + a citation verification workflow for grant proposals.
Today's edition: what agencies are saying about AI and citations, plus what's new across DOT, NSF, NIH, and ARPA-H.
What’s Inside:
DOT: SBIR opens June 3, proposals due July 7.
NSF: $1.5B X-Labs initiative for independent research teams.
ARPA-H, Dept. of Ed, NSF pitches: June deadlines and status updates.
NIH + NSF: New policies on AI use and citation integrity.
Citation verification workflow: Three categories of reference errors, a step-by-step check, and current agency disclosure requirements.
🗓️ Funding & Connection Opportunities Round-Up
DOT SBIR opens June 3 - proposals due July 7
What: The Department of Transportation opened its FY26 SBIR pre-solicitation period, running through May 29. Full submissions are expected to open June 3, with proposals due July 7.
So what: DOT is one of the smaller SBIR agencies, with fewer applicants per topic and less competition than DoD or NIH. If your technology addresses transportation safety, infrastructure, autonomous vehicles, or mobility, it’s worth checking whether the topics align. The narrower topic set means fewer opportunities, but also a less crowded field.
Do: Review the pre-release topics. If there’s a fit, you have about five weeks from open to deadline. How to apply.
NSF launches $1.5B X-Labs for independent research teams
What: NSF announced X-Labs, a new initiative funding independent research organizations (not universities) with large, multiyear awards. The mechanism is Other Transactions Agreements, not standard grants. First topics: Scientific Instrumentation for Sensing and Imaging, and Quantum Systems (Interconnects and Integrated Photonics). Total commitment: up to $1.5B over the next decade, with additional topics expected in the coming weeks.
So what: This is a different model from SBIR. Bigger awards, longer timelines, milestone-driven, and designed for entrepreneurial teams pursuing platform-level scientific capabilities. If you’ve been building outside the university system and SBIR felt too small or too short for your work, X-Labs may be a better fit. The Sensing and Imaging topic is broad enough to cover a range of instrumentation approaches.
Do: Read the funding opportunity on SAM.gov and register for the introductory webinar at the NSF X-Labs page. Future topic announcements are expected soon.
May funding radar + a citation verification workflow for grant proposals.
Today’s edition: what agencies are saying about AI and citations, plus what’s new across DOT, NSF, NIH, and ARPA-H.
Read online Estimated Reading Time: 7 minutes
What’s Inside:
DOT: SBIR opens June 3, proposals due July 7.
NSF: $1.5B X-Labs initiative for independent research teams.
ARPA-H, Dept. of Ed, NSF pitches: June deadlines and status updates.
NIH + NSF: New policies on AI use and citation integrity.
Citation verification workflow: Three categories of reference errors, a step-by-step check, and current agency disclosure requirements.
🗓️ Funding & Connection Opportunities Round-Up
DOT SBIR opens June 3 - proposals due July 7
What: The Department of Transportation opened its FY26 SBIR pre-solicitation period, running through May 29. Full submissions are expected to open June 3, with proposals due July 7.
So what: DOT is one of the smaller SBIR agencies, with fewer applicants per topic and less competition than DoD or NIH. If your technology addresses transportation safety, infrastructure, autonomous vehicles, or mobility, it’s worth checking whether the topics align. The narrower topic set means fewer opportunities, but also a less crowded field.
Do: Review the pre-release topics. If there’s a fit, you have about five weeks from open to deadline. How to apply.
NSF launches $1.5B X-Labs for independent research teams
What: NSF announced X-Labs, a new initiative funding independent research organizations (not universities) with large, multiyear awards. The mechanism is Other Transactions Agreements, not standard grants. First topics: Scientific Instrumentation for Sensing and Imaging, and Quantum Systems (Interconnects and Integrated Photonics). Total commitment: up to $1.5B over the next decade, with additional topics expected in the coming weeks.
So what: This is a different model from SBIR. Bigger awards, longer timelines, milestone-driven, and designed for entrepreneurial teams pursuing platform-level scientific capabilities. If you’ve been building outside the university system and SBIR felt too small or too short for your work, X-Labs may be a better fit. The Sensing and Imaging topic is broad enough to cover a range of instrumentation approaches.
Do: Read the funding opportunity on SAM.gov and register for the introductory webinar at the NSF X-Labs page. Future topic announcements are expected soon.
Also on your calendar
ARPA-H has multiple programs with June deadlines. HEARING proposer day is June 8 (register by June 3 for in-person, June 5 for virtual), with solution summaries due June 29. IGoR solution summaries are due June 25. Full list at ARPA-H events and open funding opportunities.
Dept. of Education SBIR - Phase IA and Phase IB proposals due June 29 (11am ET). Direct to Phase II also due June 29 (2pm ET). Ed-tech and learning science, $250K awards for 9 months. Solicitation info.
NSF Project Pitches remain paused. As of May 18, NSF is not accepting new pitches. Both the new solicitation and the pitch window are expected “soon,” but no date has been given.
Pipeline (no current deadlines): USDA Phase II expected fall 2026, Phase I early 2027. DHS new solicitation likely this summer. DOC-NIST Phase II September, Phase I early 2027. DOC-NOAA Phase I this fall. EPA Phase I this summer.
NIH and NSF update policies on AI use and citation integrity
What: NIH published official guidance on AI use in grant applications, with a focus on fabricated references. Their position: citations pointing to nonexistent papers can constitute data fabrication under research misconduct rules. Key statement: “Applications that are either substantially developed by AI or contain sections substantially developed by AI are not considered the original ideas of applicants and will not be considered by NIH.” NSF separately updated its misconduct definition to explicitly include AI tools. A peer-reviewed paper (Resnik & Hosseini, 2026) provides the legal analysis: fabricated citations can meet the federal misconduct standard when the researcher acts with “recklessness”: indifference to a known risk.
So what: If improper AI use is detected after an award, NIH may refer the case to the Office of Research Integrity (ORI). But this is also relevant beyond AI: citation errors affect proposal quality regardless of their source. A 2025 meta-analysis found roughly 17% of citations in peer-reviewed medical publications contain inaccuracies, about half of them major. The error rate in proposals, where no editor verifies references, is likely higher.
Do: Build citation verification into your proposal workflow as a distinct step - the practical approach is below.
✅ Reference verification: a quality gate for your proposal’s citations
New to SBIR? Every abbreviation in this checklist is explained in plain English in our SBIR Glossary.
The situation
A biotech PI is preparing an NIH R43 application. Sixty-plus references across Specific Aims, Research Strategy, and the background section. Assembled from a reference manager, a colleague’s shared library, and a recent literature search. The references were added incrementally as the proposal developed. Some were carried forward from a previous submission. Some were added late in the process to support a reviewer’s likely question.
This is how most proposal reference lists come together: incrementally, from multiple sources, formatted toward the end of the writing period.
Why this matters
Why treat citation verification as a formal quality gate rather than part of the final proofread?
Reviewer credibility. A reviewer who recognizes an incorrect publication year, a misspelled author name, or a citation that doesn’t support the claim you’re making will question the rigor of the rest of your proposal. This has always been true.
Compliance risk. NIH and NSF have published updated policies on fabricated references. Under the current framework, a researcher who submits a proposal containing ghost references and who didn’t verify them could meet the standard for “recklessness”: acting with indifference to a known risk of fabrication. This applies whether the fabrication was intentional or the result of using tools (including AI) without checking the output.
Error baseline. If 17% of citations in peer-reviewed medical papers contain errors (Baethge & Jergas, 2025) - and those are papers that went through editorial review, the rate in proposals, where no editor checks references, is likely higher. Most of these errors are preventable with a systematic pass.
Three categories of citation errors
A useful framework from Resnik & Hosseini (2026) distinguishes three categories:
1. Ghost references - the citation doesn’t exist
A fabricated author, title, journal, or DOI. This is the highest-risk category under the updated NIH and NSF standards, as it can constitute data fabrication. AI tools are known to produce these - plausible-looking citations that combine real author names with invented titles - but they also arise from garbled notes, misremembered papers, or copying citations from secondary sources without verification.
2. Wrong metadata - real paper, incorrect details
The paper exists, but the publication year is wrong, an author name is misspelled, the journal title is incorrect, or the DOI doesn’t resolve. This won’t trigger a misconduct investigation, but it signals a lack of attention to detail. Reviewers who know the field will notice.
3. Misattributed claims - real paper, doesn’t support the assertion
The paper exists and the metadata is correct, but it doesn’t actually say what you’re citing it for. This is the most common category, with or without AI involvement. It typically happens when you cite based on an abstract without reading the relevant section, or when you cite a paper for a claim it mentions in passing rather than as a finding.
Verification workflow step-by-step
Step 1. Export your full reference list.
Pull every citation into a single document. Next to each one, note the specific claim it supports in your proposal. This makes Step 4 (claim alignment) faster. Don’t overlook references in figure captions, table footnotes, and budget justifications.
Step 2. Verify existence.
For each reference, confirm the paper exists. Paste the DOI into doi.org, or search the title in PubMed or Google Scholar. If a DOI doesn’t resolve and the paper doesn’t appear in any database, don’t include it until you can confirm it’s real.
This is the minimum verification step under the updated standards. Ghost references are the primary compliance risk.
Step 3. Verify metadata.
For each confirmed reference: are the author names spelled correctly? Is the publication year right? Is the journal name accurate? Do the volume and page numbers match?
Papers published online ahead of print may show different years across databases. Use the DOI as the canonical identifier when there’s a discrepancy.
Step 4. Verify claim alignment.
For each reference, return to the actual source and re-read the relevant section. Does it support the specific claim you’re making?
Common issues to watch for: citing a review article when you should cite the primary source. Citing a paper based on its abstract when the full text presents a more nuanced conclusion. Citing a paper for a claim it mentions in passing, not as a finding.
Step 5. Check agency disclosure requirements for AI tools.
If AI tools were used for any part of your proposal preparation (formatting, literature search, bibliography management, text polishing), review your agency’s disclosure policy.
NIH: Disclose AI tool use in your methods. Use of AI for routine editing and polishing is acceptable when aligned with institutional policies. Proposals substantially developed by AI will not be considered.
NSF: The research misconduct definition now explicitly includes AI tools. Follow institutional and program-specific disclosure policies.
DoD: No blanket AI policy currently, but individual components may include requirements in their submission instructions. Check your component’s guidance.
If a question arises later, documented disclosure shows good practice. Undisclosed AI use that surfaces after the fact looks like concealment.
Step 6. Archive verification evidence.
Save a PDF or screenshot of each verified source, with the date accessed. If a citation is later questioned, this provides documentation that you verified it at the time of submission. This is standard practice in clinical research and worth adopting for proposals. Watch for preprints that may be retracted or papers that receive corrections after you cite them.
The checklist
Per-reference verification (run on every citation):
☐ Exists - DOI resolves, or found in PubMed / Google Scholar
☐ Metadata correct - authors, year, journal, volume/pages match
☐ Claim aligned - source supports the specific assertion I’m making
☐ Evidence archived - PDF or screenshot saved with access date
AI disclosure (run once per proposal):
Did AI tools touch any part of this proposal’s preparation?
☐ No AI used → no disclosure needed, still verify all references
☐ AI used for formatting / polishing / bibliography → disclose per agency policy, verify all references
☐ AI used for literature search or text drafting → disclose per agency policy, verify all references with particular attention to ghost references
☐ Unsure whether a tool qualifies → check your agency’s definition, err toward disclosure
Agency-specific:
☐ NIH: AI use described in methods section
☐ NSF: Institutional and program-specific policies followed
☐ DoD: Component submission instructions checked for AI requirements
Common questions
“I only use AI for grammar and formatting. I write my own references from papers I’ve read. Is this relevant to me?”
Yes. Wrong metadata and misattributed claims occur without any AI involvement. A paper you cited two years ago may not say exactly what you remember. The verification workflow catches errors regardless of their source.
“Has NIH actually rejected or rescinded a grant over AI-generated references?”
No public SBIR case as of this writing. The summer 2025 cases were flagged by review administrators. The ORI referral process takes time. But the policy is published, the legal framework is established (Resnik & Hosseini, 2026), and agencies are developing detection methods. Building verification into your workflow now is considerably easier than responding to an inquiry later.
“I apply to DoD, not NIH. Does this matter for me?”
DoD hasn’t published a blanket AI policy, but individual components may include requirements in their submission instructions. Systematic reference verification improves your proposal quality at every agency. A reviewer who checks a citation and finds it doesn’t support your claim will downgrade your score, AI policy or not.
“I have 70 references. This sounds like a full day of work.”
Less than you might expect. DOI resolution checks take seconds per reference. PubMed title searches are fast. The time-intensive step is claim alignment (Step 4), but there you’re scanning the relevant section of the paper, not re-reading the entire publication. For a typical 60-80 reference proposal, a systematic pass takes roughly 2-4 hours. Spreading this across the writing period (verifying each citation as you add it) is more efficient than a batch check at the end.
“Is this just agencies trying to discourage AI use?”
The NIH submission cap (six per PI per year) was partly a response to the volume of AI-assisted applications. But citation integrity is a distinct issue. Inaccurate citations waste reviewer time and undermine the credibility of your technical arguments, however the errors originated.
One thing to do this week
Pick a proposal you’ve submitted in the past year. Pull up the reference list and choose five citations at random. Run them through Steps 1-4: DOI check, metadata check, claim alignment check. If all five are clean, your current process is working. If you find errors, you’ll know what to add to your workflow before the next submission.
If this verification workflow is useful for your team, forward it to a colleague who's preparing a submission.
Till next time,
—Lana.