Most AI failures in training design are not technical. They are editorial.
The same mistakes appear whenever AI enters the design workflow:
- Starting with the course, not the performance gap. A model will happily write modules for a problem nobody diagnosed.
- Rewarding speed. Drafting in minutes is only valuable if review still happens.
- Accepting fluent prose as sound structure. Confident language hides weak sequencing and missing practice.
- Removing the subject expert too early. AI summarises what experts know; it cannot tell you what they notice.
- No quality bar. If you cannot define good, you cannot judge generated output either.
- Leaving data, copyright and confidentiality rules to chance.
Trainers notice this first, because facilitation depends on content they trust. Instructional designers need explicit review criteria. L&D managers must protect review time in project plans, even when stakeholders now expect faster turnarounds. HR leaders need a policy before the pilot, not after it.
The question is not whether AI can write the course. It is who signs off on it, and against what standard.