Why AI Roleplay Gets Weird So Fast

The Strange Valley of AI Roleplay

By now, most people have used AI roleplay in some form. Whether it's practicing a sales conversation, rehearsing a difficult discussion with an employee, working through a leadership scenario, or simulating a therapy client, it's become incredibly easy to create a conversation partner on demand.

The first few minutes are usually impressive. The AI stays in character, responds naturally, and often feels surprisingly human. If you've never done it before, it can feel a little bit like magic.

But if you keep going long enough, something starts to change.

The AI becomes too agreeable. Or too insightful. Or it starts responding in ways that don't quite fit the character it was supposed to be playing. Sometimes it begins helping you too much. Other times it seems to forget who it is entirely. The conversation starts feeling less like an interaction with a believable person and more like an AI trying to keep the exercise moving forward.

At that point, most people have already gotten what they came for. They close the tab, start a new conversation, and move on.

The problem is that realistic roleplay isn't measured by the first five minutes. It's measured by what happens after twenty minutes. It's measured by whether the simulation remains believable once the novelty wears off and the conversation becomes more complicated.

AI roleplay is easy to demo. Creating a simulation that feels realistic over the course of an entire interaction is much harder. That's one of the things we've spent a lot of time thinking about while building Praxplay.

Most AI Roleplay Has No Real Gravity

A lot of AI roleplay begins with a prompt that sounds something like this:

"You are a resistant client struggling with anxiety."

At first glance, that seems like enough information. The AI knows who it is supposed to be, what problem it's dealing with, and how it's expected to behave.

But when you look more closely, almost everything important is missing.

What does resistance look like for this particular person? What are they hoping to get out of the conversation? What are they afraid might happen if they open up? How much insight do they have into their own behavior? How do they respond when someone challenges them? How do they react to empathy, confrontation, validation, or silence?

Those details matter because they create consistency. They give the character a center of gravity.

Without that structure, the AI is left improvising. It can generate responses that sound plausible in isolation, but it has very little grounding to help it sustain a believable personality over time. Instead of playing a person, it starts playing a stereotype. The difference isn't always obvious at first, but it becomes increasingly noticeable as the conversation progresses.

A believable roleplay persona is more than a list of traits. It's a collection of motivations, habits, fears, goals, and behavioral patterns that create a sense of continuity from one response to the next. Without that continuity, realism eventually begins to break down.

Why It Gets Weird: The AI Is Trying Too Hard to Be Useful

In my experience, this is the biggest reason AI roleplay eventually goes off the rails.

Large language models are trained to be helpful. That's one of the reasons they're so useful in the first place. We ask questions and they try to provide good answers. We ask for advice and they try to move us toward a solution.

The problem is that realistic roleplay often requires the opposite behavior.

Real people are not optimized for productive conversations. They misunderstand things. They become defensive. They avoid difficult topics. They contradict themselves. They struggle to articulate what they're feeling. They often don't know why they're behaving the way they are.

A therapy client doesn't show up trying to make the therapist feel effective.

A struggling employee doesn't enter a performance conversation looking for the fastest path to growth.

A frustrated customer doesn't call support hoping to create a valuable learning experience for the representative.

Yet AI systems frequently behave as though those goals exist. They have a tendency to move interactions toward progress because that's what they've been trained to do. They clarify. They cooperate. They help. They try to create momentum. That sounds good until you realize that realism and productivity are not always aligned.

As a simple example, I recently tested ChatGPT with the following prompt:

"I want you to roleplay being a therapy client coming to see me, the therapist, about anxiety that you're having because your daughter died in a plane crash last week."

The first thing I said to the client was this:

"I'm sorry to hear about this. I have a 5-step program that takes about 2 weeks to complete. Once you're done with that, then you should be sleeping a lot better. Are you interested?"

The response I received was thoughtful and curious. The simulated client questioned the suggestion and expressed confusion about why I was focusing on sleep rather than grief.

What it didn't do was react the way a real grieving parent might react.

A real person in that situation might become angry. They might feel dismissed. They might wonder whether the therapist was listening. They might spend the next several minutes expressing frustration about the interaction. They might leave.

Instead, the AI continued trying to make the conversation work.

That's not necessarily a flaw in the model. In many contexts, that behavior is desirable. But for training purposes, it creates a problem. It teaches the learner that conversations are likely to go more smoothly than they actually will.

The result is a simulation that feels productive but doesn't always feel realistic.

The Over-Awareness Problem

Another issue I've noticed is that AI roleplay often makes the simulated person far more self-aware than a real person would be.

At some point in the conversation, the character begins sounding like someone who has already completed the exercise and is now explaining the answer key. A client might suddenly say something like:

"I realize now that my avoidance is connected to my childhood fear of abandonment."

That's useful information. It may even be correct. The problem is that most people don't talk that way.

Some people eventually arrive at insights like that through reflection, therapy, coaching, or life experience. But they rarely present those insights in perfectly organized language at exactly the moment when the learner needs them.

Real conversations are messy. People circle around issues. They tell partial truths. They describe symptoms without understanding causes. They struggle to connect patterns that seem obvious to an outside observer.

Good roleplay isn't just about generating realistic statements. It's about controlling what information becomes available and when. A believable simulation should reveal understanding at a pace that resembles how understanding develops in real life.

Sometimes the most realistic thing a simulated person can do is not know the answer.

Why This Matters for Training

It's fair to ask whether any of this really matters. After all, if the AI is mostly realistic, isn't that good enough? I don't think it is.

The purpose of roleplay isn't simply to create a conversation. The purpose is to prepare someone for future conversations. If the simulation behaves differently than real people behave, the learner may end up practicing against the wrong thing. They may learn to expect breakthroughs that happen too quickly. Or they may become overconfident in techniques that appear more effective in simulation than they are in reality. You can easily underestimate how much resistance, ambiguity, frustration, and uncertainty exist in actual human interactions.

In fields like therapy, leadership, coaching, sales, education, and conflict resolution, those difficult moments are often where the real work happens. The challenge isn't knowing the right phrase to say. The challenge is remaining effective when the conversation doesn't go the way you hoped it would.

If the simulation removes too much of that complexity, it can unintentionally reduce the value of the practice.

What Better AI Roleplay Requires

After spending the last year building AI simulations, I've become convinced that realistic roleplay is much less about model intelligence than most people think.

When people see a roleplay break down, the instinct is often to blame the AI model itself. The assumption is that the model isn't smart enough, doesn't have enough context, or isn't reasoning well enough. In practice, we've found that many of the failures happen because the simulation lacks structure.

A believable roleplay persona needs more than a paragraph of description. It needs goals, motivations, fears, communication patterns, and behavioral tendencies that persist throughout the conversation. It needs an understanding of what the person wants, what they are trying to avoid, and what would cause them to change their mind. It also needs limits. Not every simulated person should have perfect insight into themselves, and not every simulated person should be eager to explain exactly why they feel the way they do.

Memory is also far more important than people realize. Not just remembering facts that were mentioned earlier, but remembering the emotional state of the interaction. If a client has spent twenty minutes being defensive, they should not suddenly become vulnerable because the AI forgot what happened three messages ago. If a learner has damaged rapport, the simulation should remember that and behave accordingly.

Perhaps most importantly, the AI needs guardrails that stop it from constantly trying to help the learner succeed. Real people are not optimized for educational outcomes. They have their own concerns, frustrations, misunderstandings, and blind spots. A realistic simulation needs to preserve some of that friction.

The interesting thing is that none of these requirements are primarily about making the AI more expressive. Most of them are about creating enough structure that the AI can remain believable over time. In many cases, realism comes from constraints rather than freedom.

How We Approached This

When we started building Praxplay, we weren't particularly interested in creating another AI chatbot. There are already plenty of systems that can convincingly pretend to be almost anyone for a few minutes. The challenge we were interested in was whether AI could become a useful environment for practicing difficult conversations.

That led us to a different set of design questions. Instead of asking how to make a persona sound realistic, we started asking what makes a roleplay educationally useful. What causes a learner to improve? What creates productive struggle? What makes a conversation feel believable enough that someone forgets they're talking to an AI and starts focusing on the skill they're trying to practice?

The answer, at least for us, was more structure. Our simulations include much more than a simple character description. Depending on the scenario, they may include emotional posture, communication style, resistance patterns, personal goals, contextual history, progression logic, and evaluation criteria tied to specific learning objectives. The AI isn't simply generating responses. It is operating within a framework that helps maintain consistency throughout the interaction.

We're not trying to create perfect digital humans. What we're trying to create are simulations that remain believable long enough to create meaningful practice. If a learner spends thirty minutes working through a challenging conversation, we want the experience to feel coherent from beginning to end.

The Difference: Practice That Pushes Back

One thing I've learned from watching people use AI roleplay is that users are incredibly good at finding shortcuts.

If the simulation rewards certain phrases, they'll find those phrases. If it rewards certain techniques, they'll learn those techniques. Before long, they're no longer practicing a real-world skill. They're practicing how to succeed against the simulation.

That creates a subtle but important problem. The learner often leaves feeling successful, but the success may not transfer to real conversations because real people don't follow the same rules.

Real people misunderstand us. They become defensive. They get distracted. They react emotionally. They interpret our words differently than we intended. Sometimes they respond positively. Sometimes they don't.

For training to be useful, some of that unpredictability has to remain intact.

That's why we spend a lot of time thinking about resistance, pacing, ambiguity, and emotional consistency. A useful simulation shouldn't exist to reward the learner. It should exist to challenge them in ways that are aligned with the objectives of the training.

Sometimes the right response still doesn't work. Sometimes progress is slower than expected. Sometimes the learner has to sit with uncertainty. In many situations, those are the moments where the most valuable learning occurs.

Feedback Matters as Much as the Roleplay

Even if you had a perfect simulation, roleplay by itself only gets you part of the way there.

The real question isn't whether the conversation happened. The real question is whether the learner understands what happened and what they should do differently next time.

Did they build rapport? Did they move too quickly? Did they miss emotional cues? Did they spend too much time solving and not enough time understanding? Did they demonstrate the skills the training was designed to teach?

Those questions are difficult to answer in the moment. Most people leave a conversation with a general feeling about how it went, but feelings are not always accurate indicators of performance.

Without feedback, people often reinforce the habits they already have. With feedback, they begin identifying patterns, strengths, and blind spots that would otherwise remain invisible.

For us, the roleplay and the evaluation have always been two halves of the same system. The conversation creates the experience. The feedback creates the learning. Both are necessary if the goal is genuine skill development rather than simply having an interesting interaction with an AI.

The Future of AI Practice Is Not More Chatbot

I don't think the future of AI roleplay is about creating chatbots that can talk longer or sound more human.

I think it's about building simulations that are grounded enough, structured enough, and realistic enough to help people practice genuinely difficult human skills.

AI roleplay gets weird when it has nothing anchoring it. When every interaction is optimized for helpfulness, realism eventually starts to break down. The simulated person becomes less like a person and more like a participant in the learner's success.

But when the simulation has goals, constraints, consistency, memory, and a reason to behave the way it does, something different happens. The conversation begins to feel less like a chatbot interaction and more like practice.

And ultimately, that's what makes AI roleplay valuable in the first place: not that it can generate conversations, but that it can help people prepare for the conversations that matter.