Last week, I described the paragraph that writes back. The one that resists you. The one where every sentence you draft feels almost right and then collapses under its own weight.
Some of you read that and recognized the feeling immediately.
It is the most natural thing in the world. The paragraph resists. The cursor blinks. And the tool is right there — a second tab, a single paste, a prompt that says: Make this clearer.
I know because I teach grantsmanship to researchers at every career stage, and this is the conversation that is happening in every cohort I run right now. Everyone is using AI. Almost no one is talking about it honestly. And the gap between what these tools can do for your grant and what they can do to your grant is a line most researchers have never been asked to find.
This is the newsletter where we find it.
The Mirror
Here is what AI does well in grant writing, and it is not what most people think.
The best use of an AI tool in your revision process is as a mirror. You paste a paragraph. The tool reflects it back. And in that reflection, you see what your own eyes have learned to skip — the hedging language, the circular logic, the moment where your reasoning jumps from point A to point D without visiting B or C.
Last week I called this scaffolding language — words like “importantly,” “it is well established,” “this innovative approach” — that hold the place where your actual thinking has not arrived yet. An AI tool can find that language faster than you can. It can flag the sentence that says nothing. It can identify the transition that covers a logical gap with confidence rather than evidence.
This is what I call the Mirror Model in my curriculum. The tool reflects your thinking back to you. It surfaces what is vague. It identifies where your reasoning is incomplete.
What it cannot do is complete the reasoning for you.
That resistant paragraph from last week — the one that would not come together — was resistant because the thinking underneath it was unfinished. An AI tool can tell you the paragraph is unclear. It cannot tell you why your conceptual model does not yet support the mechanism you are proposing. It cannot tell you that your preliminary data actually points in a different direction than your hypothesis. It cannot tell you that the approach you described in Aim 2 does not follow logically from the question you posed in Aim 1.
Those are scientific problems. They require your expertise, your years in this field, your intimate knowledge of the data. And when you let an AI tool resolve them for you — when you accept its smoothly written alternative — you have not fixed the paragraph. You have buried the problem under better prose.
The reviewer will find it anyway.
The Line
NIH drew the line in NOT-OD-25-132, released in the summer of 2025 under the title “Supporting Fairness and Originality in NIH Research Applications.”
The policy states it plainly: applications substantially developed by AI will not be treated as reflecting the applicant’s original ideas. Such applications may be administratively withdrawn.
There is no percentage threshold. No per-section rubric. No word count that separates “assisted” from “generated.” The standard is whether the scientific thinking in your application is yours.
Here is where this gets practical.
Using AI to check grammar, tighten a sentence, or improve the clarity of a paragraph you already wrote — that is editing. NIH permits it. You wrote the science. The tool polished the language.
Using AI to generate your specific aims, produce your significance section, draft your research strategy, or construct your conceptual framework — that is drafting. The ideas came from the machine (or as my son says, "the robot"). NIH considers that a violation of the originality standard, regardless of how much you edited the output afterward.
The distinction is not about the tool. It is about the direction of the thinking.
When you write first and then ask AI to sharpen what you wrote, the intellectual trajectory moves from your mind to the page, with AI as a refinement step. When you prompt AI first and then edit what it produced, the trajectory moves from the machine to the page, with you as a filter. The science looks the same in both cases. The origin of the ideas does not.
And in May 2026, NIH made the consequences explicit. The Extramural Nexus reminder confirmed that existing research misconduct definitions — fabrication, falsification, plagiarism — apply to AI-generated content in grant applications. An AI-fabricated citation is a fabricated citation. An AI-generated passage that closely paraphrases someone else’s published work without attribution is plagiarism. And the standard for misconduct is “intentionally, knowingly, or recklessly” — which means that failing to verify what an AI tool produced is itself a serious lapse.
The phrase that should stay with every researcher reading this: “The AI got it wrong” is not, on its own, a defense.
One more thing you need to know. Three weeks ago, NIH issued NOT-OD-27-011, implementing Executive Order 14434. The federal government has officially replaced the term “Artificial Intelligence” with “Super Intelligence” across all agency communications and policy documents. You will start seeing “SI” where you used to see “AI” in NOFOs, guidance, and reporting requirements.
The underlying definition has not changed. The policies governing how you use these tools in your grant applications have not changed. NOT-OD-25-132 still draws the line between editing your own work and generating someone else’s. The misconduct framework still applies.
The rebrand matters for one reason your research office may not tell you: when terminology shifts, researchers stop paying attention. They assume the new language means new rules, and they wait for someone to explain it. Meanwhile, the old rules — the ones that actually govern what you can and cannot do — keep running. Do not wait for a new guidance document. The guidance already exists. It is called NOT-OD-25-132, and everything I described above still applies, regardless of what anyone calls the tool.
The Voice Problem
There is a third dimension to this that the policy does not address and that almost no one in the grantsmanship space is talking about.
Your grant has a voice.
Experienced reviewers know this intuitively even if they have never articulated it. They read hundreds of applications. They develop an ear for the difference between a researcher who is writing from deep operational knowledge and a researcher who is describing a study they have read about. The confidence shows up in specific places — the way you describe your recruitment strategy, the precision of your analytic plan, the small methodological choices that reveal someone who has actually run this kind of study before.
AI-generated prose has a texture. It is fluent, organized, and comprehensive. It covers all the expected points. It uses transitions that connect paragraphs in ways that feel logical. And it sounds like no one in particular.
That is the problem.
When a reviewer reads a research strategy that is technically correct and stylistically generic, something registers. The operational specificity is missing. The choices that reveal a real investigator behind the science are absent. The writing is smooth in a way that actual expertise rarely is — because real expertise is specific, sometimes awkward, and always grounded in the particular details of a particular study run by a particular team.
I am not suggesting reviewers are running detection software. I am telling you that an experienced reviewer can sense the difference between a section written by someone who knows this work in their bones and a section that reads like a well-organized summary of the field. And when that instinct fires, what erodes is not a score on a rubric. It is the reviewer’s confidence that you can actually do what you are proposing.
That confidence lives in Factor 2. It lives in Factor 3. And once it is gone, no amount of methodological rigor in your approach section can rebuild it.
What This Means for Your Draft Right Now
If you are in the middle of a revision — and many of you are, with the February cycle approaching — here is how to use AI without crossing the line and without losing your voice:
Use it as a mirror. Paste the paragraph that is giving you trouble. Ask the tool to identify where the logic is unclear, where the transitions are weak, where the claims are unsupported. Read what it reflects back. Then close the tool and do the thinking yourself.
Do not accept AI-generated replacement text. If the tool rewrites your paragraph, that rewrite carries the tool’s reasoning, not yours. It may be clearer. It will not be more scientifically sound. And it will not sound like you.
Keep records. Document which sections involved AI assistance and at what level. There is no formal disclosure field in the NIH application yet, but your institution’s research office may already be developing internal documentation requirements. The researchers who will be best positioned when formal disclosure arrives are the ones who started keeping records before they were required to.
And read your draft out loud. This is the oldest writing advice in the world, and it is the best AI detection tool that exists. If a paragraph sounds like something you would say to a colleague over coffee, it is yours. If it sounds like something a committee wrote, look again.
The tool is here. It is powerful. And it is not going away.
The question is not whether you will use it. The question is whether you will use it in a way that strengthens your scientific voice or replaces it.
I built two new modules in my Lost In Translation curriculum on exactly this — how to use AI as a revision partner without crossing the integrity line, and what NIH policy actually requires. Because this is not a conversation that can wait until someone gets caught.
It is a conversation that belongs in your writing process right now.
What is your line? Where do you draw it — and has anyone ever asked you to articulate it before?
Lisa Carter-Bawa, PhD, MPH, APRN, ANP-C, FAAN, FSBM, is a behavioral scientist who bridges the gap between what research knows and what communities experience. She is the creator of the Lost In Translation (LIT) Grantsmanship Curriculum, a behavioral science-based approach to NIH grant writing for researchers at every career stage. The LIT course is available as a self-paced online course — Module 1 is free at lisacarterbawa.com/module1. I also run small cohorts timed to NIH submission deadlines. The next cohort (Spring 2027 cohort tied to a June 2027 submission) is accepting applications now — visit https://www.lisacarterbawa.com/cohort for details and if you would like to apply, visit this link.
