When AI Removes Friction from Interaction, Do We Lose Something Important?

At its best, AI can serve as an effective learning tutor. At its worst, it does our thinking for us and replaces genuine interaction with a mere illusion of connection. New research raises the question of how we can use AI without losing skills that are essential for learning, such as independent thinking, critical judgement, concentration and the ability to have difficult conversations with other people.
Päivi Lohikoski

In this blog post, I reflect on a couple of trends that I have encountered recently. I am openly curious about AI, and I have also used it as a sparring partner while writing this piece. I am a trainer in interaction and leadership, which is why I wanted to explore the topic from the perspectives of learning and interaction.

Two interesting terms I came across are metacognitive laziness and pseudo-interaction in relationships and learning. MIT professor and researcher Sherry Turkle has studied the impact of AI on human relationships and discussed the topic in various forums. According to Turkle, social media has taken our attention, and AI is now taking our capacity for attachment. She is particularly concerned about what relationships with AI do to our ability to be with other people. The concern relates to uses in which a machine imitates a loving and caring relationship. Emerging trends suggest that this can make genuine friendship, collaboration and romantic relationships difficult for some people. This is especially evident among young people and lonely adults.

Friction Builds Resilience and Competence

When AI provides immediate, conflict-free interaction and artificial empathy, we may begin to regard this as a sufficient level of interaction and start avoiding real relationships. We learn to expect the same immediate, effortless responses from people and become disappointed by their slowness, imperfections and contradictions. Real people are complex, unpredictable and demanding, and they do not always respond as we would like.

One of Turkle's key concepts is friction. AI offers the illusion of intimacy and the fantasy of easy relationships without demands. In reality, the challenges of human conversation provide the foundation for the interaction and problem-solving skills needed for learning and the development of a professional identity. Our level of competence is defined in relation to other professionals in the field.

Another interesting concept is metacognitive laziness, introduced by Fan and fellow researchers in their 2025 study. It means that learners rely on AI so heavily that they outsource the regulation of their own thinking to it. AI handles the planning, monitors understanding and evaluates the results. The task is completed and the performance may look good, but little or no learning may actually take place.

The antidote is to keep metacognition in our own hands. We should assess and monitor for ourselves what we ultimately understand. In a university setting, teachers should independently assess what a student appears to understand. Sometimes this only becomes clear through conversation. A report alone may not reveal where or how AI has been used.

Dialogue Promotes Learning

Lifelong learning is based, among other things, on discussion, learning at work and thinking together. According to Turkle, generative AI offers artificial empathy and sufficiently good answers, which can lead people to avoid difficult conversations and lose their ability to engage in long, focused dialogues. This weakens our capacity to build knowledge together, tolerate disagreement and learn in social environments. Learning becomes more difficult when a learner constantly receives ready-made, excessively positive answers, does not practise evaluating sources, does not learn to assess arguments in the moment and cannot tolerate uncertainty or even anxiety. Learning, development and new competence often lie beyond the anxiety zone, outside our comfort zone.

I believe it could be useful to assess and develop competence by occasionally returning to the classroom to talk with one another, practise difficult interaction situations face to face, debate, learn how to disagree and even write answers by hand.

Lifelong learning requires sustained concentration, the ability to stay with difficult problems without constant interruptions and the capacity to construct complex chains of thought. It also involves social problem-solving in practical tasks. At its worst, AI weakens precisely these skills. Learning together requires effort, failure, tolerance of frustration and the ability to ask for, receive and give feedback without an AI-generated filter. Professional identity, ethical judgement and professional growth, for example, develop in relation to other people in workplaces and other communities. People also need other people in order to learn and grow into experts.

A Wise Strategy for Using AI in Learning

Experiments related to Kestin's work on mathematical problems have been conducted around the world, and results are beginning to emerge. For example, Kestin and colleagues found in a Harvard study that a good AI tutor guides students to think rather than merely giving them answers. An AI tutor should ask questions, provide hints, request justification and help students correct their thinking from their current level of understanding.

Such an agent is designed with pedagogy first. It requires teacher-defined objectives, tasks and guiding questions, a feedback logic and a staged learning structure that applies the principles of active learning. In the model studied, the tasks targeted levels such as understanding, applying and analysing in Bloom's taxonomy.

Teachers should therefore prevent AI from becoming a crutch. If students can use AI to bypass their own thinking, the learning benefit may disappear. A tutor should activate thinking rather than do the thinking for the student. A good AI agent provides continuous, personalised feedback and creates a learning situation in which students must think, apply and justify for themselves.

AI Traffic Lights in Teaching

At the University of Oulu, teachers may restrict or prohibit the use of AI on the basis of learning objectives, but they may also encourage or require its use. Students may also be required to explain which AI application they used and how they used it. The purpose of the AI traffic-light model is to make visible how AI may or must be used in a learning task.
An interesting option arises when the use of AI is a compulsory part of the task. The aim is then also to practise using AI appropriately, critically and responsibly.

For example, a marketing student could first write promotional copy for a new service independently and then create another version with AI. Finally, the student would compare the versions and explain which of the AI's suggestions they accepted, which they rejected and why.

In this case, AI does not do the learning for the student. The AI-generated output becomes an object of learning that the student must be able to evaluate, challenge and improve.

The purpose of AI is not to free us from thinking or from one another. We experience joy, fear, shame, love, disappointment and meaning, and trust between us grows from shared history, reciprocity and the knowledge that we matter to one another. A university teacher notices things in a student's gestures, presence and behaviour that data alone cannot reveal.

Instead of asking, “Should AI be used, and if so, how?”, it may be more useful to consider how AI can help us analyse feedback, prepare for and conduct difficult conversations, or identify different options. Building trust, genuinely encountering another person, making value-based choices and taking responsibility for the consequences of everything we do remain human responsibilities.

Tips:

  1. Try it yourself first. Complete the task or write a draft before using AI. Bringing AI into the process after you have started may be better than beginning with AI from the outset.
  2. Use AI as a tutor. Ask for hints, questions and help finding errors, but not for a finished solution. Do not automate the parts you need to practise, such as problem-solving, justification and writing in your own voice. Automate routine work instead.
  3. Test yourself. If you cannot explain the topic in your own words or complete the task without support, you have not yet learned it.
  4. Keep metacognition in your own hands. Assess and monitor your understanding, and direct your own work. Do not follow a bot's feedback blindly.
  5. Check the results. Language models make mistakes, and detecting them requires your own expertise.
  6. Nurture your relationships. Remember to call a friend, meet people face to face and make room for uncertainty and all kinds of conversations with real people.

Author: Päivi Lohikoski, Ph.D., Education Manager, University of Oulu Kerttu Saalasti Institute

Sources:

Turkle, S. (2026). Connected but alone. TED Talk Machines.

Boston Globe (15 October 2026): interview with Turkle; MIT News coverage of the same topic; NPR (2024): interview with Turkle.

Bastani, H., Bastani, O., Sungu, A., Ge, H., Kabakcı, Ö., & Mariman, R. (2025). Generative AI without guardrails can harm learning. PNAS, 122(26), e2422633122.

Fan, Y., et al. (2025). Beware of metacognitive laziness. British Journal of Educational Technology, 56(2), 489–530.

Kestin, G., et al. (2025). AI tutoring outperforms in-class active learning. Scientific Reports, 15, 17458.

Digital Promise (2025). Hearing from Students.

Photo: Minna Kilpeläinen

Created 17.11.2026 | Updated 17.11.2026