When refund requests rise in an AI course built around a respected professor, test the learning promise before blaming the professor or rebuilding the clone. The decisive question is whether students can complete the task they paid to learn, using guidance they can trust.
At 4:42 on a Friday afternoon in Accra, Kojo had seven refund emails open and a launch report due to his co-founder before dinner. Kojo is an illustrative composite: a course founder who still answers support tickets himself, usually with cold coffee beside his laptop and two browser windows full of lesson analytics.
One email stopped him. A student had asked the professor’s AI clone the same question twice and received two different recommendations. Her assignment was due that evening. If she submitted the wrong approach, she could fail the assessment. If Kojo refunded her immediately, he would still have six other requests and no explanation for why they had arrived together.
The professor’s name had sold the course. That Friday, the name could no longer carry the experience.
The refund emails pointed to a broken promise
Kojo first suspected the offer. Perhaps the sales page had promised too much. Students expected direct access to the professor and discovered a clone answering routine questions instead.
Then he suspected the professor. The recorded lectures were dense, and several examples assumed knowledge the students did not have. A respected expert can still teach at the wrong altitude.
Finally, he suspected the clone. It answered quickly, confidently, and sometimes differently when a student changed a few words. That pattern mattered. A student can work through a difficult lecture. Conflicting guidance creates a different problem: they no longer know which answer deserves trust.
Kojo could have rewritten the landing page, asked the professor to record new modules, or changed the model behind the clone. Each option would consume runway. None would prove which part had failed.
He closed the refund dashboard and opened the assignments.
Student work revealed where confidence collapsed
The useful evidence sat between the lesson and the refund. Kojo reviewed where each student stopped, what they had attempted, and what question they asked before leaving.
The pattern was narrower than the refund emails suggested. Students were completing the early lessons. They understood the professor’s examples. Trouble started when they had to apply the method to a new case and the clone supplied a polished answer without showing how it reached the recommendation.
That exposed the real promise of the course. Students had paid to develop judgment. The product was giving them conclusions.
The professor had not failed because his name carried less authority. The promise had failed at the moment authority needed to become a repeatable decision process. The clone made that gap harder to see because its language sounded certain.
This is the same risk behind conflicting answers from an AI faculty clone. Consistency alone would not solve Kojo’s problem, though. A clone can repeat the same weak answer every time.
Kojo changed the test before changing the product
With the weekend approaching, Kojo chose a smaller intervention. He took one assignment that had triggered three refund requests and wrote down the reasoning the professor expected students to demonstrate.
Then he turned that reasoning into checkpoints. What evidence supports the recommendation? Which assumption could change it? What would make the opposite choice sensible?
The clone would ask those questions before offering a conclusion. When evidence was missing, it would say so. When two paths remained defensible, it would explain the tradeoff instead of selecting one with false certainty.
Kojo also sent the seven students a plain message. He acknowledged that the course had given them an answer when they needed help reasoning through the assignment. He offered the revised exercise and kept the refund option open.
The bad ending was still possible. Students could decide that one broken assignment had damaged the course beyond repair. The professor could resist a teaching flow that appeared to reduce his expertise to checkpoints. Kojo could spend the weekend repairing the wrong layer.
He tested the revised exercise himself, then gave it to the student whose assignment was due. Her final submission did not copy the clone’s recommendation. She chose a different path and showed why.
That was the turn.
A famous teacher still needs a visible learning mechanism
A professor’s reputation can create attention and initial trust. It cannot substitute for evidence that the student is learning.
For an AI-course founder, rising refunds should trigger three separate checks. Compare the sales promise with the actual experience. Examine whether the teaching gives students enough context to act. Test whether the clone supports judgment or merely produces plausible conclusions.
Run those checks on one failed task before commissioning new videos or replacing the AI system. Watch a student attempt the work. Identify the exact point where they stop trusting their own decision. Then change that point and test it again.
On Monday morning, Kojo still had refund requests to process. The professor was still concerned about how the revised flow represented his method. Nothing had been wrapped up neatly.
But one student had submitted an argument she could defend, with the clone’s answer left unused in another tab. That gave Kojo a product decision grounded in student work: keep the professor, narrow the promise, and rebuild the clone around reasoning students could see.
Comments
No comments yet.