Why people practice more honestly with an AI avatar than with a colleague

Everyone who has ever given a training session knows the moment. It is quarter past two, the theory has been covered, and the trainer asks who wants to demonstrate the conversation. A silence falls over the room that you can almost touch. Eventually someone volunteers, often the same person who always volunteers, and then proceeds to have a conversation that goes remarkably smoothly. I always think of that moment when someone asks me what an AI avatar actually adds.

The latter should concern you. I have seen hundreds of these role-plays and they almost never really go wrong, while the same conversations regularly go off the rails in practice. The explanation is not that people can do it better in the training room. They are doing something different from practicing.

The strongest argument for practicing with an AI avatar therefore lies, for me, in the question of who is watching, and much less in the technology.

In the training room, people perform

Anyone who stands in front of a group of colleagues takes that group into account. You choose the wording you know will land well, you avoid the path whose outcome you are not sure about, and as soon as things become really tense, you step out of the role with a joke. Nobody does this consciously. It just happens, because fourteen people are watching who you will be sitting in the same meeting with again tomorrow.

And then there is the question of who is watching. In a management team, your own director is in the room. In a department, the manager who evaluates you is sitting there. Under those circumstances, asking someone to practice a conversation in which they are allowed to show themselves at their worst is a rather unreasonable request.

Amy Edmondson showed in Administrative Science Quarterly that psychological safety, the shared belief that a team is safe for interpersonal risk-taking, is associated with learning behavior, and that this learning behavior forms the link to team performance. Asking for help, admitting mistakes, trying something that might fail: they are all actions in which you expose yourself. In a team where that is not taken for granted, you get a role-play that looks good and delivers nothing.

What happens when the audience disappears

This is where it gets interesting, because it has been researched.

Lucas, Gratch, King and Morency published a study in 2014 in Computers in Human Behavior in which participants had a conversation with a virtual conversation partner. One group was told that there was a human behind it who was watching and directing the conversation. The other group was told that it was fully automated. To the participant, the conversation looked the same in both cases.

The difference was in what people did. The group that thought they were talking to a computer reported less fear of self-disclosure and less tendency to manage the impression they made, showed more sadness in their facial expressions, and, according to independent assessors, disclosed more about themselves.

One belief therefore made the difference, while the conversation partner was identical. What makes people shut down is the gaze of someone who can judge them, and we actually knew that all along.

For anyone who wants to train soft skills, that is a more important fact than it usually gets credit for. An AI avatar has the characteristic that nobody is watching, and that characteristic sits exactly at the point where practice in real-life settings breaks down.

Practice requires that you are allowed to get it wrong

There is a second reason why this matters, and it comes from a completely different angle.

Ericsson, Krampe and Tesch-Römer described in Psychological Review what makes deliberate practice effective. You have to work at the edge of what you can do, you have to receive immediate feedback, and the difficulty has to increase. Later research has nuanced the scope of their claims, but there is little disagreement about that mechanism.

Working at the edge of your ability means that things will go wrong frequently by definition. If things do not go wrong, you were not working at the edge. And that is where the clash with the training room occurs: an environment in which mistakes have social costs is an environment in which people will not push themselves to the edge.

Once that audience disappears, the learner’s behavior changes. People try the wording they are not sure will work. They allow a silence to last uncomfortably long to see what happens. They have the conversation again because they messed it up the first time, and it costs them nothing.

What an AI avatar does not solve

Now comes the part that does not help my own story.

Such a digital conversation partner does not give you a colleague who says afterward: “I noticed that I shut down when you said that.” That sentence can sometimes be worth more than ten exercises in a development program, because it is the moment when someone sees something about themselves that they could not see on their own. That is exactly what a good trainer facilitates and what a system cannot do.

There is also a group for whom it remains unfamiliar. Usually that passes after two conversations, especially when the scenario comes from their own work, but not for everyone. I think you have to accept that rather than glossing over it.

And team dynamics cannot be practiced this way. Four people who have known each other for years, with all the history that entails, require peer consultation or a facilitated conversation. No simulation can substitute for that.

What it does solve is repetition

What it does solve is repetition, and that is the bottleneck in almost every development program. Saks and Belcourt showed in Human Resource Management that the proportion of participants who apply the training content drops from 62 percent immediately afterward to 34 percent after one year. You stop that decline with practice moments in the weeks that follow, and those are difficult to organize in a training room.

How we have set it up

At PractAIce, the user builds their own scenario. If someone needs to address an employee about consistently arriving late, they describe the situation, choose a persona that fits this colleague and describe the behavior and competency profile as context. The conversation that follows is about this employee, in this organization, with this problem.

The avatar responds to what is actually said. Change your tone, and the response changes. Ask a closed question, and you get a short answer with all the awkwardness that comes with it. The AI coach then provides feedback on the behavior you demonstrated, linked to the competencies that were configured in advance.

If things went wrong at one particular moment, for example at the first objection, you can use a short microtraining to return to that point in the conversation instead of doing everything again. Two to four minutes, and you immediately notice what a different formulation does to the other person.

How to use this when onboarding new people is explained in Onboarding: new people don’t drop out because of the handbook. We wrote more about the broader effect of this way of practicing in AI role-plays: why practicing with an AI avatar actually sticks. For the specific case of giving feedback, where the social costs of practicing may be highest, there is more information in Feedback training.

The combination works better than either on its own

What I arrive at is a division of roles that is more of a benefit than a threat to trainers. The trainer does what they are irreplaceable at: exposing the pattern, confronting people, creating the moment when someone sees themselves. The avatar does the miles, in the weeks in between, without a schedule and without an audience.

Those who combine the two see something different happening in the sessions. People are no longer practicing whether the model works, because by then they know from experience. They talk about what happened when they tried it, and that is much more interesting.

Frequently asked questions about the AI avatar

What is an AI avatar?

An AI avatar is a digital conversation partner with a face and voice that responds in real time to what you say. In a training context, the avatar might play a customer, an employee or a patient, allowing you to practice a conversation without needing an actor or colleague.

Why do people practice more honestly with an AI avatar?

Because nobody is watching who can judge them. Research into virtual conversation partners shows that people who think they are talking to a computer are less concerned with the impression they make and disclose more about themselves.

Doesn’t talking to an avatar feel unnatural?

Usually during the first few minutes. After that, the conversation takes over, especially when the scenario comes from the person’s own work. Participants often notice that they experience the same tension as in a real conversation.

Does an AI avatar replace the trainer?

No. A trainer exposes patterns and creates insight, and a system cannot do that. What the avatar takes over is repetition, and that is precisely where almost every development program gets stuck.

What conversations can you practice with an AI avatar?

Anything that takes place between two people: giving feedback, delivering bad news, dealing with resistance, negotiating, coaching, de-escalating and selling. Team dynamics over a longer period fall outside its scope.

The strange thing about this subject is that reading tells you very little about it. You notice whether it works within three minutes of a conversation. Take the most difficult conversation you have to have this month and try it once on the AI avatar training page. If it leaves you cold, you know enough. And if you notice your heart beating a little faster halfway through, you know that too.

Sources

• Lucas, G.M., Gratch, J., King, A. & Morency, L.P. (2014). It’s only a computer: virtual humans increase willingness to disclose. *Computers in Human Behavior*, 37, 94-100. https://doi.org/10.1016/j.chb.2014.04.043
• Edmondson, A.C. (1999). Psychological safety and learning behavior in work teams. *Administrative Science Quarterly*, 44(2), 350-383. https://journals.sagepub.com/doi/10.2307/2666999
• Ericsson, K.A., Krampe, R.T. & Tesch-Römer, C. (1993). The role of deliberate practice in the acquisition of expert performance. *Psychological Review*, 100(3), 363-406. https://doi.org/10.1037/0033-295X.100.3.363
• Saks, A.M. & Belcourt, M. (2006). An investigation of training activities and transfer of training in organizations. *Human Resource Management*, 45(4), 629-648. https://onlinelibrary.wiley.com/doi/abs/10.1002/hrm.20135