When Michigan State University Extension put OpenAI's ChatGPT to work drafting a seeding rate prescription for a 2026 corn field, the chatbot returned a tidy three-step plan with rates of 34,000–36,000 seeds per acre in high zones, 32,000–34,000 in medium zones and 28,000–31,000 in low zones. The numbers looked plausible. MSU's agronomists flagged them as a starting point at best, and in some cases a profitability risk.

The comparison, published by MSU Extension, cuts to the core of the debate over artificial intelligence in precision agriculture. Large language models from Google, Microsoft, OpenAI and X can summarize data quickly, cite sources and structure a workflow. What they cannot do is account for the specifics of an individual field — row spacing, hybrid selection, landscape position, soil type — without the farmer supplying that context first.

Where the AI fell short

ChatGPT recommended building prescriptions from multi-year yield averages grouped into three to five productivity classes. MSU Extension instead points to yield stability maps, which sort fields into four classifications: high and stable, medium and stable, low and stable, and unstable. That distinction matters. An unstable zone can post a strong yield one year and a poor one the next, and seeding it like a consistently low performer wastes input dollars.

On rates, the extension service cited recent research showing a significant profitability drop when seeding rates increase, and directed growers to its Crop Budget Estimator to calculate 2026 cost of production before locking in populations. Research suggests most farmers set rates from their own experience, supplemented by private crop consultants and university extension recommendations — a hierarchy AI has not displaced.

The chatbot handled file logistics reasonably well, noting that ISOXML and shapefile are the two most common export formats. MSU added the fine print: a shapefile bundles four file types (.shp, .shx, .dbf and .prj), and platforms like John Deere Operations Center and Climate FieldView each require different import steps.

Data queries: strong on summaries

Microsoft Copilot performed better as a research assistant than as an agronomist. Asked for April precipitation at coordinates in central Michigan, it returned station-based NOAA records for Lansing: 3.53 inches in 2022, 3.89 in 2023, 3.35 in 2024 and 2.80 in 2025 — and correctly noted that monthly station totals can diverge from field-scale storm effects.

Asked for weekly urea prices, Copilot tracked the 2026 rally: roughly $595–600 per ton in early February, $611 by late February, $674 by mid-March and a sharply wider $750–840 range in late March, citing DTN Retail Fertilizer Trends and USDA-AMS Illinois production cost reports. Growers weighing spring nitrogen purchases got an accurate picture of the firming market, with the caveat that late-March quotes varied widely by region as global supply disruptions intensified.

On phosphorus, Copilot refused to give a blanket rate, correctly insisting that Tri-State recommendations are soil-test driven: below the critical level, apply starter plus broadcast; at critical, crop removal only; above critical, no phosphorus required.

The limits

MSU's bottom line is blunt on two fronts. First, neither AI nor humans can predict weather, and effective nitrogen decisions in grain crops require weighing past conditions, current weather and plausible future scenarios. Second, farmers should resist feeding LLMs their own farm data before defining clear objectives. The models also carry memory of prior conversations, so responses are never truly generic.

Basso's Digital Agriculture Lab at MSU is working with participating farmers to deliver turn-key seeding and nitrogen prescriptions for Michigan field crops. Growers interested in the program can contact the lab directly to implement the technologies on their own operations.