Every fall it arrives like weather: the stack of college recommendation letters. If you’re a high-school counselor with a caseload in the hundreds, “write a strong, personal letter for each senior” can quietly eat every weekend from September to December. So it’s no surprise that, in a 2024 survey of U.S. teachers, about one in three said they’d used AI to help write at least one recommendation letter — and most of them said the reason was simply to survive the workload.
The questions underneath that, though, are the ones that keep people up at night. Will a college catch it? Is it even allowed? Am I breaking a privacy law by pasting a kid’s info into ChatGPT? Let’s answer those honestly — because the real risks are not the ones most people worry about, and once you understand them, there’s a safe, genuinely useful way to do this.
What’s actually true in 2026
Three things, and they’re not what the panic suggests.
“Colleges will detect it” is mostly the wrong worry. AI detectors are unreliable on recommendation letters specifically — one study found experienced reviewers could tell an AI letter from a human one only about 59% of the time, barely better than a coin flip, and the popular detector tools did no better. On top of that, an analysis of 174 universities’ AI policies found that essentially none of them even mention letters of recommendation — the rules are written for student essays, not recommenders. So no, there isn’t a reliable robot waiting to flag your letter.
But a human reader still might notice — and that’s the real risk. Admissions officers read hundreds of letters from the same schools and counselors, year after year. They develop an ear for how you write. A letter that’s technically fine but sounds like generic, over-polished AI mush — all “exemplary candidate” and “unwavering dedication” — reads as off in a way a person clocks before any software would. A weak letter doesn’t get a student flagged; it just doesn’t help them, which is worse.
The genuine legal risk is privacy, not plagiarism. This is the one to take seriously. Under FERPA — the federal student-privacy law — a student’s records and identifying details are protected. Pasting a student’s name, grades, disciplinary notes, or other record details into a public AI tool can cross that line. (For scale: experts have argued that even uploading a named student essay to a plagiarism checker can be a FERPA problem.) This is the rule that actually matters, and it’s also easy to follow.
The safe workflow: AI drafts the scaffold, you write the soul
The approach that works treats ChatGPT as a first-draft accelerator, never the author. The trick is to feed it the substance without feeding it the student’s identity, and to keep the final voice unmistakably yours.
Step 1 — Start with a brag sheet. This is good practice with or without AI: have the student fill out a questionnaire — favorite project, a moment they’re proud of, a challenge they overcame, what they want a college to know. That’s your raw material, in their words.
Step 2 — Strip the identity before anything goes near AI. Replace the name with “the student,” drop grades and any record details, and turn the brag sheet into anonymous bullet points. You want the qualities and stories, not the identifiers.
Step 3 — Ask for a scaffold, not a final letter:
“Help me draft a college recommendation letter from a school counselor. Here are anonymized notes about a student: [paste de-identified bullets]. Give me a structure and a first draft — an opening that names a specific strength, two body paragraphs each anchored to one concrete example, and a closing. Plain, warm, specific. Avoid clichés like ’exemplary’ and ‘unwavering.’ Leave clear [BRACKETS] where I should add details only I know.”
Step 4 — Rewrite it into your voice. This is the part that can’t be skipped. Replace the brackets with the real anecdote, cut anything that sounds like a press release, and add the one observation only you could make — the thing you noticed in a hallway conversation in October. By the end, the AI’s fingerprints should be gone and yours should be all over it.
There’s a useful gut check buried in all this. As one counselor put it: by the time you’ve written a prompt good enough to get a truly personal letter, you’ve basically written half the letter yourself. That’s not a failure of the tool — that’s the point. The AI removes the blank-page friction and handles the boilerplate; the part that makes the letter work was always going to be you.
What this means for you
If you’re a counselor with a huge caseload: use AI for the structure and the routine scaffolding so you can spend your limited time on the personalization that actually moves a letter. A consistent brag-sheet process is half the battle; the AI draft is the other half.
If you’re a teacher who writes letters: same playbook. The students who ask you specifically want your take on them — keep that front and center, and let AI handle the “okay, how do I even start this paragraph” friction.
If you’re drowning in back-to-school comms too: the same de-identify-then-draft approach works for parent-night messages, 504 check-in notes, and schedule-change replies — anything routine and non-confidential. Just keep student records out of it.
What this can’t do
- It can’t fake your relationship with the student. The specific moment, the growth you witnessed, the “I’ve watched this kid for four years” — AI has none of that, and it’s exactly what makes a letter persuasive.
- It can’t make FERPA optional. De-identifying isn’t a nice-to-have; it’s the rule. Never paste a student’s name or records into a personal AI account.
- It sounds generic by default. Left unedited, AI letters cluster around the same flattering, hollow phrases — which is precisely what an experienced reader tunes out. The rewrite is mandatory.
- It can invent things. It might add an award the student never won or a class they never took. Verify every concrete claim before it goes out under your signature.
- It doesn’t replace your judgment about whether to write. If you can’t write a genuinely supportive letter, an AI draft won’t fix that — and shouldn’t.
The bottom line
The honest version of this is freeing: you’re not in a cat-and-mouse game with an AI detector, and you’re not breaking the rules by getting help with a first draft. You’re using a tool to clear the boilerplate so your real work — the specific, personal, only-you-could-write-it part — gets your full attention. Protect student privacy, keep your voice, and AI becomes what it should be here: a way to give every senior a strong letter without losing your whole autumn.
Want to get sharper at the part that matters — making AI drafts sound like you instead of a robot? Our AI for Writers course is all about steering tone and voice so the output reads human. First two lessons are free, no signup needed.
Sources
- Teachers using ChatGPT for recommendation letters (Foundry10 survey data) — GradPilot
- Are high-school counselors encouraging AI for college applications? — EdSurge (July 2025)
- Brain versus bot: can reviewers detect AI-written recommendation letters? — AEM Education and Training (PMC)
- What the data shows on detecting AI recommendation letters — RevolutionInAI (Mar 2026)
