Case Study 24.1 — Two Applicants, One Tool

A composite, illustrative case constructed to contrast safe and unsafe AI use. The applicants are composites; the AI failure modes and their consequences are real. Verify your funder's current AI policy.

Why this case: the same tool, opposite outcomes

The clearest way to teach the line between safe and unsafe AI use is to watch two applicants use the same tool on the same kind of task and arrive at opposite outcomes — one strengthened, one sunk. Meet Sam Okonkwo, the doctoral student preparing an F-series fellowship application (Chapters 6, 16), and a composite peer we'll call Devin, preparing a similar application. Both use an AI assistant. Sam stays the author; Devin lets the tool become the author. The difference decides their fates.

Sam: AI as a verified assistant

Sam uses AI deliberately and within bounds, exactly as Section 24.2 prescribes:

  • Brainstorming. Stuck on how to frame the significance of the dissertation research, Sam asks AI for several framings, picks two that resonate, and rewrites them in Sam's own voice, grounded in the specific science.
  • Clarity editing. Sam drafts the research strategy, then asks AI to tighten clumsy paragraphs — and checks each edit to confirm it preserved the scientific meaning (catching one place where AI's smoothing subtly overstated a result, which Sam corrects).
  • Devil's-advocate review. Sam asks AI to critique the draft as a skeptical reviewer; it flags an unclear aim and an unaddressed feasibility question — both fair, both fixed.
  • Summarizing the funder's guidelines. Sam pastes in the funder's instructions and asks for a summary to make sure nothing's missed — then verifies the key requirements against the original.

Crucially, Sam never asks AI for citations or statistics, builds the biosketch and training plan from real facts, and runs a verification pass (Section 24.4) on everything AI touched. Sam's submitted application is faster and clearer for the AI help — and entirely true, specific, and Sam's own. It reads like a real scientist who cares about real work. Sam remains the accountable author throughout.

Devin: AI as a ghostwriter

Devin, rushed and tempted, uses the same tool very differently:

  • Asks AI to "write the significance section about" the topic — and accepts the fluent, generic result with light edits.
  • Asks AI for citations supporting the field's importance — and drops the perfectly-formatted references straight into the bibliography without checking them.
  • Asks AI for statistics about the problem — and uses a confident-sounding figure as an anchor, unverified.
  • Lets the AI voice stand — generic, hedged, characterless — across much of the application.

Devin submits faster than Sam, and the draft looks polished. But it carries three time bombs: two of the AI-provided citations are to papers that don't exist (hallucinated, Section 24.3); the anchor statistic is fabricated; and the prose has the generic, voiceless quality of unedited AI output.

What happened at review

A reviewer who knows the field tries to follow up one of Devin's citations, can't find it, checks another, and realizes it's invented. The effect is immediate and total (Section 24.3's pitfall): the reviewer loses trust in the entire application — if this citation is fabricated, what else is? — and flags a possible integrity concern. Even setting the citations aside, the generic prose fails to move the panel; it reads like no specific person wrote it about no specific project. Devin's application is not funded, and the fabricated-citation issue follows Devin uncomfortably.

Sam's application, by contrast, is evaluated on its real merits — strong, specific, true — and is competitive.

What this case teaches

  1. The same tool, opposite outcomes. AI didn't help or hurt inherently; how each applicant used it decided everything. Sam stayed the author; Devin abdicated authorship.
  2. The cardinal danger is real and catastrophic. Devin's unverified AI citations weren't a small error — they collapsed the application's credibility and raised an integrity flag.
  3. Verifiability is the line. Sam used AI only where the output could be verified and the substance was Sam's; Devin trusted AI for exactly the substance you must never take on faith.
  4. Voice matters. Even apart from the fabrications, Devin's generic AI prose failed to engage the human reviewer; Sam's real voice succeeded.

🔄 Retrieve: Without rereading, name (a) the two AI uses that sank Devin, and (b) the verification habit that protected Sam. (Answers above.)