On Shame and AI in the Classroom

After a decade of designing and delivering accelerated training programs, I’ve recognized some patterns that emerge in every cohort. Because a lot of my work is in the transformational learning space (I think it sounds fancier than “reskilling”), I’m constantly ruminating on the best way to achieve great outcomes.

For my adult learning programs, I’ve been moving away from synchronous, live delivery for a while now because of the three-speed problem.

The Three-Speed Problem

In accelerated, synchronous training, you inevitably end up with three groups. These are my unscientific names for these groups.

• The high flyers. These are often people who did some self-study or had some relevant experience or high aptitude in the meta-skills related to the topic. They could be pushed farther, but are constrained by the pace of the class.
• The pace-keepers are generally on pace and on track. This is often the persona that we design learning experiences for.
• The climbers who need extra attention and are falling behind. I call them climbers because it’s often the case that they’re starting from true zero or lack some meta skills coming in that slow them down. Oftentimes, the reality is they’re starting at the base camp, while others are starting closer to the peak.

In terms of the classroom experience, equal parts of the cohort are either bored or stressed out of their minds. It’s a structural problem I’ve never fully solved, despite years of iteration. A lot of my focus ends up being on the climbers and the pace-keepers, because they’re often the ones who will benefit most from my direct mentorship.

There’s a big problem with a group of adult learners, and its name is shame.

The Weight of Not Knowing

Every time I start a new cohort, I say the same thing on day one:

“This is the only time in your professional career when you can ask anything you want. There are no dumb questions, no judgment. My only goal is your competency and confidence.”

I mean it. I enforce it. I model vulnerability by including some live, hands-on activities where I can make mistakes and have to debug my work. And if I can’t figure something out on the spot, I take it back to the lab and return with not only an answer and a work sample, but also a detailed description of how I arrived at the solution.

Learning how to research and apply information is arguably the most important skill technical professionals develop.

And yet, a subset of students simply will not ask questions. I’m frequently frustrated by this, but I have a theory as to what is going on. To illustrate, let’s use Joe, who is a fictional climber, and Katherine, who is a high flyer in the same group.

Joe knows he’s behind Katherine. He sees her nodding along while he’s still processing the previous concept. When we engage in hands-on activities, Katherine’s output is not only faster but also higher quality. She considers edge cases and anticipates problems before they occur. And so, when Joe considers raising his hand, he feels exposed. The appearance of incompetence, even in a dedicated learning environment, triggers something primal.

So instead, he keeps quiet. He nods along, and when I check in with him, he says things are “fine.” By the time his assessments reveal the truth, we’ve lost valuable weeks. In accelerated training, that gap is usually insurmountable. This is especially true in technical topics where early competencies provide the foundation for later ones.

There quickly comes a point where Joe cannot advance without shoring up his foundations.

Enter the AI Tutor

In my opinion, this is where AI becomes genuinely interesting. An AI tutor offers Joe something new: the ability to ask questions without the weight of human judgment. It has no issues explaining the same concept forty times without a hint of frustration. It never sighs, never glances at the clock, never makes him feel like he should have gotten it by now.

For someone paralyzed by shame, this could be a transformative experience.

But…

The Hidden Costs

I’ve spent years refining learning systems and observing real human behavior. Based on this, I have some concerns.

Cognitive Offloading

The first concern involves cognitive offloading. This is especially a concern in software development. It’s dangerously easy to produce working solutions without understanding. Joe could submit perfect code while having outsourced all the actual thinking to AI.

One of my mentors once told me something that stuck. “Code is an expression of a solved problem.” And he’s right, if you don’t understand the problem, you can’t write the code. Coding is a profession that requires a specific mindset, and everyone, no matter how “gifted”, must develop that mindset.

That struggle is a necessary part of becoming a skilled professional. It cannot be shortcut without cost. We can argue about the state of the job market, but there is always a shortage of skilled talent.

If you want to see me get agitated, ask me about the recent interviews I’ve had with computer science graduates who can’t write a for loop. There is a zero percent chance that they can design, debug, or evaluate the quality of code. Even if AI increases the speed of code production, the skill with the most market value will be evaluation, not generation.

Hallucinations

AI confidently delivers incorrect information. Even OpenAI is admitting the reality that I’ve been screaming about for years. LLM hallucinations are mathematically inevitable. So, let me take a moment to express my disdain for grifters who continue to say that “you’re just prompting it wrong.”

You may imagine me offering the grifters a 1-fingered salute here.

The issue here is how Joe, who hasn’t yet developed the competencies to recognize bad patterns, is supposed to identify when the AI is wrong? He lacks the very foundation needed to be skeptical.

Even the high flyers in my cohorts, when I take the restrictions off and let them start using AI assistance, have frequently reported being sent down rabbit holes or ending up in loops where the AI creates problems and can’t fix them. For the high flyers, they simply roll up their sleeves and do the research or ask me, since they rarely have the same issues of shame that others do.

But Joe? He not only can’t fix the problem, but it ends up making him even more confused.

Interestingly enough, some of my high flyers report that the AI usage, when it goes off the rails, instills self-doubt, especially when it tries to take them down a path that they don’t recognize from my materials. They have the “spider sense” that something isn’t right, but often lack the confidence to challenge it because of their inexperience. When it happens, this is a very bad thing.

Resilience and Professionalism

Not knowing things and asking for help is a common experience for every working professional. One thing I tell my learners is that managers don’t like surprises. As a team member, it is your responsibility to be transparent about timelines, blockers, and areas where you lack confidence or understanding.

So, if Joe never learns to manage his feelings of shame in a classroom, how will he handle his first code review? What about his first architecture meeting, where he needs to admit he doesn’t understand the requirements? Every time we let him avoid that discomfort, we rob him of the chance to build resilience.

Will the AI tutor prop him up enough to be successful in the classroom, only to later be fired for poor performance? That scenario is worse for all parties. The employee who gets dismissed and the employer who lost time and resources on someone who couldn’t fulfill their role.

My concern is that the students who most need to develop resilience to shame are the ones most likely to use AI to avoid it entirely. Anyone who has witnessed physical therapy can tell you that without pushing yourself appropriately, the healing process is much less effective.

And yet, I can’t ignore the reality that shame is causing real harm right now. If Joe can’t bring himself to say “I don’t understand,” then if an AI tutor can keep him engaged and learning, even imperfectly, isn’t that better than losing him entirely?

The Experiments I’m Running

I’m thinking a lot about this topic and starting to run some experiments. Here are some things I’m looking at:

• What if a learner uses AI to workshop questions before class, helping them identify the root causes/misunderstandings more clearly?

  • I haven’t been able, so far, to get learners to do this effectively. I think it has merit, but it needs work.

• How about reflection assistance? Is there value in a learner using AI to verify understanding by explaining what they know?

  • This has been generally effective in my experiences so far. The issues are hallucinations, of course, but a tougher issue is that AI often introduces topics or techniques that the learner isn’t ready for yet. They’re really bad, in my experience, at keeping things constrained to what the learner should know at a point in time.

• Can AI effectively generate more hands-on problems and solutions?

  • In my experiments so far, the answer is no, it can only produce very basic examples effectively.

Unfortunately, I don’t have clean answers. But as much as educational thought leaders like to talk about “lifelong learning”, the only people who successfully practice it are those who are comfortable with not knowing, confident in their ability to figure things out, and resilient enough to ask for help when they need it.

Those might be the traits AI can’t help with at all.