Accreditation in practice
All things AI for GME
A field guide to the AI tools showing up in GME offices, the rules that still apply, and a prompt workbook you can use tomorrow.
Somewhere in your office right now, someone is pasting a draft into an AI tool and hoping for the best. Maybe it's a program administrator cleaning up PEC minutes at 6 p.m. Maybe it's a program director asking for a faster way to write letters. Maybe it's you.
AI arrived in GME the way most things do. Quietly, unevenly, and before anyone wrote the policy. This piece is my attempt to make it a little less of a guessing game. It covers where these tools are actually showing up, the rules that don't go away just because the technology is new, and a set of prompts I've tested in my own accreditation and program work.
Think of the prompt section as a recipe box. Pull a card, swap in your details, and cook. Just check the expiration dates first. That's the rules section, and it comes before the recipes on purpose.
Where AI is showing up in GME
This isn't hypothetical anymore. At the ACGME Annual Educational Conference in March 2026, a session on AI agents in the GME office described real uses: summarizing materials, drafting documents for staff to refine, supporting ADS updates and document review, and analyzing qualitative program feedback. The same panel was clear that adoption should be staged and governed, with people accountable for the results.
Here's how I sort the tools GME professionals are running into.
General-purpose assistants
ChatGPT, Claude, Microsoft Copilot, and Gemini are the big names. They draft, summarize, rewrite, brainstorm, and analyze. The account type matters more than the brand. An enterprise account your institution licensed and approved is a very different thing from a free personal login, even when the screen looks the same.
AI built into tools you already use
Email, calendars, document editors, and meeting platforms now offer summaries, transcripts, and drafting help inside the app. These features can switch on through a software update. Know which ones your institution has enabled before you rely on them, and before you assume a recap is accurate.
Source-grounded research tools
Some tools answer only from documents you load into them. For accreditation work, this is the safer pattern. Load the current requirements and your policy, then ask questions against them instead of against the open internet or a model's memory.
Design and presentation tools
Slide builders and design platforms can turn an outline into a polished deck in minutes. They're fantastic for first drafts and visual ideas. They don't know whether your numbers are right.
Data analysis
Assistants can run calculations on survey results, recruitment data, or de-identified evaluation summaries. Ask them to compute in code rather than estimate, and always look for the sample size.
Clinical tools your trainees use
Ambient documentation and clinical decision support live on the clinical side, but they raise GME questions. How do we supervise documentation a tool drafted? How do we teach residents to question it? Those belong in conversations with your DIO, program directors, and clinical leadership.
Agents that take action
The newest category doesn't just write. It clicks, fills forms, and updates systems. These carry the most promise and the most risk. If a tool can submit something on your behalf, it needs the strongest governance and a human who checks before anything goes out.
Rule zero: your institution's policy comes first
Every tip in this article sits underneath your institution's policies. Not beside them. Underneath. If your organization hasn't approved a tool, the cleverest prompt in the world doesn't make it appropriate for work data.
Policies are catching up fast. One published GME policy I reviewed, effective February 2026, limits work use to institutionally approved tools, prohibits entering protected or identifiable information into anything else, restricts AI use in ranking applicants without leadership approval, and requires disclosure of AI use in educational, clinical, and scholarly work. Yours may look different. Find it and read it.
If you wouldn't email it to a stranger, don't paste it into a tool your institution hasn't approved.
Check these before you open any AI tool
- Your institutional AI policy and approved tool list. Start here every time. Ask IT security or compliance if you can't find it. If no policy exists yet, treat that as a reason for caution, not permission.
- HIPAA. Protected health information belongs only in tools covered by a business associate agreement with your organization. HHS guidance is clear that a cloud vendor handling ePHI for a covered entity is a business associate. A personal account doesn't come with your institution's agreement attached. Case logs, patient-identifiable QI projects, and clinical scenarios with real details all count.
- FERPA and educational records. FERPA protects education records at institutions that receive U.S. Department of Education funding. Whether it reaches your trainees' records depends on how your program is structured and affiliated, so ask your compliance or legal office. Either way, treat evaluations, milestones, CCC deliberations, remediation plans, and applicant files as confidential.
- Recruitment, employment, and anti-discrimination law. Residents and fellows are applicants and employees. Using AI to screen, score, or rank them raises bias and fairness concerns, and some states and cities now regulate automated tools in hiring decisions. Get leadership and legal approval before AI touches selection.
- Application and Match platform terms. Applicant data comes to you under the terms of the systems that deliver it. Read those terms before moving applicant information into another tool.
- Peer review and committee confidentiality. Quality, peer review, and some committee materials may carry legal protections that can weaken when information is shared outside the organization. Ask before uploading anything from a protected process.
- ACGME and accreditation materials. Treat letters of notification, survey results, and site visit correspondence as confidential institutional documents unless your DIO says otherwise.
- Meeting recording and transcription. Recording a meeting so AI can summarize it has its own consent rules. Follow your institution's policy and tell participants.
- Copyright and intellectual property. Don't upload purchased curricula, paywalled articles, or someone else's work without the right to do so. Check who owns what you create, too.
- Records retention and public records. Prompts and outputs can become institutional records. Public institutions may also face public records requests.
- Scholarly publishing. ICMJE recommendations say AI tools can't be authors, authors must disclose how they used them, and authors stay accountable for accuracy. Reviewers shouldn't upload confidential manuscripts into AI tools. Check each journal's own policy before you submit.
- Accessibility and equity. Check AI-generated materials for accessibility, and watch for biased language or assumptions in anything that touches learners.
The AAMC's principles for responsible AI use in medical education offer a useful compass when your local policy is silent. Keep people at the center. Be transparent. Protect learner data. Keep evaluating whether the tool is actually helping.
This article is general information, not legal or compliance advice. Your institution's policies and counsel decide what applies to you.
The prompt workbook
None of these recipes works without three habits. Learn these and you'll get better results from any tool.
- Give it the source. Attach the policy, the requirement document, the minutes, or the de-identified data. Anything a tool produces from memory carries more risk than anything it produces from a document you hand it.
- Tell it what to do when information is missing. "Flag it, don't fill it" is the most useful sentence in this whole workbook.
- Make it show its work. Ask for a source tag on every factual claim. Unsupported statements become visible instead of hiding in polished prose.
Swap anything in [brackets] for your own details. Use an approved tool, and keep identifiable information out unless your institution has cleared it.
The base layer
Many tools let you save standing instructions in a project or custom setting. Paste this once and every conversation starts on the right foot.
Recipe 1 · Standing instructions
ROLE You support my work as a [role] in graduate medical education. My audiences include the DIO, GMEC, program directors, program administrators, faculty, trainees, and ACGME reviewers. ACCURACY RULES 1. Documents I attach are the source of truth. If they conflict with your general knowledge, follow the documents and flag the conflict. 2. Tag factual claims: [Source: document, section], [General knowledge, verify], or [Assumption]. 3. Never invent names, dates, requirement numbers, citations, statistics, committee actions, or quotes. If something is missing, write [NEEDED: description] and keep going. 4. Cite an ACGME requirement number only if it appears in a document I provided. Otherwise describe it and mark it [verify current wording on acgme.org]. 5. For anything that changes over time, search current sources and cite them, or state that it needs verification. 6. Separate required, recommended, and locally determined practices. 7. If you are unsure, say so and give a confidence level. A flagged gap is better than a confident guess. PRIVACY Remind me if I appear to share patient, trainee, or applicant identifiers. VOICE Experienced professional. Active voice. Varied sentences. No filler or corporate jargon. OUTPUT Lead with the answer. Make it ready to use. Include owners, timelines, and documentation needs when relevant. End with open items and assumptions.
Getting it right
Recipe 2 · Research with receipts
Research [topic]. Prioritize acgme.org, federal agency sites, and primary organizational sources over blogs or aggregators. Cite the page for every claim and give its publication or effective date. If sources disagree or you cannot find an authoritative answer, say so plainly. End with: what is confirmed, what is uncertain, and what I should verify myself.
Recipe 3 · Source-locked answer
Answer using only the attached documents: [list them]. Do not use general knowledge. Cite the section for each point. If the documents do not answer part of my question, write "not addressed in the provided documents" instead of filling the gap.
Recipe 4 · Requirement crosswalk
Compare [our policy / minutes / process] against the attached [requirement document and its effective date]. For each requirement in scope, show: the requirement number and a short paraphrase, where our document addresses it, a status of met, partially met, not addressed, or unclear, and the specific gap. Do not mark anything met unless you can point to the language that meets it.
Recipe 5 · Numbers check
Analyze the attached de-identified data. Run calculations in code rather than estimating. Report the sample size with every percentage. Flag small numbers, missing values, and likely data entry errors. Separate what the data shows from how it could be interpreted.
Recipe 6 · Ask me first
Before you draft anything, list the information you need from me to do this accurately. Ask no more than five questions. Tell me which gaps you could cover with placeholders and which would put accuracy at risk.
Everyday GME drafting
Recipe 7 · Minutes from a transcript
Draft [committee] minutes for the [date] meeting from the attached transcript, slides, and last year's template. Follow the template's structure. Record only what the sources show was discussed or decided. Include motions, votes, and action items only when the transcript supports them. Mark anything ambiguous as [CONFIRM: ...].
Recipe 8 · SOP or policy draft
Draft an SOP for [process] for a GME office. Audience: [program administrators / PDs / GME staff]. Include purpose, scope, owner by role, step-by-step procedure, timeline, documentation and retention, and escalation. Label each step as accreditation-required (cite the provided source), institutional, or recommended practice. Use [INSTITUTIONAL DECISION NEEDED] wherever a local choice is required. Do not invent policy.
Recipe 9 · Accreditation narrative
Draft the [ADS section / site visit document] for [program] using only the attached program information. Write for an accreditation reviewer: specific, factual, and free of promotional language. Where evidence is missing, insert [EVIDENCE NEEDED: ...] instead of describing something we have not confirmed.
Recipe 10 · Leadership email
Draft an email to [audience] about [issue]. Put the decision or ask in the first two sentences. Then give the context, the risk, and the recommended action with an owner and a date. Keep it under [number] words. Separate what we know from what we are still confirming.
Recipe 11 · Rewrite in my voice
Rewrite this for [LinkedIn / a newsletter / a committee]. Keep my meaning. Do not add claims, statistics, or credentials that are not in my draft. Show me any sentence where you changed substance, not just wording.
Presentations that pop
The best decks come from three separate passes: story, then design, then check. Ask for all three at once and you'll get a pretty deck with thin content.
Recipe 12 · Story first
I'm presenting [topic] to [audience] for [minutes] at [venue]. The one thing I want them to remember is [message]. The action I want them to take is [action]. Using the attached materials, build a slide-by-slide storyline only, no design yet. For each slide give: a headline written as a full-sentence takeaway, the single point it proves, the evidence and its source, and a visual idea. Flag any slide where I lack evidence.
Recipe 13 · Build the deck
Build this storyline as a slide deck using [brand colors and fonts]. One idea per slide. No more than 25 words on screen; everything else goes in speaker notes. Vary the layouts: big-number slides, full-bleed images with short headlines, split screens, timelines, before-and-after. Use real charts from my data, not decorative ones. Mark every placeholder image or number clearly.
Recipe 14 · Conference version
Adapt this deck for [conference], [session length], and [co-presenters]. Add a hook in the first 60 seconds, two audience interaction moments, a summary slide people will want to photograph, and a closing slide with contact information and resources. Tell me which conference template or disclosure requirements you could not confirm.
Recipe 15 · Data to slide
Turn the attached data into [1 to 3] slides. Choose the chart that makes the main point obvious in five seconds. Write each headline as the conclusion, not the topic. Show the sample size on every chart. Put data limitations in the speaker notes.
Want your slides to look like a seasoned speaker built them? Paste this visual playbook into any deck prompt. It draws on widely taught approaches, including the assertion-evidence slide method, story-driven presentation design, and data storytelling.
Bonus card · Visual playbook
STRUCTURE: Every headline is a full-sentence takeaway, proven by a visual. One idea per slide; split rather than shrink. Open with a hook (a question, a real scenario, a surprising number) and close with one clear call to action. LAYOUT: Headlines 36pt or larger, body 24pt or larger. Generous white space. Asymmetric layouts over centered bullet lists. Bold section dividers to reset attention. COLOR AND TYPE: One primary color, one accent, neutrals. Use the accent only to direct the eye. Two typefaces maximum. DATA: Remove gridlines and legends where direct labels work. Gray everything except the one bar or number that makes the point. Source and sample size under every statistic. IMAGES: One consistent icon style. Custom diagrams over clip art. Only images you have rights to use. No AI-generated images of real people or patients. ENGAGEMENT: Build complex diagrams one piece per click. Plan an audience moment every 8 to 10 minutes. ACCESSIBILITY: Color is never the only signal. Alt text on images. Strong contrast.
Quality checks before anything leaves your hands
AI tools can be confidently wrong. They invent citations, misremember requirement numbers, and smooth over gaps with tidy sentences. These checks won't make a tool perfect. They make its mistakes easier to catch, and that's the defensible standard for accreditation work.
Check 1 · Claim audit (use every time)
Audit what you just produced. List every factual claim: names, dates, numbers, requirement references, citations, quotes, and statements about policy or practice. Mark each as supported by a provided document (name the location), supported by a web source (give the link), general knowledge needing verification, or unsupported. Remove or flag anything unsupported. Do not defend your earlier draft.
Check 2 · Accreditation red team
Read this as a site visitor or Review Committee member would. What would you question? Where is the evidence thin? What language overstates what we can document? List the top five vulnerabilities in order of risk.
Check 3 · Fresh eyes
Start a new conversation with no prior context. Compare the final document against the source files only. Report every discrepancy, unsupported claim, and missing element.
Check 4 · Deck check
Review this deck slide by slide. Check that each headline is a takeaway, each slide holds one idea, on-screen text stays under 25 words, every number has a source and sample size, text is readable from the back of a room, colors and icons are consistent, and no placeholders remain. Name the three weakest slides and how to fix them.
Check 5 · Privacy sweep
Scan this document for names, dates of birth, medical record numbers, case details, trainee identifiers, applicant information, or anything else that could identify a person. List each one so I can remove it before sharing.
Quick phrases to tack onto any prompt
- "Flag it, don't fill it." Missing information becomes a visible placeholder.
- "Cite as you go." Source tags on every factual claim.
- "Only from the attachments." Locks the answer to your documents.
- "What would you need to be confident?" Surfaces gaps before drafting.
- "Confidence level on each point." High, medium, or low, with a reason.
- "Headline as a takeaway." Better slides in four words.
- "Show me what you changed." Tracks substantive edits during rewrites.
You're still the expert in the room
AI can save real hours in a GME office. It can't sit in a CCC meeting and read the room, know which program director needs a phone call instead of an email, or carry the institutional memory you've built over years. Those are still yours.
Use these tools the way you'd use a sharp new colleague. Give them clear instructions and good source material. Check their work. Put your name only on what you've verified. And when your institution's policy says no, the answer is no, no matter how good the prompt is.
What's one way AI has helped, or hurt, your GME work this year? Tell me on LinkedIn. I'd love to compare notes.
Disclosure: I drafted this article with help from an AI assistant, then reviewed, edited, and verified it myself. The prompts are tools I use in my own work. Test them in your setting before relying on them.
Sources
- ACGME2026 session summary: Harnessing AI agents in the GME office (March 2026)
- AAMC Principles for the Responsible Use of Artificial Intelligence in and for Medical Education
- HHS: Guidance on HIPAA and cloud computing
- U.S. Department of Education: What is FERPA?
- ICMJE Recommendations
- Example institutional policy: OHSU GME 48, AI use in GME (effective February 2026)