Coaches and consultants building AI "twins" of themselves keep running into the same problem, and it never shows up at launch. The build handles questions cleanly through the first month while the client stays happy and everyone moves on to other things. Then, somewhere around month four or five, the answers have quietly stopped sounding like the person who built it.
Aiyou, a company that builds AI twins specifically for coaches, has a name for this: methodology drift. It's common enough that they've built a specific fix for it, and the fix says more about the actual problem than the name does.
What actually breaks
The build itself is fast. Feed a model a book, a set of session transcripts, a framework, and it can hold a passable conversation in your voice within days. What breaks later is a little more subtle than just a wrong answer. Aiyou describes the model drifting toward confident, generalized responses, because that's what users respond well to. However, the expert's real value rarely comes from a generalized answer. The expert's value comes from the exceptions, especially when the framework or rules need bending for a special case. That's the part that erodes first. It erodes one interaction at a time with no single conversation looking wrong enough to flag on its own.
In one case Aiyou describes, they tested a coach's AI clone against her own ten-element framework, session by session, and found the model's version had quietly diverged from hers.
Why it happens
This isn't a bug in one company's product. It's a property of how these models generate answers. Confidence reads as competence to most users, so a model under no correction pressure keeps producing confident answers whether or not the judgment behind them still holds true. A human expert catches themselves being wrong mid-sentence, because they built the exceptions in the first place. A model has no equivalent internal check. It only has whatever check someone builds around it. Thirty days of use won't show the drift. However, six months might. Volume and time are the friend of drift.
The design problem, misdiagnosed
The instinct so far has been to treat this as an engineering fix. It needs better prompting. It needs better retrieval mechanisms. Or, it needs a review process before launch. These are not wrong. It just stops short. Instructional design has run quality assurance against this question for years and under a different name: does this version of the material still teach what the original taught, or has repetition and simplification worn the edges off?
A curriculum that gets copied, adapted, and handed between instructors drifts for the same reason an AI clone drifts. The confident, simplified version of an idea is easier to produce and easier to consume than the original one. Instructional design's answer was never just to check the output after the fact. It was to build a recurring audit into the process from the start. And to build the audit with the person who owns the original methodology. Do this before the drift starts to happen, before the visible wrongs start appearing.
Nobody currently building AI clones for coaches and consultants has connected these two disciplines yet. The AI-clone industry is treating methodology drift like a problem it just discovered. Instructional design solved a structurally identical one decades ago.
The audit isn't a later step. It's part of the build.
If you're turning a book, a framework, or a curriculum into anything that answers people dynamically instead of saying the same fixed thing to everyone, decide before launch what the recurring check looks like: who reviews outputs, how often, against which version of the original material, and what happens when the model's answer and your answer diverge. Treat the first thirty days of confident-looking performance as exactly what it is, too early to tell anything, because this failure mode doesn't move that fast. Static content never had this problem, because static content can't drift. The moment your content starts answering people back, something has to hold that stability in its place. Build the audit in before you build the clone.