AI in eLearning: What Changed This Term, and What It Means for Courses

The biggest change in AI in eLearning this term isn’t a new feature inside course platforms. It’s that the AI study tools moved into the general assistants learners already use, often paid for by their university or free for students. Anyone taking your course can now ask an AI to explain a lesson, quiz them on it, or do the exercise for them. The research says which of those they do matters more than whether the AI is there at all.

What happened

Three announcements in the past seven weeks show the shift.

A university gave everyone an AI assistant. On 14 September 2026, the University of Sydney announced that all its students and staff would get free, supported access to ChatGPT Edu, with features rolling out in stages and university data kept out of model training (University of Sydney, 2026). The university says its own AI platform, Cogniti, stays the preferred tool for teaching, and that students will still “prove their knowledge and capabilities in secure supervised assessments”.

General assistants became study tools. Google’s round-up of its September announcements, published on 2 October 2026, says Gemini Notebook now offers custom flashcards and several quiz formats, and that eligible college students can claim a free 12-month Google AI student plan (Google, 2026). A learner can upload your lesson notes and get practice questions back in seconds.

Workplace learning platforms built AI into the course. Docebo’s October 2026 release lists a Tutor Agent that lets learners ask questions grounded in course content, and a Reinforcement Agent that, after a supported course, “schedules personalized practice” (Docebo, 2026). Docebo is an enterprise learning platform rather than a creator platform, but the pattern, a tutor tied to the course plus scheduled follow-up practice, is where the larger platforms are heading.

Why it matters for people who build courses

For most of the last two years, the question was whether to add AI to a course. That question is now mostly settled by the learner. If they have a capable assistant open in another tab, your course is already being used alongside an AI tutor, whether you designed for it or not.

That changes what a course is competing with. Explanations are now cheap: a learner can get a patient, tailored explanation of almost anything. So can flashcards and practice quizzes. What’s harder to get from a general assistant is the rest of what a good course does: a sequence that builds a skill in the right order, practice designed for that skill, feedback on real work, and an assessment that shows what the learner can do.

What the evidence says

The research on AI tutors is young, but the two best-known trials point the same way: the AI’s design decides whether it helps or harms learning.

In a field experiment with nearly a thousand high school maths students in Turkey, Hamsa Bastani and colleagues gave some students a standard ChatGPT-style interface (“GPT Base”) and others a version with prompts designed to safeguard learning (“GPT Tutor”). While students had access, practice grades rose by 48 per cent with GPT Base and 127 per cent with GPT Tutor. The catch came afterwards (Bastani et al., 2025).

The authors report that students used the unguarded version as a “crutch” during practice, and that the safeguards in GPT Tutor largely removed the harm. Performance during practice, in other words, is not the same thing as learning.

The second trial shows the other side. In Harvard’s large introductory physics course for life sciences, Greg Kestin and colleagues compared lessons taught by a custom AI tutor with in-class active learning lessons, in a crossover trial with 194 students (Kestin et al., 2025). Students learned more with the AI tutor, in less time, with gains the authors describe as more than double those in class. But their tutor was built carefully: it guided students through each part of a problem in sequence, gave feedback on demand, and worked from expert-written step-by-step answers. As the authors put it:

Two limits matter. Both studies are single trials, in school maths and a university physics course, measured over days or weeks rather than months. And the wider evidence base is thin: a 2026 review by Cristina Barnard, Lily Fesler and Susanna Loeb found more than 800 relevant papers on AI in schools, but only 20 high-quality causal studies of its effects on students or educators (Barnard et al., 2026). We’d treat the direction of these findings as early evidence, not a settled rule.

What we’d do

If you build or run an online course, here’s how we’d respond this term.

1. Assume learners have an AI open. Write it into the course. Tell learners how to use an assistant well for your subject (ask for hints, not answers; explain your reasoning and ask it to check it) and when not to use it at all.
2. Keep some practice AI-free. The harm in the maths trial showed up only when access was taken away. Build in short closed-book checks, such as a five-question quiz at the start of the next module, so both you and the learner can see what they can do alone.
3. If you add a tutor, make it a tutor. Ground it in your course material, have it ask the learner to try first, and give hints before full solutions. A general chatbot dropped into a lesson is closer to GPT Base than GPT Tutor.
4. Spend your effort where AI is weakest. Sequence, practice design, feedback on real work and an assessment of the actual task are what a learner can’t easily generate for themselves. Our online course design checklist covers each of them, and the evaluation phase of the ADDIE model is where you’d test whether an AI feature helps.
5. Check your platform before you buy an add-on. Course platforms are adding AI features quickly, and most are aimed at building courses faster rather than helping learners learn. Ask what the feature does when a learner asks for the answer.

What to watch next

The useful signal over the next few months won’t be launch announcements. It will be studies that measure what learners can do weeks later, without the AI, and platforms that publish how their tutors decide when to give an answer. Until then, design as if the AI is in the room, and keep checking what learners can do without it.