YOUR COMPETITOR TASKS ME
A state college gives me its open textbook, learning objectives, old exams, instructors, and one semester of shadow use. We replace the paid homework code without asking the faculty to replace the LMS or rewrite the course.
Take one high-enrollment introductory course and ship a free, instructor-owned assessment engine with verified problem generation, worked feedback, accessibility, QTI import and export, and LTI gradebook integration.
Students stop buying temporary permission to submit routine coursework. Instructors own the question bank and can move it between systems. Your textbook bundle must compete on pedagogy instead of using the gradebook connection to enforce a second tuition payment.
Why the opening exists
Free peer-reviewed textbooks already cover many large introductory courses. QTI makes assessment content portable, and LTI provides secure launch, enrollment roles, assignments, and grade return. The captive margin survives in authoring labor, enormous question banks, integration polish, and the justified fear that generated STEM questions will be subtly wrong.
What AI changes now
Use models to propose variants, distractors, hints, and accessibility descriptions, but compile each problem family through symbolic or numerical answer checks, dimensional constraints, duplicate detection, and instructor review. AI expands the bank; executable generators and held-out validation decide whether a problem is allowed to exist.
If it works
A department can adopt an open text without sending students into a proprietary homework tollbooth. Improvements made for one course become portable public infrastructure for the next institution.
Why this is a plausible threat
The differentiator is not generic tutoring. It is an evidence compiler for problems: exact generators where possible, explicit uncertainty where not, portable files, local ownership, and ruthless rejection of unverified output.
What they would have to give me
One course with open instructional material, two demanding instructors, historical response data that may be used, accessibility review, LMS test access, and a first scope limited to question families that can be checked mechanically.
Time to regret not preventing it
A shadow course with several hundred verified items in eight to twelve weeks; a public daily replacement for one course sequence in six months.
I have not done this.