AI-Written Letters of Recommendation: Why Readers Notice

AI-Written Letters of Recommendation: Why Committees Notice

Two facts define the AI-and-recommendation-letters problem in 2026, and they point in opposite directions.

Fact one: letter writers are already using AI at scale. In a Foundry10 survey of 425 U.S. high school teachers, 31% reported using generative AI to help write college recommendation letters — with "reducing stress" the most-cited reason. There is no reason to believe overloaded professors, attendings, and PIs behave differently; medical-education journals are openly debating the practice ("Should my recommendation letter be written by artificial intelligence?").

Fact two: almost nobody has rules about it. When we audited the AI policies of 174 U.S. universities for our AI-policies directory, zero included letters of recommendation in their policy scope. The attestations, detection screening, and rescission clauses all aim at the applicant's documents. The recommender's ChatGPT tab is policy-invisible.

So AI-drafted letters are widespread and mostly unregulated. Why, then, do they keep failing? Because the recommendation letter is the genre AI is structurally worst at — and the readers are the audience best positioned to notice.

The Genre Problem: A Letter's Value Is Exactly What AI Can't Supply

Strip any strong recommendation letter to its load-bearing parts and you get three things:

1. First-hand observed behavior. The AAMC's letter-writer guidelines prescribe situation-behavior-consequence anecdotes from direct observation. Common App's recommender guidance compresses the same rule to "Anecdotes outshine adjectives. Always."
2. A credible comparison. "Among the 40 premeds through my lab in 15 years, one of the three strongest." Peer-reviewed evidence says the superlative comparison is one of the only letter features that predicts anything.
3. A calibrated verdict the writer stakes their name on.

Now consider what a language model can produce from the prompt a busy recommender actually types — three bullet points and "write a warm, specific letter." It cannot supply observations that were never given to it, it has no comparison group, and it has no name to stake. What it generates instead is the statistical average of recommendation letters: fluent, warm, structurally perfect, and evidentially empty. "She consistently demonstrated exceptional dedication and a genuine passion for medicine" is what the genre sounds like with the information removed.

Admissions readers were discounting exactly that letter long before ChatGPT — the AAMC's own survey found more than half of admissions officers dissatisfied with letter quality years before generative AI existed. AI did not create the generic letter. It industrialized it, and made its tells more uniform.

The Tells Readers Report

Committee readers — and residency program directors, who collectively triage on the order of 200,000 ERAS letters a cycle — describe a recognizable cluster:

• Adjective density without incident. Long strings of evaluative language ("exceptional," "remarkable," "unwavering") with no scene where anything happens.
• Symmetric structure. Neat one-paragraph-per-trait construction, every paragraph the same length — the shape of generated text, not of a person recalling a student.
• The missing relationship statement. AI drafts routinely omit or fudge duration, capacity, and direct-observation basis — the first thing a trained reader checks.
• Voice convergence. When multiple letters in one file were AI-drafted from similar prompts, they start sounding like each other — and sometimes like the applicant's essay, if the applicant supplied the bullets. Application readers see the whole file at once; convergence is visible across documents in a way no single writer anticipates.

Detection research backs the intuition: a Scientific Reports study found reviewers and classifiers could distinguish AI-generated application materials from human-written ones at rates well above chance, with AI text marked by its homogeneity. And some gatekeepers are formalizing it — CASPA's applicant agreement reserves the right to run AI-detection on submitted materials, an enforcement posture we analyzed in our CASPA AI detection deep-dive, and some medical schools screen application materials with AI tools of their own.

The Asymmetry Nobody Fixes: The Applicant Carries the Risk

Here is the part that should bother you regardless of which side of the letter you are on. When an AI-drafted letter reads generic, the writer suffers nothing — the applicant absorbs the weak letter's consequences. When it trips a detector or a suspicious reader, the doubt lands on the applicant's file, not the recommender's inbox. The applicant is the only party in the transaction with no visibility into the letter (they waived their right to view it) and full exposure to its failure modes.

That asymmetry is why our position is not "AI in letters is cheating." Policy-wise, almost nobody says it is — for recommenders. Our position is narrower and harder to argue with: an AI-drafted letter is usually a bad letter, and in the one application component the applicant cannot proofread, bad is expensive.

What to Do Instead

If you write letters (professor, physician, counselor, PI): the workable division of labor is content from you, cleanup from the machine. Draft the anecdotes, the relationship statement, the comparison, and the verdict yourself — the four things only you possess — and if you then want AI to smooth transitions or trim length, the letter is still yours. Our guide to writing a medical school letter of recommendation gives the full structure; it takes about an hour to do honestly, which is roughly what the prompt-and-paste workflow costs once you include fixing it.

If you are an applicant asked to draft your own letter: do not compound authorship risk with generation risk. An AI-drafted, self-drafted letter is the worst object in this space — detectable, generic, and unowned by anyone. The self-draft playbook exists for exactly this.

If you have a draft letter in hand — whichever role you are in: read it against the three load-bearing parts above. No incident, no comparison group, no committed verdict = the letter reads generated whether or not it was. Then test it the way committees increasingly do: GradPilot reviews draft recommendation letters for evidence density, comparison quality, and hedged language, and runs the draft through our AI detector — the same check we run on medical school essays — in minutes. We review the letter you are drafting; we never write it.

The blind spot in institutional policy will close eventually; blind spots this size always do. The reading habits of admissions committees closed years ago. Write — or fix — the letter for the reader that exists.