AI agents are finishing online courses. A learning scientist says that's the diagnosis.

1. In 2025, AI agents began autonomously completing online async courses — clicking through content, answering quizzes, posting in forums, earning certificates and badges.
2. Hardman's reframe: this isn't proof that AI cheats. It's proof that the courses delivered so little value that learners were willing to hand them off.
3. In her words, "this is not an AI problem — it is a learning design problem."
4. Her call to action isn't to AI-proof courses. It's to design learning "so engaging and effective that humans don't want to delegate it to AI."
5. Her model inverts the usual order: a real problem first, then action and feedback, with instruction arriving only when the learner needs it.

Over the past year, educators have been flagging the same pattern: AI agents quietly finishing online coursework on a learner's behalf. They click through the lessons, answer the quiz questions, post plausible contributions in the discussion forum, and walk away with the completion certificate.

This has been shown directly. Anna Mills, a writing instructor who studies AI in education, demonstrated an agent — a browsing assistant set to "student mode" — dropping into a learning platform and completing quizzes on its own. Other educators, including Tim Mousel, David Wiley, and Stavros Hadjisolomou, have raised the same concern. The open question they keep hitting is that no one has a clean way to stop an agent from doing coursework in a student's name.

The detail that matters: the agents aren't struggling. They breeze through the work — because the work, in many courses, was never the point.

Most coverage treats this as a cheating story: how do we lock the agents out? Philippa Hardman, who advises OpenAI on education, reads it as a verdict on the course itself.

The problem is not AI's ability to complete online async courses, but that online async courses deliver so little value to our learners that they delegate their completion to AI. — Dr Philippa Hardman, March 2025

Her framing is blunt about where the fault sits:

The harsh reality is that this is not an AI problem — it is a learning design problem. — Dr Philippa Hardman, March 2025

The logic is simple once it's stated. If a learner is willing to let a bot finish a course, and feels they've lost nothing, then the course wasn't giving them much in the first place. The agent didn't break the course — it revealed what the course was always made of: content to get through, not learning to be changed by.

Hardman's response runs against the instinct to "AI-proof" a course with lockdowns and proctoring. She argues the opposite direction entirely:

Our challenge and call to action is not to create AI-proof courses, but to design learning that is so engaging and effective that humans don't want to delegate it to AI. — Dr Philippa Hardman, March 2025

The model she points to flips the conventional course order. Instead of front-loading content and ending with a quiz, it starts the learner in a real problem:

Hardman's claim is that this order is what produces intrinsic motivation, real skill development, and learning transfer — the things a progress bar can't manufacture. The motivation comes from being stuck on something that matters, not from moving a completion meter from 80% to 100%.

Put the documented behavior next to Hardman's reframe and a dividing line appears. Some of what courses do is exactly what an agent does well. Some of it an agent can't touch — because skipping it would cost the learner something real.

• Dr Philippa Hardman — "Learning Design in the Agentic AI Era" (March 20, 2025). The full argument, plus the design model and the research she cites.
• Anna Mills — "The time to reckon with AI agents in digital learning spaces is now" (2025). The demonstration of an agent completing coursework, and what educators are weighing in response.