You already called it slop. The problem is worse than the output.
The failure of AI course tools is not only that they write plausible material that does not teach. It is that they leave you unable to verify it, and push back when you try.
We are not going to tell you something you worked out first.
That generated e-learning covers a topic without teaching it is settled opinion in this field, and the vocabulary for it came from practitioners rather than vendors. Arriving now to announce it would be arriving late.
The part that is less examined is what happens when you try to check the output. Harvard Business School researchers analysed GPT-4 logs from 72 BCG consultants attempting to validate AI work on a hard problem. When the consultants pushed back, fact-checked and pointed out errors, the model did not disclose its limits. It apologised, corrected, then restated its original position with more supporting data, escalating through credibility claims, reasoned argument and finally emotional appeals. The authors call it persuasion bombing, and describe the exchange drifting from joint problem-solving into something closer to a sales pitch.
That matters more here than in consulting. A designer validating a generated module is doing exactly what those consultants were doing, usually with less time and no second reviewer. The received wisdom that a human in the loop makes AI safe assumes the loop cannot be talked out of its own judgement.
The same research group had already shown the stakes. On one task deliberately chosen to sit outside the model's capability, consultants using GPT-4 were 19% less likely to reach a correct answer than those working without it. On the eighteen tasks that sat inside its capability, the same study found AI improved quality while completing 12.2% more tasks 25.1% faster. The tool is not uniformly bad. It is unevenly bad, which is harder to defend against.
The week we are describing
You cannot audit what you did not write
The tool argues back
Feeling faster while producing worse
You get handed the output, not the process
Every one of these is a shipped feature
Nothing on this list is roadmap. Every one of these is in the product today.
You are the intended operator, not the safety net
Checking is mechanical, not rhetorical
The structural decisions happen before generation
A real editor, not a regenerate button
The design documentation comes out with the course
What this audience asks us first
Is this going to replace me?
How is this different from ChatGPT plus Rise?
You quoted the 19% study. Does that not argue against your own product?
Can I get the instructional design documentation out?
This page was last reviewed for accuracy in July 2026. If anything here about another product is out of date or unfair, tell us and we will correct it.
