Safe AI help · Updated 15 September 2026

Recommendation Letters and AI: What You Can and Cannot Do

Two things are happening at once this application season, and students keep worrying about the wrong one.

If your recommender writes your letter with help from a chatbot, that is common, usually permitted, and not your call. If you write it — even because they asked you to, even if a chatbot did the phrasing — that is the version that can break your application, and several universities say so in writing. So do not spend your energy policing your professor's tools. Spend it on the one thing that actually changes what the letter says: the specific, concrete material you hand them.

The rule that catches students is about authorship, not AI

Being asked to draft your own letter is so routine that top universities now write policy clauses anticipating it. Stanford Graduate Admissions puts it about as plainly as a rule can be put:

Your recommenders must be the sole authors of your letters of recommendation. Drafting, writing, translating, or submitting your own reference, even if asked to do so by a recommender, is a violation of our application terms.

Read the clause that follows it too, because it settles who carries the risk: "You are responsible for informing your recommenders of this policy." Your professor asking is not a defence. The duty to say no is yours.

The University of Chicago Division of the Social Sciences goes one step further and turns it into something you sign. Letters "must be written entirely by the recommender," the division does not accept letters submitted by applicants, and "Applicants must certify that they did not write or submit their own letters." That is no longer a guideline you might have missed. It is a statement you personally attest to, which is the kind of thing that gets applications withdrawn after the fact.

You probably already signed something that contradicts it

There is a second problem with drafting your own letter, and it bites even at schools with no authorship clause. Almost every graduate application asks whether you waive your right to inspect your letters, and Harvard Griffin GSAS explains why that box matters: "admissions committees may consider letters of recommendation as more credible sources of information when the applicant has waived their right to access the letter." Crucially, "both your recommenders and your program of interest can see whether you have chosen to waive your right to inspect your recommendations."

So the waiver is a signal, and the signal is confidentiality: this letter was written about me, not with me. Draft the letter yourself and you have read it in the most complete way possible — you composed every line. The waiver still says what it says.

Harvard also quietly closes the obvious escape route. Not waiving gains you almost nothing, because the right to inspect only exists "if you enroll," and applicants who are rejected or who go elsewhere never get to see anything. You give up credibility in exchange for access you will probably never have.

Your recommender is probably using AI — and that is mostly fine

The other half of the worry is real but largely out of your hands. In January 2026 the AAMC reported that faculty are routinely using Copilot, ChatGPT, Claude and Gemini to turn rough notes into finished residency letters. A 2024 foundry10 survey of 425 US high school teachers found about 31% had used generative AI on at least one college recommendation letter — and the students surveyed alongside them split three ways on whether that was acceptable, 38% saying no, 31% yes and 30% undecided. That survey covers high school teachers writing college letters rather than professors writing graduate ones, so treat it as an indication of direction, not a measurement of your department.

Some fields have drawn a line anyway. PAEA, which runs the CASPA application service used by physician assistant programmes, said in guidance to programs that letters of recommendation or evaluation must be "the original work of the person who is submitting them" and "can't be created by generative AI." That guidance was published in 2023, so confirm the current CASPA instructions before relying on it — but if you are applying anywhere in the health professions, it is worth mentioning to your recommenders, who may not know a rule like this exists.

Everywhere else, you are not the enforcement mechanism, and you should not try to be. What you can do is change the input.

What actually decides your letter

The most useful line in the whole AAMC piece is a urologist describing what he feeds the tool. Ronald Rodriguez, MD, PhD, says he tells it: "I say this person worked with me for six months on a project in which we did this or that, and I can reflect on the kind of person he is."

That is the shape of the thing. The model supplies fluency. The recollection — six months, that project, this specific person — has to come from the recommender, and it is the only part a reader cannot get from any other letter in the pile. Which means the more of your file is written by machine, the more the scarce ingredient is concrete detail about you, and the more your evidence packet is worth.

Chicago's policy page is unusually helpful here, because having told you what you may not do it tells you exactly what you may: if a recommender asks you to draft the letter, "you may provide supporting information—such as your CV, research summary, or information about your goals—but you may not write the letter yourself."

That is the whole permitted lane, and it is wider than it looks. A good packet is two pages and contains things only you can supply: the exact course, term and grade; the title of the project and what you personally did on it; the problem you got stuck on and how you got unstuck; the moment in office hours they might remember; where you are applying and what each programme says it wants. Facts they can verify and quote. Not adjectives.

Split the job

The workflow, step by step

Four prompts. None of them writes a letter or any part of one — that is the line, and these are built to stay well behind it.

1. Turn your own records into a packet. Feed it your real material: syllabus, transcript line, assignment briefs, your own notes and drafts.

You are helping me build a factual information packet for a
professor who has agreed to write me a recommendation letter.

Here is my raw material:
[paste your CV, the course details, your project description,
your own notes about what you did]

Rules:
- Do NOT write any part of a recommendation letter.
- Do NOT write anything in my professor's voice.
- Do NOT add any achievement, skill or quality I did not state.
  If something is missing, ask me for it.

Organise ONLY what I gave you into:
1. How we know each other - course code, term, my grade, contact
   hours, anything I presented or led.
2. Specific things I did, one line each, each with a concrete
   detail a reader could check.
3. Where I am applying and the deadline for each.

Then list up to 8 questions I should answer to fill the gaps.

2. Audit the packet for adjectives pretending to be evidence. This is the highest-value two minutes in the process.

Here is my information packet:

[paste it]

Do not rewrite it and do not improve the wording.

For each bullet, label it:
- EVIDENCE, if it states something specific and checkable that
  my professor could quote in a letter, or
- ADJECTIVE, if it is a claim about my qualities that a reader
  could not verify and that could appear on anyone's packet.

For every ADJECTIVE, ask me the question whose answer would turn
it into EVIDENCE. Do not answer those questions yourself and do
not invent examples.

3. Ask properly, and give them a way out. A lukewarm letter does more damage than a missing one, so the escape hatch is not politeness — it is self-defence.

Help me write a short email asking a professor for a graduate
school recommendation letter. Keep it under 200 words, plain and
not flattering.

Facts to use, and nothing beyond them:
[course, term, grade, project, what you are applying to, deadline]

It must:
- Ask whether they can write a STRONG letter, and say plainly that
  a no is genuinely fine.
- Offer the attached packet and offer to meet.
- Give the earliest deadline and how the letter gets submitted.
- Ask whether they want anything else from me.

Ask me anything you need rather than inventing it. Give me two
versions, one warmer and one more formal.

4. Say no to the draft request without damaging the relationship. Use this the moment you get "why don't you write something up and I'll sign it."

My recommender asked me to draft my own letter for them to sign.
My application terms say the recommender must be the sole author
and I have to certify I did not write it.

Write me a short, warm reply that:
- Thanks them and does not lecture them or quote policy at length.
- Says I am not permitted to draft or write any part of it.
- Immediately offers the alternative: a two-page packet of dates,
  projects and specifics, plus a call, so it costs them less time
  than drafting would have.
- Keeps the door open and does not sound like an accusation.

Under 150 words. No bullet points.

Where the line is

Everything above sits in the permitted lane: your own material, your own emails, your own logistics, with a tool doing the sorting. The one prohibited act is being the author of a reference about yourself, and no amount of editing or paraphrasing converts that into something else. If the wording on a programme's page is ambiguous, ask the admissions office in writing before you act — our email template for asking about permitted AI use adapts to that in a line or two, and the guide to writing an AI disclosure statement covers the cases where you are asked to declare what you used.

Related reading

• The rest of the application. What the rules actually say about AI on your personal statement — the essay you are allowed to write, with its own certification trap.
• If someone doubts your writing. Falsely accused of using AI, and why detector output is weak evidence.
• The job-search versions. Tailoring a resume and cover letter, internship outreach emails, and writing a LinkedIn profile that survives AI search — where retroactively improving your own record causes the same trouble.
• Before you paste anything. The privacy checklist for uploading class materials — your transcript and draft packet count.
• The general rule. Is it cheating to use AI for schoolwork?

FAQ

My professor asked me to draft my own recommendation letter. Can I?

Check the application terms before you agree, because at some schools the answer is a flat no and the violation is recorded against you rather than your professor. Stanford Graduate Admissions states that "Drafting, writing, translating, or submitting your own reference, even if asked to do so by a recommender, is a violation of our application terms," and adds that "You are responsible for informing your recommenders of this policy." The University of Chicago Division of the Social Sciences requires that letters be "written entirely by the recommender" and that "Applicants must certify that they did not write or submit their own letters." The request is common and usually well meant, but common and permitted are not the same thing.

Does it change anything if I only used AI to polish the draft?

No, and this is the part students get wrong. The clauses above are authorship rules, not AI rules. Neither the Stanford page nor the Chicago page mentions artificial intelligence at all. What they prohibit is you being the author of your own reference, so whether the words came out of your head, a template or a chatbot makes no difference to whether the rule was broken. Running your draft through a model does not convert it into your recommender's letter. It just makes your letter sound like everyone else's.

Is it a problem if my recommender uses AI to write my letter?

Usually it is allowed, it is increasingly common, and it is not your decision to make. The AAMC reported in January 2026 that faculty are using tools like Copilot, ChatGPT, Claude and Gemini to turn rough notes into finished residency letters. A 2024 foundry10 survey of 425 US high school teachers found about 31% had used generative AI on at least one college recommendation letter. Health professions are the exception: PAEA has stated that under CASPA policy, letters of recommendation or evaluation must be "the original work of the person who is submitting them" and "can't be created by generative AI." Outside rules like that one, your leverage is not over the tool. It is over the material your recommender puts into it.

Will admissions run my letters through an AI detector?

Some readers will try, and at least one admissions body has said publicly that a detector result is not enough on its own. PAEA's guidance to PA programs states that "PAEA will not investigate an applicant if the only evidence the applicant did not write his or her personal essay comes from AI detection software," on the view that "current AI detection tools are simply not reliable enough yet." Its general counsel noted that detectors have "fairly high false positive and false negative rates" and that the developers themselves "warn in their terms of service against using them as the sole basis to make important decisions about students." Do not treat that as safety, though. A letter can be discounted for reading like nothing in particular without anyone running a tool.

Should I waive my right to read my letters of recommendation?

Most applicants do, and schools say plainly why it matters. Harvard Griffin GSAS notes that "admissions committees may consider letters of recommendation as more credible sources of information when the applicant has waived their right to access the letter," and that "both your recommenders and your program of interest can see whether you have chosen to waive your right to inspect your recommendations." It also spells out the limit of not waiving: you can inspect the letters only if you are admitted and actually enrol, so declining the waiver buys you very little. The waiver is also the cleanest reason to refuse a request to draft your own letter, because you cannot honestly claim not to have seen a letter you wrote.

Bottom line

Whether a chatbot helped write your letter is your recommender's problem. Whether you wrote it is yours, and it is the one that shows up in application terms and on a certification you sign. Say no to the draft request — warmly, immediately, with a better offer attached — and put the effort into a two-page packet of dates, projects and specifics. That packet is the input to whatever process your recommender uses, machine-assisted or not, and it is the only part of this you control.