> ## Content Index > Fetch the complete content index at: https://blog.ascendr.ai/llms.txt > Use this file to discover other available public pages before exploring further. # The Verification Gap - URL: https://blog.ascendr.ai/the-verification-gap/ - Published: 2026-08-31T09:00:00.000Z - Updated: 2026-09-03T12:50:41.000Z - Description: AI is making research easier and faster. Verifying it has not kept pace. Every polished draft lands on the same desk, waiting for a signature your institution stands behind. That distance is the verification gap. The piece closes with working principles and a 90 day plan for the office that signs. - Author: Ascendr, PBC - Tags: blog *AI is making research output cheap. Verification has not gotten faster. A practical framework and a 90-day plan for the office that signs.* It is 3:30 on a Friday afternoon when the email arrives. Attached is a complete R01 draft: specific aims, biosketches, and a budget justification. Everything is polished and well-cited. The note says AI helped prepare it, but the PI reviewed it. Can we submit? It is due Monday, and your office must sign before it goes out. The question that matters is not whether AI wrote it. The question is whether your office can stand behind it. The bottleneck in research is shifting from producing the work to trusting it. That question anchored our founder’s talk at the NC RAA Annual Conference in Chapel Hill this July. What follows is a practical version, adapted for the work on your desk. --- ## **A quick test:** Which of these situations reached your office this year? - An AI-drafted proposal or specific aims page arrived for review. - A faculty member asked what AI use must be disclosed. - A citation did not check out. - Someone raised concerns about confidential data being entered into an open AI tool. - A capability or team claim could not be verified. If even one sounds familiar, this piece is for you. ## **The pressure is measurable:** In Wiley’s ExplanAItions surveys, the share of researchers reporting that they used AI in their work rose from 57% to 84% in a single year. Because these figures are self-reported, the trend is more reliable than the exact percentages. The increase in volume is harder to dispute: grant applications to major funders were 57% higher in 2025 than in 2022, while reviewer capacity did not increase. Funders felt the pressure first. In the summer of 2025, one PI submitted more than 40 applications in a single round, many with AI assistance. NIH responded by limiting each PI to six applications per year. The same notice, NOT-OD-25-132, states that applications substantially developed by AI are not considered the applicant’s own original ideas. NSF had already restricted the use of generative AI in merit review and asked proposers to disclose its use. The rules are already changing. ## **Naming the problem:** We call this the **verification gap**: the widening distance between how quickly research output is produced and how quickly an institution can confirm that it is accurate, original, compliant, and real. When text becomes cheap, trust becomes the work. ## **Where the gap appears:** That work lands on the desks of research administrators. The verification gap appears in proposal narratives, citations and references, biosketches, budget justifications, capability and team claims, subaward partner vetting, human-subjects materials, and AI-use disclosures. Each is a point where polished output meets your institution’s signature. ## **From overhead to assurance:** The risk is clear: something untrue could leave your office with your institution’s name on it. It might be a fabricated citation, an overstated capability, or an undisclosed use of AI. But the same pressure also creates an opportunity. The offices that close this gap stop being overhead and become assurance, and the automation fear flips with it. Research administration does not get automated away; it gets promoted. ## **Four working principles:** Closing the verification gap does not require a new set of AI skills. It requires strong administrative practices built around four principles. ### **1\. Evidence over assertion** Trust what has been demonstrated, not merely what has been claimed. Funded grants, published papers, and issued patents retain their meaning even when polished prose becomes cheap. Unsupported claims do not. Before asking how convincing something sounds, ask what the person, team, or tool has actually produced. ### **2\. Keep a trust boundary** **Verified** means traceable to a source, checkable by a skeptic, and safe to carry your institution’s name. **Estimated** means model-generated, plausible, and unconfirmed. It should be used only to support decision-making. Automate below the trust boundary and make decisions above it. Never let fluent language blur the distinction. ### **3\. Keep human judgment at the center** AI proposes. People decide and remain accountable. A tool can inform a decision, but it cannot carry responsibility for that decision. Every consequential output should have a named human owner before it leaves the office. ### **4\. Label honestly** State clearly what is verified, estimated, or unverified. In a flood of confident, unlabeled output, candor is rare. The trust it earns compounds over time. ## **Scale is the challenge:** The obvious objection is scale. When sources cooperate, a person may still need about fifteen minutes to verify a single claim. One proposal can contain more than 100 citations before capability claims, budgets, biosketches, and other supporting materials are considered. Multiply that work across everything your office signs in a year, while submission volume rises and headcount remains flat, and even careful manual review cannot keep up. Fight scale with scale. Let machines handle the mechanical checks: confirming that citations resolve and support the claims made, numbers reconcile, and AI use is disclosed according to the funder’s requirements. Machines can also retrieve relevant public evidence, including awards, papers, and patents. Human should retain responsibility for judgment: what a flagged mismatch means, whether the evidence supports the claim, whether the institution is prepared to sign, and who is accountable for the final decision. The technology that helped create the flood can also help perform the checks. Automate below the trust boundary; decide above it. ## **Evaluate every tool:** Before trusting any tool, ask four questions: 1. **Provenance:** Where did this answer come from, and can we see the source? 2. **Control:** Who will have access to our data? 3. **Auditability:** Can someone check the work afterward? 4. **Accountability:** Is a named human responsible for the decision? These questions should be asked for every tool and every vendor. ## **Six moves, one loop:** The plan consists of six connected moves: 1. **See the field:** Use public evidence to understand funding opportunities and the broader research landscape. 2. **Map capability:** Identify expertise through demonstrated grants, papers, and patents. 3. **Govern AI use:** Create a one-page policy defining where AI may be used and what must be disclosed. 4. **Verify before it ships:** Apply a release checklist and clear evidence labels. 5. **Invest in people:** Shift staff time toward judgment, relationships, and verification. 6. **Measure what matters:** Track quality, originality, equity, verified output, and institutional learning, not only volume and win rates. Run the loop on one live opportunity, then run it again. This work can begin with free public sources, including NIH RePORTER, NSF Award Search, USAspending, Grants.gov, OpenAlex, ORCID, Semantic Scholar, USPTO Patent Public Search, Google Patents, and Retraction Watch. The cost is discipline, not money. ## **Start with the release check:** Begin with **verify before it ships**: the final check before your institution’s name goes on the work. 1. Every citation opens and supports what the document claims. 2. Capability claims match demonstrated evidence, such as awards, papers, and patents. 3. AI use is disclosed according to the funder’s requirements. 4. A named person signs off before the work leaves the office. Then label the output as **verified**, **estimated**, or **unverified**. ## **The first 90 days:** ### **Weeks 1–4: See clearly** Open the free public sources, assign an owner to each, and draft a one-page AI-use policy. ### **Weeks 5–8: Pilot the process** Create an evidence map for one live funding opportunity and run the release checklist on one real proposal. ### **Weeks 9–12: Make it stick** Select three metrics that go beyond volume and add verification labels to everything you circulate. On Day 91, give leadership a one-page update: what changed, what the process caught, and what comes next. **Start here**: pick one live opportunity on Monday and run the loop on it. To access the full deck, source appendix, and field guide click[ here](https://ascendr.ai/NC-RAA-2026-The-Verification-Gap-George-Koomullil.pdf?ref=blog.ascendr.ai). --- *This is Part 1 of our three-part series on Verification Gap. Read the complete series at* [https://ascendr.ai/library/](https://ascendr.ai/library/?ref=blog.ascendr.ai)*.* --- ### Written by the Ascendr Team Ascendr, PBC is a public benefit corporation that believes every high-stakes decision deserves better than opinion. We wrote this series the way we build our products, with every claim traceable and every limit labeled. Through our writing, we share practical ideas that help research, IP, and technology teams make clearer, more defensible decisions. ###