Ten years ago, this would have sounded absurd: learners now have to think about their thinking and the machine’s output too.
We’ve already made the case that metacognition — the ability to plan, monitor, and evaluate one’s own reasoning — is one of the strongest predictors of learning success. But AI adds a new layer. When learners use AI to support their work, they are not only monitoring their own thinking. They are also deciding whether the AI’s output is accurate, complete, and worth trusting. That is a different cognitive task, and most training programs have not caught up to it yet.
In our earlier piece on metacognition in training design, Why Training Fails: Learners Don’t Know How They Learn, we covered the two core components: metacognitive knowledge — awareness of how you think and learn — and metacognitive regulation — the ability to use that awareness to plan, monitor, and evaluate in real time. Both are teachable. Both require intentional design.
AI introduces a new area of metacognition: evaluating the reasoning of a system that can sound reliable and certain even when it is wrong. When a learner reads a textbook, they bring critical reading skills to bear. When a learner receives AI output — a summary, a draft, an explanation — the same set of skills may not work. The AI output looks authoritative. It doesn't hedge or visibly struggle the way a peer would. And that surface credibility can short-circuit the critical monitoring that good metacognitive practice requires.
AI use can produce better outputs while producing worse thinking. Students who rely on AI for writing tasks may produce stronger final products but retain significantly less of the underlying knowledge — and report reducing their own cognitive effort in the process.1 Students produced stronger final products but retained less of the underlying material, and reported putting in less cognitive effort to get there.
Over-reliance on AI output isn't simply a habit problem. It's a response to a system that is structurally designed to sound certain.2 When AI provides confident-sounding recommendations with explanations, users increase their trust even when the AI is wrong.3 The confidence itself becomes the signal — and it's a misleading one.
What learners actually need is metacognitive sensitivity: the ability to distinguish between when AI is reliable and when it isn’t.4 That’s a learnable skill. It has to be designed into the experience.
The fix is to design the evaluation skill into the task itself, rather than designing AI out of it.
The difference can be one design decision. Submitting an AI summary only asks the learner to finish a task. Asking them to compare it against the source, mark where it drifts, and explain in writing what they kept or cut changes the cognitive demand entirely — same tool, same output.
Image alt. text: Four-step workflow showing an AI-generated topic summary, AI critique, learner-edited summary, and instructor feedback, connected by arrows.
Other design moves that develop trust calibration:
None of these require learners to reject AI.
These skills map directly onto the Learning Environment Model. Evaluate Before asks what learners already know — and in an AI-rich context, that prior knowledge becomes the reference point for evaluating output. Monitor extends beyond strategy to include a new question: is the learner checking the tool’s output, or just accepting it? Evaluate After adds one more test: can the learner demonstrate understanding independently of what AI produced for them?
To help learners evaluate AI-generated output, consider using these reflection questions:
Before Using AI
While Reviewing
After Using AI
Image alt. text: Three-column checklist showing reflection questions for before, during, and after using AI.
When we design with these phases in mind, metacognitive trust calibration stops being an abstract goal and becomes a visible design decision. You can look at a training blueprint and ask: where are learners required to evaluate AI output? Where do they have to justify their trust? If those moments are missing, you know exactly what to add.
For most of human learning history, the core metacognitive task was monitoring your own understanding. That task has not gone away. But AI has added a parallel one: monitoring the reliability of a sophisticated, confident, sometimes-wrong collaborator that is often rewarded for guessing rather than abstaining.
It's a new and teachable skill, and it belongs in the design of any training experience that hands learners AI tools.
AI can help us move faster, generate ideas, and expand what is possible in learning design—but it still depends on human judgment. The real opportunity is not simply using AI more often. It is learning how to use it with greater intention, reflection, and responsibility.
That is the focus of our Human Side of AI in Learning campaign, where we explore how training and development leaders can approach AI as a tool for better thinking, not a replacement for it. The campaign includes practical insights on AI, learning design, and the human decisions that shape meaningful learning experiences.
As part of the series, LX Studio’s Trevor Cox, Ph.D., sat down with Laura Dumin, Ph.D., Faculty-facing AI Coordinator at the University of Central Oklahoma, to discuss what learning leaders can take from higher education’s evolving relationship with AI.
Explore the campaign and listen to the conversation here.
View our resource library that inspired our The Human Side of AI in Learning series.
1 Hurley, K. (2025). The paradox of AI assistance: Better results, worse thinking. EDUCAUSE Review. https://er.educause.edu/articles/2025/12/the-paradox-of-ai-assistance-better-results-worse-thinking
2 Kalai, A. T., Nachum, O., Vempala, S. S., & Zhang, E. (2025). Why language models hallucinate. arXiv:2509.04664. https://doi.org/10.48550/arXiv.2509.04664
3 Naiseh, M., Al-Thani, D., Jiang, N., & Ali, R. (2023). How the different explanation classes impact trust calibration: The case of clinical decision support systems. International Journal of Human-Computer Studies, 169, 102941. https://doi.org/10.1016/j.ijhcs.2022.102941
4 Lee, D., Pruitt, J., Zhou, T., Du, J., & Odegaard, B. (2025). Metacognitive sensitivity: The key to calibrating trust and optimal decision making with AI. PNAS Nexus, 4(5), pgaf133. https://doi.org/10.1093/pnasnexus/pgaf133