Today I talked to a robot. I asked it to help me think through a solution for a project, borrowing its creativity when mine was running short. It also helped me summarize points in a document and prepare a presentation. Later, I turned to it for help with an accounting question. It's also been my companion for reorganizing my training routine and picking my next travel destination. The uses I'm describing here, simple and everyday, of language models (LLMs, a type of artificial intelligence) are practically ubiquitous in modern life. But in an age of artificial intelligence, what risks do these tools pose to our own intelligence?
We recently watched Dr. Allison Koenecke's presentation on auditing speech-to-text transcription models — a relatively simple tool, but with a lot to unpack. The central point was that transcription "success" could lead to failures by ignoring the end use. As a key example: a doctor or speech therapist analyzing an AI-transcribed speech sample would never be able to diagnose a speech disorder! That's because all the best current models strip out "imperfections" and markers of spoken language, like hesitations ("um" and "uh"). The success of an AI use case is relative to context.
Once the tool is built and ready, we can call in as many speech doctors and therapists as we want — they won't be much help at that point. The technology's usefulness is limited by not incorporating expert perspective on its use during its construction. Something similar happens across the AI industry, when the human element of interacting with these tools isn't taken into account. We expect technological advances to produce certain effects on the user, like learning and productivity gains, and to avoid negative impacts like dependency and distress. And yet behavioral experts are typically brought in only at the very end of development, once the product already exists, to mitigate such unwanted effects.
This perspective is laid out by influential behavioral science researchers in a recent open letter position paper on AI (Sacher et al., 2026). If we're building technologies that, among their functions, change behavior, it's reckless to define their success metrics and design them without user-focused behavioral scientists at the table. More specifically, the researchers argue that while AI development has been extensively debated and includes metrics for bias, technical accuracy, privacy, and data security, the systematic behavioral impacts on users are neglected, rarely evaluated or regulated. The authors contend that AI doesn't just complete tasks or provide information — it actively shapes how people think, feel, and make decisions. As a result, behavioral risks tend to be identified late, if at all. Beyond response accuracy or user satisfaction, understanding the cognitive and relational impact of AI systems matters enormously from a social perspective.
Even technically flawless systems can produce harmful consequences — like emotional dependency or reduced human trust — that are rarely assessed in a structured way before large-scale rollout. There are even cases of suicide linked to AI use and romantic relationships between humans and AI companions.
In this context, the authors propose the concept of behavioral safety as "the extent to which an AI system avoids causing predictable harm to people's decision-making, motivation, emotional wellbeing, sense of autonomy, or help-seeking behavior over time." The open letter ultimately calls for behavioral science to be built into the AI development, testing, and deployment cycle as core infrastructure — not just for ethical reviews once a system is already in wide use — with behavioral scientists on product teams, dedicated funding for monitoring long-term impacts, and evaluation frameworks focused on the "psychological competence" of AI systems. The end goal is AI that's not just effective at our tasks, but safe and trustworthy in its long-term effects on us.
When people talk about LLMs, the case for using them usually comes down to remarkable productivity gains: everything gets done in a fraction of the previous time. Writing a piece of text, summarizing content, coming up with an idea. But what's the impact of that apparent gain?
The fundamental appeal of LLMs is the removal of friction from a wide range of activities. Friction can be understood as the resistance or difficulty we encounter while doing a task: the effort of structuring a thought, the frustration of not finding the right word, the fatigue of summarizing a dense report. AI promises — and delivers — a low-resistance world. But behavioral science has long known that friction can play an important role, and removing it indiscriminately can cause harm, especially in contexts that require self-control. For example, one study found that giving people larger popcorn buckets — reducing the friction of accessing the food — made them eat more regardless of enjoyment, even when the popcorn was stale and unappetizing (Wansink & Kim, 2005). Along the same lines, positive friction is the deliberate introduction of an obstacle or extra step into a process, to prompt reflection, reduce impulsive behavior, and lead to decisions more aligned with a person's actual interests. One of the most-studied examples is the use of "cooling-off periods" — like a tool that analyzes the emotional content of an email before sending and, if it detects a high level of anger, holds the message for five minutes before notifying the user it will go out a minute later, giving a window for reconsideration. In the same vein, a 2024 paper by Chen and Schmidt makes the case for using positive friction in human-AI interaction specifically. In it, they propose a model of the contexts in which such friction can benefit both users and developers: (i) self-control situations, (ii) interrupting automatic behaviors to activate attention, (iii) prompting action, and (iv) deprioritizing efficiency in favor of divergent thinking.
Friction, especially when we're talking about cognition and productivity, isn't a flaw in the process, nor is it necessarily a problem. More than that, it's sometimes the very point of engaging in an activity. Learning science has talked for decades about "desirable difficulty": the principle that certain levels of challenge are necessary for memory and understanding to consolidate, and for new skills to develop. Just as with the transcription example above, automating a student's writing or analysis process through a chatbot can destroy the very opportunity for transformation that the student would need to process in order to learn. In that case, reducing friction blocks growth. For instance, at least one pre-print study has already shown that people who learn to perform a task without AI assistance give up less often on similar tasks, when tested without permission to use the technology (Liu et al., 2026).
To navigate this new era, we need to establish a "functional audit" of our AI use. The criterion for deciding the risk of delegating a task to an LLM should depend on the answer to a simple question: what is the primary function of this activity? If a task's function is purely its end product — a routine email, formatting a table, meeting minutes for the record — delegation carries low risk. Here, friction doesn't serve us. But if the activity's function is to transform the person doing it — writing an academic paper, deliberating over an ethical dilemma — delegation becomes irrational. The goal isn't the finished text, but the new neural connections, the writing practice, and the clarity of thought that only emerge through effort. If the goal is to learn or understand something, the process and the effort are crucial. After all, it's a mistake to expect that a set of experiences based on supervising what a chatbot wrote will produce an intellectual change similar to the experience of writing from scratch. The effort of searching, the doubt, and working through cognitive impasses are the very components that shape the thinker. When we replace the creative process with a curatorial one, we're changing the nature of the experience — and, as a result, we should expect to learn, at best, to become good curators.
The risk analysis isn't limited to using chatbots for productivity gains. Another central strand is the dynamics of relationships. Another recent position paper (Zohar et al., 2026) demonstrated the central and transformative role friction plays in human relationships. For example, if we soothe our loneliness with an artificial "companion" that agrees with everything and never challenges us, we remove the friction from real human relationships. After all, interacting with another person carries risk and cost. But in doing so, we lose the incentive to seek out complex connections that, while difficult, have the potential to make us better people — a very different benefit from the instant comfort of a chatbot's reply. The social friction of a conflict with a friend is often exactly what signals the need for change or repair. The friction of a political debate is a core part of building solutions to social problems.
So the conclusion for developing and using these technologies is clear: understanding the function of the effort and friction we put into different activities matters more than ever. If language models start offering a short-term advantage to those who avoid the friction necessary for development, we risk institutionalizing perverse incentives. This kind of insight isn't the end of a behavioral scientist's work — it's where it begins. Plenty of questions follow: in which contexts should engagement with technology be made easier? In which should it be made harder? How do we build contexts where workers are more productive in the long run? How do we design products that minimize the risk of psychological or intellectual dependency in their users? How do we equip students, professionals, and society as a whole to know when — and when not — to turn to AI? In some contexts, a chatbot's success should be measured by how much friction it manages to preserve, to make sure the user is still exercising their own critical capacities.
Here at the Brazilian Institute of Behavioral Sciences, beyond experimenting with these tools to improve our own processes and services, we're committed to investigating the human relationship with new technologies and to designing tools and solutions that preserve users' cognitive skills and decision-making autonomy amid drastic technological change. We intend to return to this cross-cutting AI agenda regularly, in response to what we're observing in the market, in society, and in the projects we're involved in.
Sacher, P. M., Michie, S., Hauser, O. P., et al. (2026). The missing discipline in AI: a call for behavioural science. Wellcome Open Research, 11:152.
Zohar, E., Bloom, P., & Inzlicht, M. (2026). Against frictionless AI. Communications Psychology, 4, 39.
Wansink, B., & Kim, J. (2005). Bad popcorn in big buckets: Portion size can influence intake as much as taste. Journal of Nutrition Education and Behavior, 37(5), 242–245.
Munoz, M., Nogueron, A., & Neira-Tovar, L. (2012). AngryEmail? Emotion-Based E-mail Tool Adaptation. In Computers Helping People with Special Needs, LNCS 7383. Springer.
Chen, Z., & Schmidt, R. (2024). Exploring a Behavioral Model of "Positive Friction" in Human-AI Interaction. In Marcus, A., Rosenzweig, E., & Soares, M. M. (eds), Design, User Experience, and Usability. HCII 2024, LNCS vol. 14713. Springer, Cham.
Liu, G., Christian, B., Dumbalska, T., Bakker, M. A., & Dubey, R. (2026). AI Assistance Reduces Persistence and Hurts Independent Performance. arXiv:2604.04721.
Enjoyed this? Subscribe to Café Comportamental for weekly essays in Portuguese, or explore more in Resources.