Chapter 24: Grant Writing with AI — Using LLMs as Tools Without Crossing the Line

A few years ago this chapter would not have existed. Today it is unavoidable: powerful AI language models can draft prose, summarize literature, brainstorm ideas, and edit for clarity in seconds, and grant writers everywhere are using them — some to great benefit, some to their peril. Used well, AI is a genuine productivity tool that can make you a faster, clearer writer and a sharper thinker. Used badly, it can introduce fabricated citations, generic prose, confidentiality breaches, and integrity violations that sink a proposal or damage a career. This chapter is about the line between the two, and how to stay firmly on the right side of it — getting the genuine speed and clarity AI offers while never surrendering the truth, specificity, and accountability that make a proposal fundable.

The technology is changing fast, and any specific claim about what a given model can do, or what a given funder permits, may be outdated by the time you read it. So this chapter focuses on the durable principles — the ones that will hold regardless of how capable the models become or how the policies evolve. And those principles converge on a single threshold concept: AI can sharpen your argument but cannot own it — accountability stays with you. An AI model is a tool for thinking, structuring, and editing, but it cannot supply the original ideas, the real data, the funder judgment, or the accountability that a fundable proposal requires. No matter how fluent the output, you remain the author — responsible for every claim, every number, and every word. Internalize that, and AI becomes a powerful assistant; forget it, and AI becomes a liability wearing the mask of help.

In this chapter we'll explain what AI language models actually are and why they fail in characteristic ways, map where they genuinely help in grant writing and where they're dangerous, confront the confidentiality and peer-review red lines, navigate evolving funder policies and the disclosure question, and arrive at a best practice: AI as a tool that augments a human author who stays fully accountable. We'll thread our anchors lightly — Hernandez, RYCC's Denise, Lighthouse, and Sam each using (or misusing) AI — because the principles apply across every funder and sector you've met. A note on this book's own stance: the disciplines here — verify everything, protect confidentiality, keep a human accountable — are not anti-AI; they are how to get AI's real benefits without its real harms. The goal is neither to forbid a useful tool nor to wave it through uncritically, but to use it the way a serious professional uses any powerful instrument: knowing exactly what it does well, what it does badly, and where the lines are.

24.1 What AI Language Models Actually Are (and Why They Fail)

To use AI well, you need a working mental model of what it is — not the engineering, but enough to predict where it helps and where it betrays you.

A large language model (LLM) is a system trained on vast amounts of text to predict plausible continuations of language. Given some input, it generates output that is statistically likely to follow — which, remarkably, produces fluent, often useful, frequently impressive prose. But notice what that description does not include: the model is predicting plausible language, not retrieving verified truth. It has no built-in mechanism for knowing whether what it generates is accurate; it generates what sounds right. This single fact explains most of how AI helps and how it fails.

The most important failure mode follows directly: hallucination — the model's tendency to generate confident, plausible-sounding content that is simply false. It will invent citations to papers that don't exist, complete with realistic authors and journals. It will produce statistics that sound authoritative but are fabricated. It will describe study findings that were never found. And it does all this in the same fluent, confident voice it uses for accurate content, so you cannot tell from the writing which parts are real. Hallucination is not an occasional bug being engineered away; it is intrinsic to how these systems work — they generate plausible language, and plausible-but-false is, by construction, within their range. Newer models hallucinate less and tools that connect models to real sources help, but the fundamental lesson is permanent: an AI model's fluent confidence is no evidence of accuracy.

It helps to hold a second limitation alongside hallucination: an LLM's knowledge is bounded and undated from your perspective. It was trained on text up to some point and may not know recent developments — a funder's policy change, a new rule, this cycle's priorities — yet it will answer about them as confidently as anything else, sometimes with outdated or invented specifics. It also has no access to your private world: your real data, your institution's actual indirect rate, your program officer's verbal steer, your pilot's results. So beyond "it can be confidently wrong about facts," remember "it cannot know your specifics or the latest specifics" — which is precisely why the substance of a proposal (your evidence, your funder intelligence, your numbers) has to come from you and your research, never from the model's general training. The model is a language tool, not an oracle and not a database of your project.

🚪 Threshold Concept: AI can sharpen your argument but cannot own it — accountability stays with you. This is the organizing idea of the chapter, and it follows from what an LLM is. The model can help you think, structure, phrase, and polish — genuinely valuable work. But it cannot be accountable for the result, because it has no stake, no judgment about your specific funder, no access to your real data, and no way to guarantee its own accuracy. When a fabricated citation or a wrong number reaches a reviewer, "the AI wrote it" is not a defense; it is a confession that you abdicated your authorship. Cross this threshold and you treat every AI output as a draft from an unreliable assistant that you, the accountable author, must verify, correct, and make your own. The author who owns the proposal can use AI freely and safely; the one who lets AI own the proposal has already lost control of the one thing that was always theirs — responsibility for the truth of what they submit.

🧩 Productive Struggle: Before reading on, consider a paradox: AI is most dangerous precisely where it is most fluent. Why would better-sounding output be riskier? Jot your thinking. The resolution, developed through this chapter, is that fluency is exactly what disarms your skepticism — a clumsy error you'd catch, but a smoothly written fabricated citation or a confidently stated false statistic slides past your guard because it reads like competent work. The danger isn't that AI writes badly; it's that AI writes plausibly whether or not it's telling the truth, so the better it sounds, the more vigilant — not less — you must be. This inverts the normal relationship between polish and trust, and learning to distrust fluent AI output is a core skill of this chapter.

24.2 Where AI Genuinely Helps

With that foundation, here is the good news: there are real, substantial ways AI helps grant writers, and they cluster around tasks where plausible language is genuinely useful and where you can easily verify the output.

Brainstorming and overcoming the blank page. AI is an excellent thinking partner for generating options — possible framings of a problem, angles on significance, names for a program, ways to structure an aim. You're not asking it for truth; you're asking it for possibilities you'll judge. When Denise at RYCC is stuck on how to frame her program's significance, asking AI for ten different framings can break the logjam — and she picks and refines the ones that fit, discarding the rest.

Outlining and structuring. AI is good at proposing structures — an outline for a narrative, a logical order for sections, a way to organize a complex argument. You bring the content and judgment; it helps with the scaffolding.

Clarity and line editing. This is among AI's most reliable uses: improving the clarity, concision, and flow of prose you wrote. "Make this paragraph clearer," "tighten this," "vary these sentence openings," "this reads as jargon-heavy — simplify it." Because you're editing your own real content and can judge whether the edit preserved your meaning, the hallucination risk is low and the benefit is real. Sam, drafting a fellowship application, can use AI to sharpen clumsy paragraphs — checking each edit kept the science accurate.

Summarizing and digesting (with verification). AI can summarize long documents — a funder's guidelines, a body of literature you provide, your own draft — to help you digest them. But here verification matters: a summary of a document you supply is far safer than asking AI to summarize literature from its training, which invites hallucination. Summarize what you give it; verify what it tells you.

Devil's-advocate review. A genuinely valuable use: asking AI to critique your draft as a skeptical reviewer would. "What weaknesses would a reviewer find here?" "What's unclear?" "What questions does this raise that I haven't answered?" AI is a tireless, low-stakes first reviewer that can surface problems before a real reviewer does (Chapter 9's pitfalls thinking) — and because you judge whether each critique is valid, the risk is low.

Jargon and accessibility checks. AI can flag where your writing assumes too much, where a term needs defining, where a lay reviewer might get lost — useful given Chapter 2's tired, possibly non-specialist reviewer.

Watch the safe pattern in action with Denise at RYCC (composite). Facing a blank page on her significance section, she asks AI to suggest ten ways to frame the importance of after-school coding for under-resourced middle-schoolers. The list comes back: some framings are generic, a couple are clichéd, but two spark something — one connecting digital-skills access to long-term economic mobility, another framing it as meeting families' own expressed demand. Denise chooses those two, discards the rest, and rewrites them in her own voice, grounded in RYCC's specific neighborhood and the real families who asked for the program. Then she asks AI to play skeptical reviewer on her drafted needs section; it flags that she asserted demand without evidence — a fair critique she addresses by adding her waitlist numbers. At no point did AI supply a fact, a citation, or a number; it supplied options and critique that Denise judged, and the substance stayed entirely hers. That is AI used exactly as it should be: a thinking partner whose every output she owns or discards. Notice, too, what she did not do — she did not ask it for statistics about the digital divide (she'd have had to verify those anyway and risked a fabrication), and she did not let its phrasings stand as her voice.

💡 Key Insight: Notice the pattern in every safe, valuable use: AI helps most where its output is easy for you to verify and where you supply the substance. Brainstorming (you judge the options), editing your own prose (you check meaning survived), summarizing what you provide (you have the source), devil's-advocate critique (you assess each point) — in each, you remain the judge of truth and quality. The dangerous uses, as the next section shows, are exactly the inverse: where AI supplies substance you can't easily check — facts, citations, data, funder-specific judgment. So the practical rule that organizes everything is: use AI for tasks where you can verify the output; never trust it for substance you'd have to take on faith. That line — verifiability — separates the productive uses from the perilous ones.

📊 From the Field: A useful way experienced writers frame it: AI is a brilliant, fast, tireless intern with a serious flaw — it sometimes makes things up with total confidence and never tells you when. You'd happily hand an intern a draft to tighten, a document to summarize, or a request for ten brainstormed options, and you'd find that genuinely helpful. But you would never submit an intern's work without reading it, never trust an intern's cited statistic without checking it, and never let an intern make the final call on your funder strategy or sign your name. Treat AI exactly so: delegate the tasks you'd delegate to a capable intern, apply the same verification you'd apply to an intern's work, and keep for yourself the judgment, the accountability, and the authorship that were always yours. The intern framing gets the relationship right — real help, real limits, and you firmly in charge. And like a good supervisor, you stay responsible for the intern's output: if it goes out under your name, its errors are yours to have caught.

24.3 Where AI Fails — and Where It's Dangerous

Now the hazards, which are serious and which have already damaged real applications and careers. These cluster where AI supplies substance you can't easily verify, or where using it at all crosses a line.

Fabricated citations and statistics — the cardinal danger. Ask AI for sources to support a claim, or for statistics about a problem, and it will often produce them — confidently, formatted perfectly, and sometimes entirely invented. A proposal containing a citation to a paper that doesn't exist, or a fabricated statistic, is a catastrophe: it destroys your credibility instantly if a reviewer catches it (and reviewers, who know the literature, often do), and it violates the citation honesty that Chapter 8 made central. Never include an AI-provided citation or statistic without independently verifying it against the real source. This is the single most important rule in the chapter, and the most violated — and it costs only seconds to honor.

Fake preliminary data or results. AI can generate plausible-sounding "results," data, or findings. Including any AI-generated data, or letting AI fabricate or embellish your actual results, is research misconduct — fabrication, full stop. Your preliminary data must be real (Chapter 9). AI has no access to your lab, your pilot, or your program's outcomes, and anything it "produces" in that vein is fiction.

Inaccurate budgets and technical specifics. AI doesn't know your actual costs, your institution's indirect rate, your funder's budget rules, or your real timeline. A budget is a precise, verifiable, accountable document (Chapters 11–12); AI guesses, and its guesses will be wrong in ways that matter. Build your budget from real numbers, not AI estimates.

Generic, voiceless prose. Even when accurate, AI prose tends toward the generic — fluent but bland, full of the kind of hedged, characterless writing that makes a tired reviewer's eyes glaze (Chapter 2). A proposal that reads as AI-generated boilerplate fails to convey the specific passion, voice, and concrete detail that distinguish fundable proposals. The remedy is the one-voice discipline of Chapter 23: if you use AI in drafting, you must rewrite it into your own specific, vivid voice, grounded in your real project's details.

Missing funder nuance. AI doesn't understand the specific room you're writing for (Chapter 2) — this institute's priorities, this foundation's relationships, this program officer's steer, this rubric's weighting (Part III). The funder-specific judgment that wins grants is exactly what AI lacks, because it's not in any training data; it's in your research, your relationships, your program-officer conversations, and your reading of the specific opportunity. A proposal written to a generic notion of "what funders want" rather than to this funder's actual priorities reads as exactly that — unanchored, one-size-fits-all, and unconvincing to the specific reviewer who knows their program.

Confidentiality and privacy breaches. Pasting sensitive material into a public AI tool can expose it — unpublished ideas, others' confidential information, personal data, or proprietary content (Section 24.5). This is a distinct and serious hazard independent of accuracy.

🔄 Check Your Understanding: An applicant asks an AI tool, "What percentage of people returning from incarceration are unemployed one year after release?" and it returns a specific, confident figure with a citation. Lighthouse wants to use it to anchor the needs section. What should Lighthouse do, and why?

Answer

Lighthouse should not use the figure or citation as given — it must independently verify both before either enters the proposal. The AI's confident, specific number and formatted citation are exactly the pattern of a possible hallucination (Section 24.3): the statistic may be fabricated or distorted, and the citation may point to a paper that doesn't exist or doesn't contain that figure. Lighthouse should treat the AI output as a lead: find the real, authoritative source (a government dataset or peer-reviewed study), confirm the actual figure, and cite that — or, if no real source supports it, not make the claim. Using an unverified AI statistic risks a credibility-destroying error and violates citation honesty (Chapter 8). The rule: AI can point you toward facts to check; it cannot supply facts you take on faith.

⚠️ Common Pitfall: Trusting an AI-provided citation because it looks real. This deserves repeating because it is the most common and most damaging AI failure in grant writing. AI fabricates citations that are formatted flawlessly — real-sounding author names, plausible journal titles, believable years, even realistic-looking identifiers — and they sit in your reference list indistinguishable from genuine ones. The reviewer who tries to look one up, fails to find it, and realizes it's invented does not think "honest mistake"; they think "this applicant fabricated a citation," and your entire proposal's credibility collapses, sometimes with integrity consequences beyond the single application. The fix is absolute and non-negotiable: every citation in your proposal must be one you have personally verified exists and says what you claim it says — found in a real database, read (at least the abstract), and confirmed. If AI suggested a source, treat that as a lead to verify, never as a citation to use. The thirty seconds of verification is the cheapest insurance you will ever buy.

🗣️ From the Review Panel: (A reviewer reflects.) I've started seeing it: proposals with a citation I can't find anywhere, or a statistic with no traceable source, or prose that has that smooth, generic, slightly-empty quality of unedited AI output. When I hit an invented citation, I don't just dock a point — I lose trust in the whole application, because if this reference is fabricated, what else is? And the generic-AI-voice proposals, even when nothing's wrong, simply don't move me; they read like anyone could have written them about any project, which is the opposite of what wins. Here's what I'd tell applicants: I don't care whether you used AI to help you write — I care whether what you submitted is true, specific, and yours. Use the tool if it helps, but verify every fact, ground every claim in your real work, and make the writing sound like a human who actually cares about this project. The ones who do that are fine. The ones who submit AI's unverified, voiceless first draft are the ones I catch.

24.4 The Human-in-the-Loop Discipline

The principle that makes AI safe has a name: human-in-the-loop — a human remains actively engaged, judging and verifying, at every step, never letting the AI's output flow unchecked into the proposal. This is the operational form of the threshold concept, and it has a few concrete commitments.

You verify every fact, citation, number, and claim. Nothing AI produces about the external world enters your proposal without independent verification against a real source. This is non-negotiable and is the single most important practice.

You supply the substance and the judgment. The ideas, the real data, the funder strategy, the specific details of your project — these come from you. AI helps you express and organize them; it does not originate them.

You own the voice. Whatever AI drafts, you rewrite into your own specific, vivid, accountable voice (Chapter 23's one-voice owner, applied to an AI "contributor"). The submitted proposal must sound like you and reflect your real project, not generic AI prose.

You remain accountable. When you submit, you are certifying that the work is yours and true. "The AI did it" is never a defense for a fabrication, an error, or a confidentiality breach. The accountability that was always yours stays yours.

To make the discipline concrete, here is a verification workflow a careful writer follows when AI touches a draft:

📋 Template — the human-in-the-loop verification pass (run before any AI-touched text is final): 1. Flag every factual claim AI contributed or edited — statistics, citations, dates, names, technical specifics, budget figures. 2. Verify each against a real source — find the actual paper in a real database and confirm it says what's claimed; confirm each statistic at its origin; rebuild any number from your real data. Delete or replace anything you can't verify. 3. Check the meaning survived — for AI edits to your prose, confirm the edit didn't subtly change your scientific or programmatic meaning (AI sometimes "improves" a sentence into inaccuracy). 4. Rewrite for voice — convert generic AI phrasing into your specific voice, grounded in your real project's concrete details (Chapter 23's one-voice owner). 5. Scan for funder fit — confirm nothing AI added contradicts the specific funder's priorities, rubric, or rules, which AI doesn't know. 6. Confirm confidentiality — verify nothing confidential was exposed in producing the text.

Only text that has passed this pass belongs in the proposal. The pass is the operational meaning of "accountability stays with you" — and it is fast once it's habit.

🔍 Why Does This Work?: Why is "human-in-the-loop" the right frame rather than, say, "use AI only for small tasks"? Because the dividing line that matters isn't the size of the task but whether a human is judging the output. AI can help with large tasks (restructuring a whole draft) safely if you scrutinize the result, and AI can harm you on a tiny task (one fabricated citation) if you don't. Keeping a human in the loop — actively verifying, judging, and owning every output — is what converts AI from a liability into a tool, regardless of task size, because it restores the one thing the AI inherently lacks: accountable judgment about truth and fit. The frame works because it targets the actual failure mode (unchecked output reaching the proposal) rather than a proxy (task size), and it scales with AI's capability: as models get more powerful, the human-in-the-loop discipline remains exactly as necessary, because more capable models still can't be accountable.

🪞 Learning Check-In: Notice your own temptation with AI. The seductive move is to let it do more than you verify — to accept the fluent draft, trust the confident citation, ship the generic prose because it's "good enough" and you're tired. That temptation is strongest exactly when you're most rushed, which is exactly when verification matters most. Ask yourself honestly: when I use AI, am I staying in the loop — reading, checking, rewriting, owning — or am I outsourcing judgment I'm still accountable for? The discipline isn't about using AI less; it's about never letting its output escape your judgment. If you can feel the pull to abdicate and resist it, you can use AI freely and safely. If you can't feel the pull, you're not paying enough attention to catch the fabrication that's coming.

24.5 Confidentiality, Privacy, and the Peer-Review Red Line

Beyond accuracy, AI raises distinct hazards around confidentiality that every grant writer — and every grant reviewer — must understand, because some uses cross bright ethical and sometimes legal lines.

Don't paste confidential or sensitive material into public AI tools. Many AI services may retain, log, or train on what you submit. Pasting in unpublished research ideas, proprietary information, personal data about identifiable people, others' confidential material, or a not-yet-public proposal can expose it — a privacy breach, a loss of intellectual-property protection, or a violation of others' confidentiality. Treat a public AI tool as a potentially non-private channel, and don't put anything into it you wouldn't want disclosed, unless you're using a tool with contractual privacy guarantees your institution has vetted.

The peer-review red line. This one is categorical and worth stating starkly: a reviewer must never upload a confidential proposal they are reviewing into a public AI tool. A proposal under review is confidential — entrusted to the reviewer by the funder, containing the applicant's unpublished ideas and data. Putting it into a public AI system breaches that confidentiality, potentially exposes the applicant's ideas, and violates the integrity of peer review. Major funders, including the NIH and NSF, have moved to prohibit reviewers from using AI tools in ways that involve uploading or exposing the applications they review. If you serve as a reviewer (a role this book encourages, Chapter 19), know and follow your funder's rules — and the default is: the confidential application does not go into any AI tool. The applicant's trust, and the integrity of the system that will one day review your proposals, depends on it.

📊 From the Field: The peer-review confidentiality line has become one of the clearest AI rules in the funding world precisely because the breach is so concrete. When a reviewer pastes an application into a public chatbot — even just to help summarize it or draft the critique — the applicant's confidential, unpublished ideas have left the secure review process and entered a system that may retain them. Imagine it's your unfunded-but-brilliant idea, exposed because a reviewer wanted to save time. That's why funders have drawn the line hard, and why it's not a gray area: reviewers handle applications within the funder's secure systems and their own judgment, not third-party AI tools. The same logic extends to anyone handling others' confidential proposals — a colleague's draft, a partner's unpublished data. Confidentiality entrusted to you is not yours to feed to an AI. This is among the few genuinely bright lines in a fast-changing area, and it is firmly drawn.

24.6 Funder Policies, Ethics, and Disclosure

The policy landscape is evolving quickly and varies by funder, so the durable skill is knowing how to navigate it rather than memorizing today's rules.

Funder policies vary and change — so check them. Funders are actively developing positions on applicant and reviewer use of AI. Some address whether and how applicants may use AI in preparing proposals; many more sharply restrict reviewer use (Section 24.5). These policies change, so for any given proposal, check your funder's current stance, just as you would any other rule (Chapters 15, 19). The meta-rule: don't assume; verify the funder's current AI policy.

The accountability principle is stable even as policies shift. Whatever a funder's specific AI rules, one thing holds everywhere: you are accountable for the content you submit. No funder accepts "the AI wrote it" as an excuse for a fabricated citation, false data, or plagiarism. So even where AI use is permitted, the human-in-the-loop discipline (Section 24.4) is what keeps you safe — because the accountability is yours regardless of the tool.

Disclosure. Whether to disclose AI use is an evolving norm, and the answer is increasingly "follow the funder's policy, and when in doubt, lean toward transparency." Some funders may require disclosure of AI use; others are silent. Where a funder requires it, disclose; where unsure, transparency is the safer ethical default, and it's never wrong to be honest about your process. What is not acceptable is using AI in a way a funder prohibits, or hiding a use you suspect would not be permitted. The project checkpoint below has you write an AI-use note for exactly this reason: thinking it through in advance keeps you on the right side of both the rules and the ethics.

📐 Project Checkpoint — Write an AI-use plan and disclosure note for your proposal: For your project, write a short, honest AI-use plan: (1) Where you'll use AI — list the specific, verifiable tasks (e.g., brainstorming framings, editing your own prose for clarity, devil's-advocate critique, summarizing documents you provide). (2) Where you won't — the lines you won't cross (no AI-provided citations or statistics without verification, no AI-generated data, no budget guesses, no confidential material pasted into public tools, no funder-prohibited uses). (3) Your verification commitments — how you'll check every fact, citation, and number, and how you'll rewrite AI drafts into your own voice. (4) Disclosure — what your funder's current policy requires, and what you'll disclose. Save it in your "My Proposal" document. This plan is your guardrail; written in advance, it keeps you disciplined when deadline pressure tempts you to cut corners.

🎓 Going Deeper — the deeper accountability and authorship question: Underneath the practical rules sits a real question about authorship: in what sense is an AI-assisted proposal yours? The durable answer rests on accountability and origination. A proposal is yours when the ideas, the judgment, the real data, the funder strategy, and the final responsibility are yours — even if you used tools (including AI) to help express them. You've always used tools: spell-checkers, reference managers, colleagues' edits, a writing-center consultation. AI is a more powerful tool, but the authorship test is the same: did you originate the substance and do you stand behind every claim? Where AI merely helped you express your own ideas more clearly, the work is authentically yours. Where AI originated substance you didn't verify or understand, or supplied data or citations you can't stand behind, the work is no longer authentically yours — and that's the line, more than any specific policy, that should govern your use. The question to ask of any AI-assisted passage is: can I stand behind every word of this as my own accountable claim? If yes, you've used the tool well. If no, you've crossed the line, whatever the current policy says. Accountability, not the tool, defines authorship.

24.7 Strategy: A Tool That Augments an Accountable Author

Pull the threads together into a practice you can hold as the technology changes. Use AI for what it does well and can be verified — brainstorming, structuring, editing your own prose, summarizing what you provide, devil's-advocate critique. Refuse it for what it does badly or dangerously — supplying citations, statistics, data, budgets, or funder judgment you'd have to take on faith. Keep a human in the loop at every step, verifying every fact and owning every word. Protect confidentiality absolutely, and never breach the peer-review red line. Follow your funder's current policy and lean toward transparency. Above all, hold the threshold concept: AI can sharpen your argument but cannot own it — accountability stays with you.

Set the productive and perilous uses side by side as a working reference:

The deepest point is one this whole book has been building: a fundable proposal is the product of genuine thought, real evidence, specific funder understanding, and an accountable human author who stands behind every word — and none of those can be outsourced to a tool that predicts plausible language. AI can make that author faster and clearer. It cannot replace them, and the moment you let it try, you've traded away the very things — truth, specificity, voice, accountability — that make a proposal fundable. Used as a tool by an author who stays in charge, AI is a real asset. Used as a substitute for the author, it is a fast path to a fabricated citation and a ruined reputation. Stay the author; use the tool.

✅ Best Practice: Get more from AI, and stay safer, by how you ask. A few habits help. Give it your real material to work with rather than asking it to supply facts from nowhere — "tighten this paragraph I wrote," "summarize this document I'm pasting," "critique this draft" are far safer and more useful than "write me a needs section about X" (which invites generic, unverifiable output). Ask for options, not answers when brainstorming — you want a menu to judge, not a verdict to accept. Explicitly ask it to flag uncertainty — though even then, treat every factual claim as unverified until you check it; an AI's self-reported confidence is not reliable either. Iterate in your own voice — take its draft as raw clay and reshape it, rather than accepting its phrasing. And keep a clean separation in your own mind between the tasks where AI supplies expression and structure (safe, with verification) and the tasks where it would supply substance and fact (do these yourself). The writers who get the most from AI are not those who delegate the most to it, but those who delegate the right tasks and bring rigor to the output — turning a plausible-language generator into a genuine aid to their own accountable thinking.

🔄 Check Your Understanding: Sam uses AI in two ways while drafting a fellowship application: (a) to tighten and clarify three clumsy paragraphs Sam wrote about the research, and (b) to generate a list of citations supporting a claim about the field's significance. One use is safe; one is dangerous. Identify each and explain why, citing the chapter's organizing principle.

Answer

(a) is safe; (b) is dangerous. Editing Sam's own prose for clarity is a verifiable use — Sam wrote the content and can check that the edit preserved the science's accuracy and meaning (the low-risk pattern of Section 24.2). Generating citations is the cardinal danger (Section 24.3): AI fabricates plausible, perfectly-formatted references to papers that may not exist, and including any without independent verification risks a credibility-destroying fabrication. The organizing principle: use AI where you can verify the output and supply the substance; never trust it for substance you'd take on faith. Sam may use AI's suggested citations only as leads to verify against real databases — confirming each exists and says what's claimed — never as citations to drop in. Editing real prose keeps Sam the accountable author; trusting AI citations surrenders exactly the accountability the threshold concept says stays with the human.

Spaced Review

Retrieve these from earlier chapters without looking back, then check against the collapsed answers.

1. (From Chapter 23) How does the "one voice from many authors" discipline apply when AI is one of the "authors" contributing to a draft?
2. (From Chapter 8) Why does AI's fabricated-citation problem make Chapter 8's citation-honesty lesson even more critical, not less?
3. (From Chapter 2) Why is AI's tendency toward generic, voiceless prose a particular liability given who actually reads your proposal?

Answers

1. Exactly as with human co-authors: one accountable human owner must integrate any AI-drafted material into a single, coherent, specific voice that reflects the real project — rewriting generic AI prose into the author's own voice and verifying its content — so the proposal reads as one human's accountable argument, not a patchwork that includes unverified machine text. 2. Because AI makes fabricated citations easier to introduce and harder to spot (they're fluently formatted and confident), the Chapter 8 discipline of citing only real, verified sources that say what you claim becomes a frontline defense — every AI-suggested citation must be treated as a lead to verify, never a citation to use, or you risk the credibility-destroying fabrication Chapter 8 warned against. 3. Your reader is a specific, often tired, human reviewer (Chapter 2) who is moved by concrete, specific, passionate writing and bored or unconvinced by generic boilerplate; AI's characteristic voiceless prose reads as "anyone could have written this about any project," which is precisely the opposite of what engages a reviewer — so AI-drafted text must be rewritten into the specific human voice that actually wins.

Chapter Summary

Key Takeaways

• An AI language model (LLM) predicts plausible language, not verified truth — so it generates fluent output that may be confidently false. Hallucination (invented citations, statistics, findings) is intrinsic, and fluency is no evidence of accuracy.
• Threshold concept: AI can sharpen your argument but cannot own it — accountability stays with you. It cannot supply original ideas, real data, funder judgment, or accountability; you remain the author of every claim.
• AI genuinely helps where output is easy to verify and you supply the substance: brainstorming, outlining, editing your prose, summarizing documents you provide, devil's-advocate critique, jargon checks.
• AI is dangerous where it supplies unverifiable substance: fabricated citations and statistics (the cardinal danger — verify every one), fake data (fabrication/misconduct), inaccurate budgets, generic voiceless prose, missing funder nuance, and confidentiality breaches.
• Human-in-the-loop is the safe discipline: verify every fact/citation/number, supply the substance and judgment, own the voice, and remain accountable — "the AI did it" is never a defense.
• Confidentiality is a bright line: don't paste sensitive material into public AI tools, and never upload a confidential proposal you're reviewing (the peer-review red line, prohibited by major funders).
• Funder policies vary and evolve — check the current one; disclose per policy and lean toward transparency. Accountability is stable even as rules shift.

Action Items

1. Adopt the rule: use AI only where you can verify the output; never trust it for substance on faith.
2. Verify every AI-suggested citation, statistic, and number against a real source before it enters your proposal.
3. Rewrite any AI-drafted prose into your own specific, accountable voice grounded in your real project.
4. Protect confidentiality — keep sensitive material out of public tools; never put a reviewed proposal into AI.
5. Check your funder's current AI policy and write an AI-use/disclosure plan before you start.

Common Mistakes

• Trusting an AI citation because it looks real — the most common and most damaging error; fabrications are flawlessly formatted.
• Including AI-generated data, results, or budget numbers — fabrication and inaccuracy with serious consequences.
• Submitting generic, voiceless AI prose that no specific human seems to have written.
• Pasting confidential material into public AI tools, or (as a reviewer) uploading an application under review.
• Assuming a funder permits AI use (or doesn't require disclosure) without checking the current policy.

Decision Framework — "Should I use AI for this task?"

1. Can I easily verify the output? → If yes (editing my prose, brainstorming, summarizing what I provide), proceed with verification. If no, stop.
2. Am I supplying the substance, or is AI? → I supply ideas, data, funder judgment; AI helps express/organize. If AI would originate unverifiable substance, don't.
3. Is anything confidential involved? → Keep sensitive material out of public tools; never upload a reviewed proposal.
4. What does my funder's current policy say? → Follow it; disclose as required; lean toward transparency.
5. Can I stand behind every word as my own accountable claim? → If yes, I've used the tool well. If no, I've crossed the line — fix it before submitting.

🔁 Carry this forward: AI is Part IV's third cross-cutting skill — a tool that, used with discipline, serves every funder and sector you've met. Next, diversity, equity, and inclusion in grant writing (Chapter 25) takes up questions that are both about what you propose and how the funding system works — questions of who gets funded, whose communities are served, and how to write proposals that advance equity authentically. The accountability and authenticity disciplines you built here — own your claims, write in a real voice, don't perform what you don't practice — carry directly into that terrain.