Grant Proposal AI Writing: A Real Test and the Review It Needs

Grant proposals are a natural place to use AI for a first draft, and a risky place to stop there. We ran an AI-drafted project summary for an agriculture grant through Humanizer v2. All five detector models voted AI on the draft. On the output, all five voted human, and every sentence read as human.

By detector numbers, it is the cleanest result in our case-study series. It also lost a project duration, picked up a word reviewers distrust, and changed one of the trial's outcome measures. This post shows all of it, because the review step is the part that decides whether the summary is fit to submit.

The draft and the run

The summary follows the standard arc: the problem, the aim, the partners, the design, a qualitative strand, the deliverables, the impact. It has the vocabulary of a thousand funded proposals: "mounting challenges", "interconnected challenges", "rigorously assess", "has the potential to significantly enhance". All seven sentences were flagged.

1. Scanning for AI signalsSeven of seven sentences flagged; all five models voted AI.
2. Rewriting flagged sentencesAll seven sentences were rewritten.
3. Re-checking with AI detectorsSeven of seven sentences read as human; all five models vote human.
4. Saving your result1,352 characters at 500 characters per credit, rounded up: 3 credits.

• AI-drafted grant summary5 of 5
• Humanizer v2 output0 of 5

Source: Humanizer v2 production run. Report verdict moved from AI-generated to Human-written.

The climate change scenario is also changing the agricultural environment in semi-arid regions and smallholder farmers are facing increasing uncertainties related to rainfall variability, soil quality degradation, and lack of effective information. To counteract this situation, this research aims to develop and evaluate an innovative low-cost SMS-based advisory service which provides localized planting and fertilization advice for farmers in two districts of northern Ghana. We will collaborate with 1,200 households, local extension officers and a regional meteorological agency to co-develop our advisory content and deliver it. We will perform a cluster randomized controlled trial to measure the effect of our approach on maize yield, inputs used, and food security. We will also conduct qualitative interviews with farmers to better understand their trust in digital advice and the barriers to adopting it, especially those specifically targeted towards women farmers. The project aims to produce a validated advisory model, open source data sets, and policy briefs for national agricultural agencies. This project should strengthen the resilience and livelihoods of smallholder farmers and offer valuable lessons that can be applied by other digital agriculture initiatives in sub-Saharan Africa.

Real output from our test run. All seven sentences read as human.

Read that output once and it sounds fine. Read it against the draft and it does not.

The review step: what was lost and what was added

Here are the two sentences that carry the project's commitments: what will be built, with whom, and for how long.

Before

This project aims to address these interconnected challenges by developing and evaluating a low-cost, SMS-based advisory system that delivers localized planting and fertilization recommendations to farmers in two districts of northern Ghana. Over a three-year period, we will partner with 1,200 households, local extension officers, and a regional meteorological agency to co-design the advisory content and delivery schedule.

After

To counteract this situation, this research aims to develop and evaluate an innovative low-cost SMS-based advisory service which provides localized planting and fertilization advice for farmers in two districts of northern Ghana. We will collaborate with 1,200 households, local extension officers and a regional meteorological agency to co-develop our advisory content and deliver it.

Three problems in two sentences:

• The three-year duration is gone. "Over a three-year period" was dropped entirely. In a grant summary, the timeframe is a core commitment that reviewers check against the budget and work plan. It has to go back in.
• "Innovative" was added. The draft did not call the system innovative; the rewrite did, without evidence. Unsupported claims of novelty are a common reviewer complaint, and this one is not the author's claim.
• The co-design scope narrowed. Partners were going to co-design the content and the delivery schedule. Now they co-develop content, and the project simply "delivers it". Co-designing timing with farmers is a real methodological choice, and it vanished.

The 1,200 households, the two districts and northern Ghana all survived. So did the core idea of an SMS advisory. The damage is in the details a reviewer scores.

The other five sentences

The rest of the output needs the same pair-by-pair reading. What we found:

• Problem statement. The rewrite opens with "The climate change scenario", a framing the draft never used, and "limited access to timely agronomic information" became "lack of effective information". Timeliness was the point of an SMS service.
• Trial outcomes. "Input use efficiency" became "inputs used". Those are different measures: one is a ratio of output to input, the other a quantity. "Household food security" lost "household", the level at which it would be measured.
• Gender focus. "With particular attention to women farmers" became "especially those specifically targeted towards women farmers", which now says the barriers target women. That is a different and unsupported claim.
• Deliverables. "Open-access datasets" became "open source data sets". Open access is about who can read the data; open source usually describes software licensing.
• Impact. "Has the potential to significantly enhance" became "should strengthen", and "across sub-Saharan Africa" became "in sub-Saharan Africa". The first change is arguably an improvement in tone; still, read it as a claim you are making to a funder.

Restoring the specifics

None of these fixes requires undoing the run. Each is a phrase-level edit inside a sentence that now reads as human. For example, an author could restore the partnership sentence like this (our own edit, not Humanizer output):

Over three years, we will collaborate with 1,200 households, local extension officers and a regional meteorological agency to co-design our advisory content and its delivery schedule.

Then delete "innovative", put "input use efficiency" and "household food security" back, rewrite the women farmers clause, and change "open source data sets" to "open-access datasets". After hand edits, re-check the text. Restored phrases can bring back some of what detectors reacted to, and the report will show you if they did; you can humanize just those passages again and review the new pair. Our guide to humanizing AI text by hand covers this kind of phrase-level edit.

Before you submit, check your funder's current rules on AI use in applications. Several funders now publish them, and they differ in what they allow and what must be disclosed.

Why we published the flaws

It would have been easy to show the bar chart and stop: five AI votes to zero. But a grant summary is a set of promises, and a humanizer that reads as human while quietly dropping a three-year commitment has not done the job on its own. The detector result and the meaning check are two separate tests, and a summary has to pass both.

Academic Humanizer is designed so the second test is fast. The AI Detection Report pairs every flagged sentence with its rewrite, so a check like the one above takes minutes, not a full re-read. Our guide on preserving meaning when humanizing lists what to look for, and the responsible revision workflow shows where that check fits before a proposal goes out.