2026 Recommendation Letter Conventions and AI

The Common Application has issued guidance for counselors and recommenders that addresses generative-AI use across both applicant essays and recommendation letters. The position, as articulated in successive updates to the Common Application's counselor and recommender resources, is that recommendation letters should reflect the recommender's own substantive observations and judgement. Recommenders may use AI tools as writing aids (for structure, tone, or polishing) but must not rely on AI to generate the substantive content of the letter.

The reasoning the Common Application has articulated: the value of a recommendation letter to admissions committees lies in the recommender's personal knowledge of the candidate, expressed in the recommender's own voice. AI-generated content from minimal recommender input does not carry that value, and submitting such content as a recommendation undermines the application process's integrity. The applicant is not held responsible for the recommender's choices about AI use; the recommender is the responsible party.

The Common Application has not implemented automatic AI-detection on uploaded recommendation letters as a default screening step. Individual member institutions may use detection tools as part of their own application review; the practice varies across the more than 900 member schools. Applicants and recommenders should assume that some level of detection is in use at any given school and should not rely on AI-generated content being undetected.

The Law School Admission Council (LSAC) issued a public statement on the use of generative-AI tools in the law-school admissions process, addressing both the applicant essay and the recommendation-letter components. The position, articulated in LSAC's policy resources, parallels the Common Application's: recommenders may use AI as a writing aid but the substantive content must originate with the recommender's own judgement.

The legal-education context has an additional dimension that law-school admissions offices and law-school faculty have emphasised: the legal profession itself is grappling with the appropriate use of generative-AI in practice, and the law-school admissions process functions as one of the early opportunities to set professional norms. Recommenders writing for law-school applicants carry a small but real weight in the formation of those norms, which is part of the reason the legal-education community has been more public than some other fields in articulating expectations.

For applicants, the LSAC dossier service handles the mechanical distribution of letters to participating law schools but does not screen letters for AI-generated content. The screening, if any, happens at the individual law school's admissions office. The applicant is not in a position to verify whether a given recommender has used AI inappropriately; the responsibility sits with the recommender. See the law school recommendation guide for the broader law-school framework.

The Association of American Medical Colleges has addressed generative-AI use in the medical-school application process (AMCAS) and in the medical-residency application process (ERAS). The AAMC's position on letter writers tracks the general convergent position: AI as an editorial assistant is permissible; AI as a generator of substantive content is not.

The medical-education context has a specific concern that the AAMC has emphasised in its guidance: clinical performance evaluation requires the letter writer's direct observation of the candidate's work in clinical or research settings, and AI-generated content cannot substitute for that observation. The ERAS letter of evaluation in particular relies on the letter writer's authority as a clinical or research preceptor; AI-generated letters from preceptors who did not directly observe the candidate's work undermine the structural premise of the residency-application process.

For applicants, the practical implication is that medical-school and residency applicants should ask their letter writers explicitly to write from their own observations and notes. The candidate's brief to the letter writer should include the specific cases, projects, or clinical experiences the writer observed, to support the writer in drafting from their own direct experience rather than from a general impression that could be padded with AI-generated content. See the medical school recommendation guide and the medical residency guide for the broader frameworks.

AI-detection tools (Turnitin's AI-detection feature integrated into its plagiarism detection product, GPTZero, Copyleaks AI Content Detector, Originality.ai, and others) produce classification confidence scores on submitted text. The published independent academic research on their reliability has identified several recurring limitations: meaningful false-positive rates (genuine human writing classified as AI-generated), meaningful false-negative rates (AI-generated text classified as human-written), and significant variation in accuracy across writing genres, language backgrounds, and revision patterns.

Published research from groups at Stanford, MIT, and several other universities has documented patterns of bias in detection tools, including elevated false-positive rates for writing produced by non-native English speakers and for writing in technical or formulaic genres. The 2023 paper by James Zou's group at Stanford on detector bias against non-native English writing is among the most-cited; subsequent work has explored both detection-tool improvements and the parallel adversarial work that aims to defeat detection.

The practical implication for admissions offices: detection is one input among several and should not be the decisive signal in adverse decisions. Most institutions using detection tools combine the scores with other factors (consistency between the applicant's writing style across components, alignment between the letter's content and the applicant's record, the recommender's verifiable identity and relationship to the candidate). The implication for recommenders: AI-generated content may be detected, but more importantly, even undetected AI-generated content tends to be generic and hurts the candidate's application by failing to provide the substantive evidence admissions committees need.

For recommenders who want to use AI tools responsibly, an honest co-drafting practice has emerged from the converging guidance. The practice starts with the recommender's own substantive notes: the specific observations about the candidate, the evaluative judgements the recommender wants to make, the comparative claims grounded in the recommender's broader experience, the concrete vignettes that anchor the letter. These notes can be sketchy (a bullet list of points, not full prose); they are the substantive content the letter will carry.

The recommender then drafts the letter from the notes, in their own words, at the substantive level. The draft does not need to be polished; it needs to contain the substantive observations the recommender wants to convey. At this stage, AI tools can be useful as editorial assistants: asking an AI to suggest structural improvements, to tighten paragraphs, to check the letter for flow, to suggest alternative phrasings for specific sentences. The AI is the polishing layer; the recommender remains the substantive author.

The practice that violates the convergent guidance: pasting the candidate's resume into a chatbot and asking for a recommendation, then submitting the output with minor edits. The substantive content in this practice originates with the AI rather than with the recommender, which is the line the major admissions services have drawn. Recommenders who find themselves tempted by the shortcut should consider whether the underlying issue is that they do not have substantive knowledge of the candidate to draw on; if so, the cleaner response is to decline the request. See the how to decline guide for the decline framework.

The phrase the writer's voice appears in the major admissions services' guidance and has become a shorthand for the substantive-content question. The writer's voice in the 2026 context does not mean the writer's specific prose style at the sentence level; it means the recommender's substantive judgement and personal observation of the candidate. A letter polished by AI editorial assistance can still carry the writer's voice if the substantive content originates with the writer; a letter generated by AI from minimal recommender input does not carry the writer's voice even if the recommender accepts it under their signature.

The shift in conception matters for how recommenders should think about the AI question. The pre-AI norm was that the recommender wrote every sentence themselves; the polishing assistance an AI now provides is structurally similar to the polishing assistance a colleague or a professional writing editor would have provided. The substantive responsibility, by contrast, has not changed: the recommender is responsible for the observations, judgements, comparative claims, and personal anchors that make the letter a substantive document. AI tools that operate in the editorial layer do not erode that responsibility; AI tools that operate in the substantive layer do.

For the broader recommendation-letter ecosystem, the convergent position has settled the question more cleanly than many observers expected three years ago. The conventions are reasonably clear, the major admissions services have aligned on a common position, and the practical co-drafting practice is workable for recommenders who want to use AI tools responsibly. Future guidance is likely to refine the boundaries rather than redraw them. See the conventions on related topics in the how to write and common mistakes guides.