AI can produce some remarkably convincing conversations.
That makes large language models (LLMs) an obvious tool for training simulations. If you want someone to practice a job interview, difficult conversation, or another interpersonal skill, why not simply give an LLM a role and tell it to respond like another person?
The problem is that a convincing conversation is not necessarily good training.
Sounding Human Is Only Part of the Job
Last year Google ran an ad called “Dream Job” during the Super Bowl to promote Gemini.
It was a well done piece, featuring a father using Gemini to prepare for a big job interview while also raising a child. It pulled on the heartstrings and was very persuasive.
Shame it was a complete fiction.
LLMs are not great training partners because they aren’t designed to be.
AI is built to keep conversations flowing. It flatters us, softens its feedback with flattery, and encourages us whenever possible. That’s great when I’m using it to rewrite an email, but what about for training?
Our conclusion, back when we started SkillSaige, was that coddling people was doing them a disservice.
Training Needs to be Realistic
Chatbots are too nice.
Imagine a learner practicing how to de-escalate with an angry customer.
You could roleplay that with a typical LLM. It would do a convincing job, but that’s the problem.
An LLM would work with you to de-escalate the situation. It would be receptive to your attempts to solve the situation and provide a path to a satisfying conclusion.
Have you ever dealt with an irate customer before?
Sometimes they don’t want to calm down and work with you. Sometimes there is no chance of a happy ending. It’s not the outcome we hope for, but it is one worth training around.
Therein lies the problem with LLMs. They can create convincing dialogue, but at the end of the day, they are too eager to please. They will build up your confidence, which can be valuable, but they fail to prepare learners for real challenges.
Realistic Training Requires Rules
Good simulations need an underlying model of the situation.
That model might track objectives, emotional state, progress, mistakes, or other conditions that determine how the interaction develops.
Here at SkillSaige, we use a proprietary system we call the Logic Lattice to manage all of that, since LLMs are not great at tracking all of that.
Without a robust underlying model, conversations with AI can quickly turn into an exercise in ego. The AI can keep the conversation moving because continuing the conversation is what it is designed to do.
That can even reward bad behavior.
A sufficiently flexible chatbot may find a plausible way to accommodate something a real customer, employee, or manager never would.
The conversation still sounds good, but the training does not necessarily teach the right lesson.
Language Models Are Still Incredibly Useful
None of this means language models are bad tools for simulations.
Quite the opposite.
They solve one of the hardest problems in simulated training, which is that language is messy.
People phrase the same idea in thousands of different ways. We imply things, make jokes, and use incomplete sentences. Traditional software struggles to account for all of that.
LLMs are extraordinarily useful for interpreting that ambiguity and producing natural responses.
We just shouldn’t confuse that capability with the entire training experience.
At SkillSaige, we use language models as one part of our larger system. They help us understand and produce human language, while our proprietary systems determine how the situation actually progresses.
That separation matters because the goal isn’t merely to make the AI sound like a person.
The goal is to create a realistic opportunity to practice.
If you’re interested in digging deeper into this topic, our CEO, Sayre Blake, has a longer post about it on her blog.
