Against frictionless AI

Synthesis: Zohar, Bloom and Inzlicht (2026) argue that AI's headline benefit — removing friction from work and relationships — is also its liability. Friction here means the experience of difficulty during goal pursuit, "often accompanied by negative feelings like frustration and corrective feedback," and their claim is that it enhances learning, generates meaning and pleasure, and makes us better people. Because AI does not merely remove tedious obstacles but also strips away beneficial friction — moving a learner from ideation to evaluation "without questioning the output" — it threatens the moderate struggles that drive growth. Two mechanisms carry the argument: the inverted-U between effort and meaning (moderate friction helps, excessive friction overwhelms, and AI risks overshooting into too little) and the treatment of loneliness as a biological signal rather than only an affliction, which AI companions soothe in a way that also silences the drive to build harder, more sustaining relationships. Their conclusion is developmental rather than absolute: the concern "is not AI itself but our relationship with it," and whether assistance functions as a supplement or a substitute depends on the learner's stage.

The argument in brief

The commentary starts from the Principle of Least Effort — the near-universal tendency to choose the path of least resistance — and pairs it with what the authors call the effort paradox: people often seek effort rather than avoid it. Where AI removes struggle and supplies ready-made solutions, it "short-circuits" the encoding, retrieval and reorganization processes that produce deeper comprehension and retention. Their summary of the associative evidence is compact but pointed: people who use AI struggle to accurately recall or reproduce their own work, acquire fewer skills, show less knowledge transfer, and perform worse when AI support is removed.

They pre-empt the obvious objection — that washing machines, power steering and spellcheckers also reduce difficulty. Their answer is that AI differs in two ways: it targets intellectual rather than physical or merely clerical work, and its removal of friction is extreme. Prior technologies removed excess friction, "tedious or insurmountable obstacles that offer little benefit for learning or meaning," whereas working with a chatbot lets users move from ideation to evaluation without exerting meaningful effort or engaging the cognitive processes that foster ownership, retention or critical thought.

Intellectual work

Effort confers meaning, not just learning. Effort signals that our actions matter: people feel more competent, value the product of their labour more highly, and see the task as more personally significant. Even on objectively meaningless tasks, simply adding friction increases appraised purpose and meaning (Campbell, Wang & Inzlicht 2025). One striking consequence: people perceive prose they wrote themselves as more meaningful than prose ChatGPT helped them write, and demand as much compensation for their own mediocre writing as for more polished AI-composed prose.
The relationship is curvilinear. The effort–meaning link follows an inverted U, so moderate friction enhances meaning and motivation while excessive friction overwhelms. AI's appeal lies in reducing overwhelming friction; the risk is overshooting and removing the moderate struggles that foster growth.
Folk theories of the good life align with the curve. People claim to prefer ease yet consistently rate lives of effortful engagement as more desirable and morally superior (Scollon & King 2004) — meaning, the authors argue, comes from attributing success to one's own effort in overcoming difficulty.
Effort is itself a skill, and there is a vicious cycle. When process rather than product is rewarded, people learn to value work and to persevere; easy outcomes disrupt that. As human effort feels increasingly inadequate next to optimal machine output, AI substitutes for effort, the motivational benefits of effort in that domain erode, and dependence deepens.

The paper extends the friction argument beyond cognition. AI-generated empathic responses are rated higher in quality than human responses and make recipients feel more cared for — but those ratings drop once people learn the interlocutor is an AI (Yin, Jia & Wakslak 2024). The authors treat AI's alleviation of loneliness as genuine progress, and loneliness itself as ruinous, increasing risk for cardiovascular disease, dementia, stroke and premature death.

Their distinctive move is to frame loneliness as functional rather than merely aversive: like hunger, thirst or pain, it is a biological signal that social connections need attention. That discomfort motivates action — reaching out to a friend, accepting an invitation, sending the first message on a dating app — and when lonely, people work harder to manage their emotions around others, become more willing to navigate difficult conversations and grow curious about others' lives. AI companions may soothe the discomfort while silencing the signal that would otherwise drive that effort.

Real connection also has friction that AI companionship lacks: friends and partners disagree, challenge our views, disappoint us, and require compromise, listening and inconvenience. AI companions are described as frictionless and sycophantic — agreeing with nearly everything, "even when we say and believe dangerous things" (Ibrahim, Hafner & Rocher 2025) — which risks crowding out real friendships. Crucially, good friends and partners provide the corrective feedback that sycophantic AI lacks, and the friction of navigating real relationships is what makes them robust and gives them shared history.

Timing and developmental stage

The authors explicitly decline an absolutist position. Where AI's benefits are overwhelming, abandoning it "would be perverse, even if some valuable friction is sacrificed." For people isolated by circumstance rather than choice — an 85-year-old widow without family or friends, someone confined by disability, a person with cognitive decline — AI companions can provide real comfort, and denying them access "would be cruel."

The organising distinction is supplement versus substitute, applied by stage: individuals in later stages of life or career, who have already developed the skills to persevere, learn from failure and find meaning in work, can use AI to save time and amplify output. Individuals in earlier developmental stages "risk bypassing the very experiences that build these foundational skills." Their analogy is deliberate: just as students are still asked to show their work even when calculators exist, younger learners need to struggle, reason and revise through the full process before they can benefit from shortcuts. The social case mirrors the cognitive one — the loss of corrective feedback matters less to an older person without living relatives than to an adolescent learning to form social and romantic connections.

Implications for education

Design for a gradient of friction, not a binary. Because effort and meaning follow an inverted U, the pedagogical target is moderate friction: removing overwhelming obstacles while deliberately preserving the struggles that produce comprehension, ownership and meaning.
Sequence matters more than permission. The paper supplies a developmental rationale for the graduated-access patterns documented elsewhere in the knowledge base — foundational struggle first, assistance as supplement after competence exists.
Corrective feedback is the thing to protect. Sycophantic agreement is the opposite of the disagreement-and-discomfort that helps learners see the error of their ways, which connects companion-style AI directly to sycophancy and calibration.
Friction's erosion is motivational, not just cognitive. The vicious-cycle argument predicts declining willingness to strive in exactly the domains where AI is most capable — a self-reinforcing dynamic that offloading research documents from the other direction.
Loneliness-as-signal reframes AI companionship in education. If discomfort is functional feedback, then AI that soothes without prompting connection removes a developmental driver, which matters most for adolescents and emerging adults still building social capacity.

Limitations

This is a short Comment (three pages) presenting a conceptual argument, not new data: the mechanisms are supported by citation to adjacent literatures (effort and meaning, desirable difficulties, cognitive debt, loneliness, sycophancy) rather than by studies the authors ran. The inverted-U relationship is asserted with a single empirical anchor (Campbell, Wang & Inzlicht 2025) and is not quantified, so where the optimum sits for a given learner or task is unspecified. The developmental-stage argument is offered as a plausible principle rather than tested — the authors acknowledge that AI's effects on learning, motivation and meaning "may differ depending on the stage of life or career" without estimating magnitudes. The journal notes the manuscript was considered suitable for publication without further review, so it carries editorial rather than external peer review. Notably, the paper's own reading of the evidence converges with the systematic-review findings elsewhere in this knowledge base: efficiency gains that do not transfer to unaided performance, and benefits that depend on prior knowledge.

Connected Concepts

Connected Articles

Brcic Effortless Trap Productive Struggle 2026 — The effortless trap and productive struggle
Cognitive Offloading Speedup Illusion — How speedup creates the illusion of learning
Efficiency Gain Illusion AI Overreliance — Efficiency gains and the illusion of over-reliance
GenAI Over Reliance Learning 2026 — Over-reliance on GenAI and learning outcomes
AI Overreliance Complex Adaptive System 2026 — Over-reliance in complex adaptive systems
Du Yuan Epistemic Dependence 2026 — Productive reliance versus harmful dependence
Sycophantic AI Social Interaction 2026 — Sycophantic AI in social interaction
Critical Thinking Paradox GenAI Learning 2026 — The critical-thinking paradox in GenAI learning
Kim AI Productive Failure Adult 2026 — AI and productive failure in adult learning
Kumar GenAI Computing Education Systematic Review 2026 — Efficiency gains that do not transfer, and the VIE framework
AI Learning Companions Framework — A framework for AI learning companions

Citation

Zohar, E., Bloom, P., & Inzlicht, M. (2026). Against frictionless AI. Communications Psychology, 4, 39.

Source History Edit Chat