How AI Is Changing Dossier Generation—and Where It Still Falls Short

Two things are true at the same time in this year's promotion cycle. At one university, a faculty member uploads a curriculum vitae into a faculty activity system and watches an artificial intelligence parser turn twenty pages of appointments, grants, and publications into structured records in under a minute. At another institution, a colleague receives written guidance saying ChatGPT may not be used anywhere in a tenure dossier, and that the same prohibition must be shared with external reviewers. The software and the rulebook are both new. They point in opposite directions.

The dossier is the least public document in an academic career and the one that decides its biggest questions. It holds the candidate's statements, the external letters, the teaching evidence, the service record. For decades it was assembled the slow way: CV entries typed by hand into whatever platform the institution had purchased, narratives drafted alone at a desk. Generative artificial intelligence (AI) is now compressing both ends of that process, writing and reading, before institutions have agreed on what either should look like.

The writing side: parsing, drafting, polishing

Start with the least glamorous part, because it's the part working now. Watermark's Faculty Success platform added AI Parsing to general availability in July 2026: a faculty member uploads a CV or pastes a list of citations, and the assistant maps entries into the institution's data structure for review. Watermark says the import takes about a tenth of the time of manual entry, that records are reviewable before saving, and that the data is never used to train models. Interfolio, now part of Elsevier, has been doing a quieter version of this for longer through its Faculty Activity Reporting and Review, Promotion and Tenure modules, which pull publication data into review cases and generate institution-formatted CVs from a central activity hub.

The drafting side is messier and more personal. Rather than waiting for a platform, many faculty turn to a general-purpose large language model (LLM) such as Claude or ChatGPT. A faculty writing coach at a large public research university, call him Dr. M, uploaded his eight-page CV into Claude and asked what his main contributions to his field were. The model returned five, each phrased more generously than anything he'd been willing to write about himself. His published account describes blushing with pride, then with shame at not having seen it himself. That is the real appeal for a pre-tenure academic. The hard part of a dossier is rarely the evidence; it's writing about yourself in the tone the genre demands while imposter syndrome does its work.

The base rate: most faculty are already experimenting. In Ithaka S+R's national survey of 2,654 instructors, 72 percent said they had used generative AI for at least one instructional purpose, though no single use had become standard and only 14 percent felt confident using the tools for teaching. Dedicated dossier use is rarer. Among biomedical researchers in a separate Ithaka survey, six in ten had tried generative AI in their research, but fewer than one in ten used it regularly. The exception, for now, is a small set of institutions that have built the assistant themselves rather than waiting for a vendor.

The reading side: summaries without consent

While candidates are learning to write with the tools, committees are quietly starting to read with them. In a faculty discussion this year, one professor asked colleagues how they would feel if told an AI system had summarized their promotion materials. The answers were unsparing: summaries 'belie values and priorities,' candidates whose impact can't be reduced to numbers would be 'flattened,' and opting out could feel punitive because refusing might invite extra scrutiny. The worry isn't hypothetical. One faculty union has already bargained a boundary into contract language. A 2025 memorandum of understanding between the University of Central Florida and the United Faculty of Florida states that generative AI 'shall not be used as the sole or primary reviewer' of promotion materials, that no AI-generated summary may be used in a supervisor's evaluation of a unit member, and that promotion decisions require 'direct and substantive human oversight' by designated officials.

That is a meaningful line, because summarization is the feature administrators reach for first. A 30-page dossier is expensive to read, and a machine can compress it in seconds. The problem is that a summary is an interpretation: which committee service counts, which teaching evidence matters, which numbers get left out. Those choices carry values, and faculty who have thought about it are not convinced a model should be making them.

What the tools still get wrong

The tools are most reliable on the mechanical layer and least reliable exactly where a dossier is judged. Hallucination is the best-documented failure. Guidance for faculty at the University of Central Florida puts it plainly: large language models still frequently hallucinate sources, DOIs, and publication details, which is a polite way of saying a model will invent a citation that looks real and isn't. In a promotion file, one fabricated DOI or one wrong award date reads as carelessness at best and dishonesty at worst.

Homogenization is the quieter problem. Reviewers describe reading batches of statements that share the same cadence and the same hedging adjectives, the tell of a widely shared prompt. When every candidate asks a model to 'highlight strengths,' the outputs converge, and the committee is left trying to distinguish people through a layer of machine polish that smooths away the exact idiosyncrasies tenure is meant to assess.

Then there's bias, and this one has the clearest data. In a single-blind study of promotion letters of recommendation, 32 academic physicians identified whether letters were human- or AI-written correctly 59.4 percent of the time, barely above chance, with a confidence interval from 58.6 to 60.2 percent. The same study found the AI letters leaned on more male-associated words while the human letters used more female-associated words, and letters believed to be human were rated higher in quality and persuasiveness. A separate study in orthopedic surgery found reviewers identified 40 percent of human letters and 63 percent of AI letters correctly. The tools don't only get facts wrong on occasion; they subtly shift who a candidate sounds like.

One institution's ban is another's assistant

Policies are splitting into camps, and the split is sharper than the usual slow diffusion of campus rules. Virginia Tech's promotion and tenure workshop slides for the 2025-26 cycle put the restrictive position in one sentence: 'Artificial Intelligence (AI) Prohibited (e.g., ChatGPT) in dossiers,' with AI barred from evaluating or summarizing dossiers and from drafting or generating reports, and the prohibition shared with external reviewers.

The permissive camp looks like the University of New England in Australia, whose Academic Promotions Assistant, built for ChatGPT-4.1, walks staff through workload confirmation, evidence collection, and narrative drafting for teaching, research, and service, and is in daily use. The vendor camp sits between them: Watermark's AI Parsing is optional at both the institutional and individual level, reviewable before anything is saved, and barred from training on faculty data.

Underneath the patchwork sits a structural lag. A 2026 study of journal editors found publication policies adapting faster than tenure systems, a governance misalignment in which journals demand disclosure while promotion committees still have no shared standard. That gap is where candidates get caught: what a journal permits, a tenure committee may prohibit, and silence on either side reads as a rule of its own.

What this means for your review file

None of this is an argument against using the tools. It's an argument for treating the dossier like a document whose audience has opinions about them.

Read the rulebook before you type. Promotion guidelines now differ on this point as much as they ever did on page counts. Search your faculty handbook for 'artificial intelligence,' 'generative AI,' or 'ChatGPT.' If the answer is silence, don't assume permission; ask your chair or faculty affairs office in writing and keep the reply.

Keep the machine on the input side. The lowest-risk uses don't touch your voice: parsing a CV into an activity system, checking a statement against the handbook's required sections, arranging six years of service into a chronology. Your CV is the raw material every parser reads, so make it current and complete before you let a tool at it.

Write first, prompt second. Draft the narrative yourself, however badly, then use a model to find the through-lines you underplayed. That ordering matters: the model should reflect your judgment back at you, not supply it.

Verify everything. Assume any citation, date, enrollment figure, or dollar amount the model produced is wrong until you've confirmed it against your own records:

• Citations and DOIs against the actual journal pages
• Dates, enrollment figures, and award amounts against your own files
• Award names and grant numbers against the original notices
• Page limits and required sections against the current handbook, not the version the model was fed

If you sit on a committee, don't summarize a candidate's file with AI unless the institution has told you to and the candidate has seen and agreed to the result. A summary is an interpretation, and interpretation is the one part of the process that has to stay human.

One concrete next step: open your institution's promotion and tenure guidelines today and search for the word 'artificial intelligence.' Whatever that sentence says, prohibition, disclosure, or silence, it sets the terms for every other choice in this article. If it says nothing, your first drafting task isn't a statement. It's an email to the person who owns the policy.

Photo by Woliul Hasan on Unsplash