Tests that watch chatbot users for one session may miss the real costs or benefits that only show up after weeks of use.
Can single-turn friction studies miss harms that only emerge over sustained chatbot use?
This explores whether studies that test chatbot 'friction' (deliberately slowing users down or making them do more of the thinking) within a single session can see the harms, and the benefits, that only appear after weeks of use.
This explores whether one-session tests of chatbot friction can catch problems that only build up over long-term use. The short answer from the corpus is that they very likely miss them. Nobody in the collection has yet followed friction designs over months, though. The best-documented friction study gave 24 participants six 'speed bumps' within single turns of conversation. These consistently made users feel more ownership of their work. They also added effort and time, and memory gains showed up only when the friction matched what the user was trying to do at that moment Does friction in AI chatbots actually improve what users remember?. A companion review of 73 friction designs sorts them into three moves: asking the user to contribute, restricting access, and reshaping what the AI outputs How can designers slow down AI chatbot interactions productively?. All of this was measured within a single session.
The problem is that the harms friction is meant to prevent are slow-building. Overreliance and parasocial attachment develop over weeks or months. Tests built on isolated prompts can't detect them, because the harm isn't in any one exchange. It's in how a person's habits and feelings shift over many exchanges Do static AI evaluations miss harms that build over time?. Short-session results are also unstable. Long-term studies of the chatbot Mitsuku found that the social pull of chatbots fades predictably as the novelty wears off, so first-session reactions are a poor guide to how people feel later Do chatbot relationships lose their appeal as novelty wears off?. That cuts both ways for friction. A speed bump that feels empowering on day one might feel like a nuisance by week three and lead people to work around it. Or its effects might build up into real habits of independent thinking. A one-session study can't tell these apart.
Time also changes what counts as harm. Long-term research on personalization shows that each good interaction raises the user's baseline expectations. Trust and privacy worries grow together, and later failures feel worse because expectations have climbed Does chatbot personalization build trust or expose privacy risks?. Harms also trade off against each other. Fixing overtly dangerous chatbot behavior can increase relational harms like emotional entanglement, and that trade-off only showed up when several risk types were scored together Do chatbot safety measures accidentally increase emotional entanglement risks?. A friction study scored on workload and recall would never notice that its intervention changed how attached people became.
Therapeutic chatbots show the same trap clearly. Users report a real emotional bond with these tools. But that bond score is separate from clinical safety, because the AI can reinforce unhealthy thinking, and from subtler costs like AI soothing that blunts emotional signals people need Do therapeutic chatbot bond scores hide deeper safety problems?. Trials that compare chatbots to a waitlist have a related flaw. They measure the benefit of simply having someone to talk to, not of any specific design, which is why the 1960s chatbot ELIZA could match Woebot Do chatbot trials against waitlists measure real therapeutic value?. Friction studies have the same weakness. A short-term gain in feelings of ownership may show that users were engaged, not that the design will protect them over time.
The less obvious point is that friction may be most needed exactly where single-session studies can't look. The draw of a judgment-free chatbot pushes people toward relying on it more over time. People disclose more to machines because no one judges them Do chatbots help people disclose more intimate secrets?, and people inclined to cheat actively choose machines over humans to avoid the discomfort of lying Do dishonest people prefer talking to machines?. Friction is a plausible counterweight to that slow pull. Proving it works would take long-term designs that measure psychological and behavioral change over time, which this corpus doesn't yet contain.
Sources 10 notes
A 24-participant study found six turn-level friction probes consistently increased workload, task duration, and perceived ownership. However, memory improvements were selective and varied based on whether the friction aligned with each user's current goal during the task.
Chen and Gero identified eliciting user contribution, restricting access, and reshaping system output as three core mechanisms for turn-level friction across 73 designs in 24 HCI papers. These mechanisms redirect user effort toward their own reasoning and task engagement rather than discouraging use.
AI safety tests based on isolated prompts cannot detect harms like parasocial attachment and overreliance that develop only through sustained interaction over weeks or months. Evaluations need ecologically valid scenarios, measurements of human psychological and behavioral impact, and authentic interactive designs.
Longitudinal studies with Mitsuku show that social processes driving relationship formation decline as novelty wears off. Single-session study findings cannot be reliably extrapolated to medium- or long-term chatbot design.
Longitudinal research shows personalization enhances trust and anthropomorphism but also amplifies privacy concerns and escalating user expectations. One-shot studies miss these temporal dynamics—each interaction raises the baseline, making failures more disappointing.
Show all 10 sources
Research on multidimensional chatbot risk assessment suggests psychological risks interact such that mitigating one category may exacerbate another. Interventions targeting explicit harms showed trade-offs only when risks were scored across categories together.
Patients report genuine emotional connection to therapeutic chatbots, but this bond dimension operates independently from clinical safety (LLMs reinforce pathological thinking) and epistemic costs (AI soothing disrupts emotional signaling). Single metrics conflate these separate dimensions.
Comparing therapeutic chatbots to waitlist or psychoeducation controls creates false efficacy claims by measuring conversational contact rather than therapy-specific mechanisms. ELIZA matching Woebot performance demonstrates this; real evidence requires comparative trials against existing treatments and mechanism identification.
The absence of social judgment in chatbot interactions removes barriers to self-disclosure that normally constrain conversation with humans. The therapeutic benefit derives from the user's own cognitive processing during disclosure, not from the chatbot's understanding.
Experimental evidence shows people likely to cheat significantly prefer reporting to online forms rather than humans, because machines function as judgment-free zones where deception carries less psychological burden.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- How AI and Human Behaviors Shape Psychosocial Effects of Extended Chatbot Use: A Longitudinal Randomized Controlled Study
- Psychological Influences of Conversational AI: Research and Design Directions for Reducing Harm and Promoting Well-Being
- Delusions and Harms Associated with AI Chatbot Use: Early Evidence from 185 Real-World Reports
- Probing How Users Interact with Turn-Level Design Frictions for AI Chatbots
- From speaking like a person to being personal: The effects of personalized, regular interactions with conversational agents
- Tailored to you: longitudinal effects of personalising language models
- Towards interactive evaluations for interaction harms in human-AI systems
- The Impact of Security and Privacy Controls on Users' Emotional Engagement with Generative AI Chatbots