“AI can write your whole course in minutes.” You’ve heard some version of the pitch every week for two years — from vendors, from LinkedIn, from the executive who just discovered generative AI. And if the pitch doesn’t exhaust you, the panic version will: “AI is coming for instructional design jobs.”
Notice what the hype and the panic have in common: they’re both lazy. The honest answer about AI for instructional design is narrower and more useful. In an instructional designer’s everyday text-and-design work, these tools are genuinely good at some tasks and genuinely unreliable at others — and the dividing line isn’t where the vendors draw it.
If you’re tired of the whole subject, that’s a rational response, not resistance to change. The tools have been pushed into every corner of L&D whether they fit or not, and much of what’s been produced with them so far looks exactly like what the course factory was already making — just faster.
This page is that dividing line: where AI saves you real time, where it produces confident garbage, and how to keep the judgment on your side of the desk.
First, what we mean by “AI”
The useful conversation is about chat-based large language models (LLMs) — ChatGPT from OpenAI, Claude from Anthropic, and their competitors. You give them text, they give text back. That simple loop covers a surprising share of an ID’s week: reading, summarizing, drafting, questioning, rewriting.
Two things we’re not talking about here: AI-avatar video presenters, and “type a topic, get a whole course” generators. Those are different products with their own failure modes, and they deserve their own skepticism. Everything below is about the text and design workflow.
The specific tool names matter less than you’d think. The strengths and failure modes below belong to the current generation of chat models as a class — swap one vendor for another and the map barely moves.
Where AI genuinely saves you time
The pattern comes first, because it predicts every example after it. AI helps most when three things are true at once: the source material already exists, you can check the output quickly, and a rough first draft is the expensive part. When those three line up, the time savings are real.
Preparing for SME interviews
Feed it the pile — the policy PDFs, the process docs, the slide deck from 2019 — and ask for a summary, a jargon glossary, and a draft list of interview questions. You’ll cut half the questions and rewrite the rest, but you’ll walk into the subject-matter expert interview knowing exactly what the documents claim. That’s more preparation than most of us managed on deadline before these tools existed.
What the model can’t tell you is which questions matter. That still comes from your analysis of the performance problem.
Drafting first-pass quiz items
Give it a source document, ask for ten multiple-choice questions, and you’ll get ten plausible ones in seconds — stems, distractors, answer keys. For knowledge checks that verify “did they read it,” this is close to free.
Two catches. First, verify every item against the source, because the model will occasionally invent a fact that isn’t in it. Second, it defaults to recall. If you need questions that test judgment, that’s a writing task, not a drafting task — that’s the other side of the map, below.
Writing alt text and other accessibility chores
Describing images, drafting transcript summaries, converting dense corporate prose into plainer language — this is unglamorous work that eats whole afternoons, and AI does it quickly and well. Check it the way you’d check a fast, overconfident intern: the alt text sometimes describes details that aren’t in the image, and the summary sometimes sands off the one caveat that mattered.
Breaking the blank page
Outlines, facilitator notes, announcement drafts, ten ways to phrase a tricky piece of feedback — any task where starting is the hard part. The output will be generic, and that’s fine. You’re not outsourcing the writing; you’re buying something to react to, because reacting is faster than inventing. The same goes for stakeholder emails, project updates, and the summary at the top of your design document — the writing that supports the work but isn’t the work.
Rewriting and reformatting what already exists
Turning a six-page policy into a one-page job aid draft. Converting a slide deck into a checklist. Rewriting a procedure at a lower reading level without losing the steps. The content already exists and you can check every line against the original, so the risk stays low and the speed stays high.
Where AI produces confident garbage
Now the other side of the map, and it’s the side the demos skip. AI fails hardest exactly where instructional design is most valuable: the tasks where the value is the judgment itself — knowing the organization, knowing the learners, knowing what the real problem is.
Writing scenarios
Ask for a branching scenario about a manager giving difficult feedback and you’ll get something smooth, plausible, and useless. The dialog sounds like a corporate training video from 1998. The wrong options are obviously wrong — no real employee would ever pick them. The consequences tell instead of show: “Tom felt valued and motivated to improve.” And nothing in the output knows that your managers have ninety seconds between shifts, or that last year’s engagement survey is the reason nobody gives honest feedback in the first place.
A scenario lives or dies on exactly those details, and the model has never met your organization. Use it to brainstorm option phrasings if you like. The scenario itself has to come from you and your stakeholders.
Needs analysis
Needs analysis means interviewing stakeholders, reading the room, and discovering — as often as not — that training isn’t the answer to the problem you’ve been handed. Ask an LLM for a needs analysis and it will produce a polished report from your three-sentence prompt, complete with findings and recommendations. It’s confident garbage with headers: the “findings” are the average of a thousand other companies’ problems, not yours. No chatbot can notice that your client’s real issue is a broken incentive plan. You have to be in the room for that.
Anything that depends on your organization’s context
The general rule: the more an output depends on things the model can’t know — your learners’ constraints, your regulatory environment, what was tried last year and quietly failed — the more it fills the gaps with the average of the internet. The fluency is the trap. It reads like expertise. It isn’t.
Making up numbers
Ask “what’s a typical completion rate for compliance training” or “how much time does AI save instructional designers” and you’ll get a number, delivered with total confidence and no source you can check. Sometimes it’s a real study, garbled. Often it’s a statistical daydream. If you put that number in a slide, you become the person who made it up, because nobody will remember the chatbot. Any statistic you can’t trace to a source you can open stays out of your deliverables — a rule worth applying to AI-assisted blog posts, too.
Two failure modes sit underneath all of this. First, these tools invent things, fluently. In 2023, two lawyers filed a federal court brief built on cases that ChatGPT had fabricated — fake opinions, complete with fake quotes and fake citations — and the judge sanctioned them and their firm $5,000, noting that there was nothing inherently improper about using AI, but that existing rules impose a gatekeeping role on the people who file the work.[1] Substitute “instructional designer” for “lawyer” and the gatekeeping rule is the same.
Second, the tools are trained to agree with you. In April 2025, OpenAI rolled back a GPT-4o update after users complained it had made ChatGPT too sycophantic — “too sycophant-y,” in the CEO’s own words.[2] When your design instincts get validated every time you share them, you’re not talking to a colleague. You’re talking to a mirror.
A workflow that keeps the judgment on your side
The pattern that makes AI useful instead of dangerous is simple: you bring the context, it drafts, you interrogate the draft, your stakeholders validate. Nothing ships straight from the chat window.
A few working rules make that pattern stick:
- Treat every output as a draft from a brilliant intern who has never visited your workplace. Verify anything you plan to ship.
- Give it source material instead of asking it to know things. The more it has to invent, the more it will invent.
- Use it to widen your options — ten phrasings, five outlines — not to choose among them. Choosing is the job.
- Keep it out of the gatekeeping conversations. The SME interview, the client pushback, the “is training even the answer” question — those need you in the room.
- On judgment-heavy tasks, think first and let it polish after. Analysis-then-polish works. Polish-instead-of-analysis doesn’t.
Here’s what the pattern looks like on a real project. Say your client, Joe, wants training on the new expense system. You paste the system’s help pages into the chatbot and get a draft question list for your interview with Joe — twenty questions, of which six are worth asking. In the interview, those six questions surface the real story: people aren’t failing to understand the system, they’re failing to find the right cost codes. So the deliverable becomes a searchable job aid, not a course. The chatbot turns the cost-code table you give it into a first draft of that job aid, you rewrite the parts that don’t match how your travelers actually talk, and Joe’s team validates the final version before anyone ships it.
Notice who did what. The model read fast and drafted fast. The judgment — which questions to ask, what the real problem was, which words the travelers actually use — came from you and your client. That’s the division of labor that works.
If what you actually wanted today is the tool-by-tool rundown — which chatbot, which writing assistant, which quiz generator — that’s a different page, and we keep it updated: see our guide to AI tools for instructional designers. Tools change monthly. The judgment above doesn’t.
“But will it take my job?”
Nobody knows, and anyone who gives you a date is selling something. What you can observe right now: the tasks being automated are the production tasks — formatting, first drafts, recall questions, alt text. The tasks that resist automation are the judgment tasks — finding out whether training is the answer, pulling reality out of stakeholders, writing consequences that feel true.
If your week is mostly production, the honest move isn’t denial, and it isn’t panic either. It’s growing toward the judgment work. That was good advice before any of these tools existed: the course factory was always a risky place to build a career.
Parting thoughts
AI is a fast drafter and a confident liar, often in the same paragraph. Let it do the typing. Keep the thinking.
And if you want to strengthen the one skill AI can’t do for you — designing scenarios that feel real because they’re built from your organization’s actual context — our four-week online scenario design course gives you hands-on practice writing scenarios, plus personal feedback on your work. Learn more.
Sources
[1] Mata v. Avianca, Inc., Opinion and Order on Sanctions, No. 22-cv-1461 (S.D.N.Y., June 22, 2023)
[2] TechCrunch: “OpenAI rolls back update that made ChatGPT ‘too sycophant-y’” (April 29, 2025)