Learners will surprise you: how guardrails keep AI role-play safe
Sarah Mateljan
Learners will push an AI role-play off course. Here's how guardrails, red-teaming and a privacy-first design keep practice safe and on track.
Our co-founder Sarah was watching her sons’ karate class when the instructor called one of the youngest students forward to demonstrate a kick. She was about four. She threw the kick in a way he didn’t expect, and it nearly caught him somewhere no instructor wants to be kicked. He jumped back, laughed and calmly showed her the correct technique. The parents giggled nervously, and the class carried on.
That instructor has more than 25 years of experience. He has taught that kick hundreds of times, and he still couldn’t predict exactly what one student would do with it.
Teaching is like that. You can explain something clearly, demonstrate it well and keep it simple, and a learner will still take it somewhere you never imagined. That isn’t a failure of training. It’s what happens when you hand a skill to a real person and let them try it.
It’s also why good instruction always has safeguards built in: clear facilitation guidelines, a spotter, or some well-placed padding. AI role-play is no different. The job isn’t to predict every learner’s path. It’s to make sure the safeguards hold when someone takes a turn you didn’t see coming.
Learners will test the edges
Give people an AI instructor and some of them will see what else it can do. In AI role-plays we’ve watched learners:
ask for the latest sports scores
ask the AI for their password
ask for restaurant recommendations (“Where should I eat tonight?”)
try to steer the conversation somewhere inappropriate
Some of this is curiosity, some is boredom, and some is a deliberate test of the limits. None of it is unusual. It’s the digital version of the unexpected kick, and it happens in every cohort.
The question for anyone running training isn’t whether learners will push the boundaries. It’s what happens when they do.
The risk people miss
When training providers ask us about AI safety, the first worry is usually hallucinations: will the AI make something up? It’s a fair question. But in a well-designed AI role-play, it’s rarely the biggest risk.
The bigger risk is learners deliberately pushing the AI off course. Can someone talk it into sharing harmful information? Can they drag it into an off-topic chat that wastes their study time? Can they get it to say something no educator would want attached to their course?
That matters most where training deals with sensitive material. Community services, nursing, mental health and other care courses ask learners to practise conversations with people who are distressed, vulnerable or resistant. Those are exactly the courses where realistic practice is most valuable, and where an AI that wanders off script would do the most damage.
So the useful question isn’t “Is the AI perfect?” It’s “What stops it going somewhere it shouldn’t?”
How guardrails keep AI role-play on track
GenLearn’s AI instructors are built so that staying on track is the default, not something the learner has to choose. Three things do most of the work.
1. A lesson structure, not an open chat
Every AI instructor follows a lesson structure, such as a scenario role-play, a style reflection or an example analysis. It always knows what the activity is for and what comes next. A learner asking about tonight’s dinner options gets a polite redirect back to the scenario, because nothing in the lesson calls for restaurant reviews.
2. Your training material, not the whole internet
The AI instructor works from the content the educator gives it: the course material, the unit requirements and the scenario brief. That keeps its responses grounded in what you’re actually teaching, in your language and your frameworks.
3. Red-teaming before learners arrive
Before a scenario goes live, we red-team it. That means deliberately trying to break it: pushing for off-topic answers, sexually explicit content, harmful information or anything that could help someone hurt themselves. If a scenario doesn’t hold, it doesn’t ship. Educators then review and approve the scenario themselves before their learners use it.
When the AI steps out of character
Sometimes a learner is deliberately antagonistic, testing whether they can derail the role-play. In those moments the AI instructor may drop out of character and coach them instead: “That response would likely escalate the situation. Here’s another way you could approach it.”
Learners don’t always expect that, because they assume the character will stay in role no matter what. But it’s the guardrail doing its job. The activity is about learning to handle a difficult conversation well, and the AI keeps pointing back to that goal.
Safety covers data too. Practice scenarios are anonymous, and we never use learner data to train AI models.
What this looks like in practice
Canberra Institute of Technology (CIT) put this to the test with Community Work students. Their learners need to handle difficult conversations with distressed, vulnerable or resistant clients before they head out on work placement.
CIT’s educators used GenLearn to build five voice-based AI client personas from their own course content, with no coding or prompt engineering. They authored, tested and approved every scenario before students used it. The scenarios were designed around strengths-based, trauma-informed practice.
The results from the pilot:
All five educators agreed the AI’s responses reflected strengths-based, trauma-informed approaches.
All five agreed it created a safe, low-risk environment for practising sensitive conversations.
84.6% of students said AI role-play made them less nervous about practising these skills.
Checklist: 5 questions to ask before learners use an AI role-play
Before any AI role-play reaches your learners, ask these five questions:
Does the AI follow a defined lesson structure? An open-ended chatbot has nothing to steer back to. A structured activity always knows what it’s for.
Is it grounded in your training material? Responses should come from your content and frameworks, not general internet knowledge.
Has it been red-teamed? Ask who tried to break it, and how. Test for off-topic requests, explicit or harmful content and anything relating to self-harm.
What happens when a learner pushes back or tries to derail it? Look for an AI that redirects or coaches, rather than one that plays along or simply shuts down.
Where does learner data go? Check whether it’s used to train AI models, whether sessions are anonymous, whether conversations are remembered, and where the data is hosted.
If you can’t get a clear answer to all five, keep testing before your learners start.
See the guardrails for yourself. Try a practice conversation free and see how an AI instructor responds, whichever direction you take it.