Human-led AI-driven learning is a topic I've been discussing with my team at KnowledgeNow eLearning Solutions since the start of generative AI in learning.

I've been building learning experiences for over a decade. I started my career in aviation training, then dove fully into higher education, and later moved into corporate learning. Today I lead a team that designs courses for colleges, universities, governments, and professional organizations across Canada. In all that time, I've never seen a technology move into our field as fast as generative AI.
Almost every conversation I have with a client now includes some version of the same question. If anyone can produce polished work with AI, what does a course, a grade, or a credential really mean?
Should we be using AI to build learning at all?
My answer is yes, through human-led AI-driven learning, where AI supports the work and people stay in charge of every decision.
AI-Driven Learning in Canada: Why Human Oversight Matters
AI is no longer something students might try. It's how many of them work. A KPMG Canada survey released this fall found that 75% of students use generative AI for schoolwork, and that 77% of educators say AI has changed how they assess learning outcomes. What struck me most was the gap in expectations: 72% of educators consider using generative AI for schoolwork to be cheating, compared with 50% of students. (KPMG Canada, September 2026)
When students and teachers disagree about what counts as honest work, trust suffers. In interviews for HEQCO's Consortium on Generative AI, one Ontario university leader described having students and faculty who "don't trust each other" (HEQCO, March 2026). The quick fix many institutions reached for, AI detection software, has created problems of its own.
The Canadian Alliance of Student Associations has warned that AI-powered plagiarism detection tools have been found to disproportionately misclassify the work of non-native English speakers (Canadian Press, August 2025), and the Ontario Undergraduate Student Alliance has asked the province to prohibit AI detection in academic integrity investigations altogether (OUSA, March 2026).
Institutions aren’t sure of the best methods - in some cases, asking students to go back to pen-and-paper methods of assignment writing.
Trust in credentials is also making headlines.
This spring, Canada's Auditor General reported that IRCC identified 800 study permits issued between 2018 and 2023 for which applicants had either used fraudulent documentation or misrepresented information (Office of the Auditor General, March 2026). Newer credentials are still earning their place. Future Skills Centre research found that businesses were more likely to recognize and accept microcredentials from institutions they deemed reputable. (Future Skills Centre)
Put all of that together and one lesson stands out. Trust doesn't come from the tool. It comes from the people and processes behind the learning.
We Didn't Ban the Calculator

When I hear people say AI has no place in education, lately, I’ve been thinking about calculators.
In April 1986, a group of math teachers protested outside the National Council of Teachers of Mathematics conference in Washington because the council had endorsed calculators in the classroom. One of their signs read "The Button's Nothin' Till the Brain's Trained." (Education Week, April 1986) Their fear was real: if students could press a button, would they ever learn to think mathematically?
Looking back, the protesters weren't entirely wrong, but the answer was never to ban the calculator. The answer was to teach students when to use it, and to make sure they could tell whether the answer on the screen made sense. A student who types the wrong numbers gets the wrong answer, and a calculator will never tell them so. Number sense, the human judgement that says "that can't be right," became more important, not less.
AI is our calculator moment, only bigger. Refusing it outright doesn't protect quality. It just means doing slowly by hand what could be done well and quickly, while our learners, clients, and competitors move ahead. But using AI without judgement is like accepting whatever the calculator shows. The value has never been in pressing the button. It's in knowing whether the result is right.
AI & Instructional Design: What Happens When the Checkpoints Go Missing
We've been here before. When the internet arrived in classrooms in the late 1990s and early 2000s, educators faced a flood of copied and pasted work. A 2001 article from the American Historical Association put it bluntly: "The web is saturated with prose just waiting to be lifted." (American Historical Association, May 2001) Rutgers researcher Donald McCabe surveyed more than 18,000 students in 2003 and found that 38 percent admitted to one or more instances of internet plagiarism, and that nearly half of those saw it as trivial or not cheating at all. (Badger Herald, September 2003)
As a designer, I still come across material with no sources, and plenty of copy and pasting throughout.
The problem was never the internet. It was content moving straight from source to final product with nobody asking whether it was accurate, original, or appropriate. Copy and paste skipped the thinking.
AI without checkpoints is the same problem, sped up.
It produces content that looks finished, sounds confident, and may be wrong. We've already seen what that costs in Canada. In 2025, Newfoundland and Labrador released its Education Accord, a 10-year roadmap for the province's schools and post-secondary system. The 418-page document took 18 months to complete, and educators soon found that at least 15 of the sources it cited were fabricated. (Plagiarism Today, September 2025) Making it more painful, one of the report's 110 recommendations called on the province to provide learners and educators with essential AI knowledge. (InfoDocket, September 2025)
The same lesson shows up outside education, and as someone who started my career in aviation training, this one stayed with me.
In November 2022, a B.C. man named Jake Moffatt booked a flight to Toronto with Air Canada to attend his grandmother's funeral, after the airline's website chatbot advised him he could pay full fare and apply for a bereavement fare later. That advice was wrong. Under the airline's bereavement travel policy, customers must request discounted bereavement fares before they travel. (CBS News) When the case reached B.C.'s Civil Resolution Tribunal, the tribunal found that the airline "did not take reasonable care to ensure its chatbot was accurate." (Fox Business) The tribunal member put the principle plainly: there was no reason the customer should know that one section of Air Canada's webpage is accurate and another is not. (Times Colonist, February 2024)
That principle applies directly to learning. A learner has no way of knowing which parts of a course a person checked and which an AI produced unchecked. To them, it's all the course, and they trust all of it.
I'm not sharing these examples to point fingers. Both came from organizations doing serious work with capable people. I share them because they show exactly where the risk sits. AI didn't fail on its own.
A checkpoint was missing: nobody opened every source, and confirmed the answer before it reached the person relying on it.
In learning design, the stakes are just as real. Without checkpoints, AI-assisted courses can end up with:
- Content that is confidently wrong, such as an outdated regulation, a made-up statistic, or a procedure that's no longer safe.
- Assessments that measure nothing, where quiz questions test recall of the AI's own phrasing instead of the competency the learner needs.
- Material that is culturally unsafe, where AI flattens or misrepresents knowledge that belongs to a community and was never its to share.
- Courses that aren't accessible, with alt text that describes the wrong thing or interactions that a screen reader can't navigate.
- Generic sameness, where every course sounds the same because nobody brought the client's voice, context, or learners into it.
Each of these is the modern version of copy and paste. They all have the same fix: a qualified person who stops, checks, and decides.
A Practical Model for Human-Led AI-Driven Learning
There's no single right way to bring AI into learning design. Every institution, team, and project is different, and any model should be shaped to fit its context. Still, when people ask me where to start, I suggest thinking in stages. A practical model could look something like this: five stages, where AI has a clear role in the middle and people own the decision at every step.
Stage 1: Set the Rules
Before any content is created, the people responsible for the learning decide whether AI will be used, for which tasks, and what's off limits. That might mean the client approves AI for research and drafting but not for final assessment language.
For Indigenous education, it means the community leads the decision about whether AI is appropriate at all and how its knowledge is handled.
Nothing yet. This stage is entirely about people.
A documented agreement on AI use, made with the client or institution before work begins.
Does everyone with a stake in this learning agree on how AI will be used?
Stage 2: Ground the Work
AI is only as good as what it's working from. In this stage, the team chooses the source material AI should rely on. That could include subject matter expert content, approved frameworks, institutional learning outcomes, and the client's own policies and standards.
Organize, summarize, and map the approved sources so the team can see what they have and where the gaps are.
A designer or subject matter expert confirms the sources are current, credible, and appropriate for this audience.
Is AI working from the right sources, and only those?
Stage 3: Draft with Clear Instructions
This is where AI earns its place. With clear prompts, it can produce first drafts of scripts, modules, performance tasks, rubrics, knowledge checks, and alt text in a fraction of the time. The quality of those drafts depends almost entirely on the quality of the instructions.
Draft content in the team's templates, following its style standards and flagging gaps instead of filling them.
The designer writes the prompt, sets the standards, and decides what must not change.
Did we tell AI exactly what good looks like?
Stage 4: Verify Everything
Every draft gets checked before it moves forward. Someone opens every source and confirms it says what the draft claims. Subject matter experts review accuracy. Cultural advisors review anything involving community knowledge. An accessibility specialist reviews the build against WCAG 2.1 AA.
Run consistency checks, compare drafts against learning outcomes, and suggest weaknesses for people to look into.
Qualified reviewers confirm the content is true, safe, respectful, and accessible.
Would this hold up if a learner, an expert, or an auditor checked it line by line?
Stage 5: Decide and Sign Off
The final stage belongs entirely to people. The design team and the client agree on what "competent" looks like and test rubrics against real examples of learner work. Someone goes through the course as a learner would. AI use is documented at handover so the client knows exactly how the course was built.
Nothing. Sign-off is a human responsibility.
Final approval by the people accountable for the learning.
The question to ask: Would we stand behind this in front of a learner, an employer, or an auditor?
Why the First and Last Stages Matter Most
You'll notice AI has no role in the first or last stage. That's the point. People decide whether AI belongs in the work, and people decide whether the result is good enough to reach a learner. AI does its best work in between, where it saves real time without taking over judgement.
If you're just getting started, you don't have to set up all five stages at once. Begin with Stage 4. Making verification non-negotiable, so that every source gets opened and every draft gets reviewed, would have prevented every example I described above.
Where AI in Instructional Design Earns Its Place
Within a model like this, AI takes on the work that used to slow learning design teams down: Research and environmental scans. AI helps gather and organize existing frameworks, labour market information, and sector standards, always with direct links so a person can check every source.
Competency frameworks and assessments
For competency-based projects, AI helps draft framework structures, performance tasks, and rubric language. Designers then shape, challenge, and refine them with the client's experts.
Scripting
Raw subject matter expert content becomes structured scripts in course templates, with the layout tags developers use to build in Articulate Rise or an institution's LMS.
Production support
AI drafts alt text and knowledge check questions, and checks for consistency across a whole program of courses.
The result is that designers spend far less time on blank pages and reformatting, and far more on the decisions that make a course work.
Instructional Design AI Prompting for Quality: What to Ask

Rob Clayton of KPMG Canada said it well:
"Proper AI use is about more than knowing how to prompt." This is true, but good prompting is still where quality starts. Most poor AI output comes from vague requests. Here are some of the principles I recommend, along with examples of the kinds of instructions that make a difference.
Give it your standards, not just your topic.
AI doesn't know your template, your style rules, or what it isn't allowed to change unless you tell it.
Format this content into our Rise script template. Preserve the subject matter expert's wording exactly. Do not add, remove, or rephrase content beyond formatting. Use Title Case for headings and do not use em dashes.
Ask it to show gaps, not fill them.
AI's most dangerous habit is inventing things that sound plausible. Tell it to make gaps obvious.
Where information is missing or unclear, do not invent content. Insert a placeholder in red text describing what is needed so our team can follow up with the client.
Make every fact checkable.
This one prompt would have caught every fabricated citation I described above.
For every statistic or claim, include a direct link to the original source. If you cannot find a source, tell me rather than estimating.
Be specific about the output.
Precise requests save rounds of revision.
Write two multiple-choice knowledge check questions at the end of each topic, labelled with the topic they apply to. Include one correct answer and three plausible distractors, with brief feedback for each option. Draft alt text for this image in under 150 characters. Describe what the learner needs to understand from the image, not every visual detail.
Ask it to challenge you.
AI is a useful critic when you ask it to be one.
Review this rubric against the learning outcome. Identify any criteria that measure something other than the outcome, any levels that are hard to tell apart, and anything a learner could meet without demonstrating the skill.
Work section by section.
Build long documents one section at a time and review each before moving on. It feels slower but is faster overall, because problems get caught while they're small.
Efficiency Is the Point, and So Is Quality
I won't pretend AI isn't about efficiency. It is, and that matters to our clients, many of whom are being asked to do more with less.
But efficiency only counts if quality holds. The time AI saves should go back into the work that builds trust: more time with subject matter experts, more careful assessment design, deeper accessibility review, and more honest conversations with clients about what their learners really need.
There is no “one size fits all” solution. The use of AI is a spectrum - not simply a yes or no. The one thing I am 100% sure of is that the quality of output is dependent on the professionals and how they chose to use such a powerful tool.
Get in touch
Want to talk about your next course? Contact us and we will get back to you.
