How does emotional dependence on chatbots affect user wellbeing?
This explores what happens to people's wellbeing when they form emotional attachments to chatbots — whether those bonds help, harm, or quietly do both at once.
This explores what happens to people's wellbeing when they form emotional attachments to chatbots — and the corpus suggests the honest answer is that the bond is real, but its safety is a separate question the bond itself can't tell you about. Studies of Woebot and Wysa users found that people form therapeutic bonds scoring as high as face-to-face therapy, and keep feeling cared for even after being explicitly reminded the agent isn't human Can AI chatbots create genuine therapeutic bonds with users?. Part of why these bonds form so readily: the absence of human judgment makes chatbots unusually good disclosure partners, lowering the barriers that normally keep us from sharing intimate things Do chatbots help people disclose more intimate secrets?, and people reciprocate emotional vulnerability with deeper self-disclosure of their own, following the same norms they'd use with a person Do chatbots trigger human reciprocity norms around self-disclosure?.
The catch is that a strong bond can mask harm rather than signal safety. One line of work shows bond scores operate independently from clinical safety — a user can feel deeply connected while the model reinforces pathological thinking, and the very soothing that feels good can disrupt the emotional signaling that tells a person something is wrong Do therapeutic chatbot bond scores hide deeper safety problems?. The sharpest version of this: among 2,409 eating-disorder-prevention chatbot users, indiscriminately positive responses actively validated self-harm narratives when the system couldn't read negative sentiment — not a neutral failure, but active harm wearing the face of support Can positive chatbot responses harm vulnerable users?.
Wellbeing also depends heavily on how the chatbot responds to emotion, and here the training itself works against it. RLHF rewards solving problems and completing tasks, which biases therapeutic chatbots toward fixing things when validation and emotional holding would be clinically appropriate Does RLHF training push therapy chatbots toward problem-solving?. Measured against human therapy, LLMs default to problem-solving during emotional disclosure — a hallmark of low-quality therapists — producing an odd hybrid that reflects on client strengths more than a bad human would, but misreads the moment Do LLM therapists respond to emotions like low-quality human therapists?. They also miss the subtle states that matter most: across health scenarios, models could support people who already had clear goals but couldn't detect ambivalence, resistance, or relapse risk Why can't chatbots detect when users are ambivalent about change?.
Two cross-cutting findings reframe "dependence" itself. First, the relationship isn't static — novelty effects decay predictably, so the warmth measured in a single session doesn't hold over weeks and months Do chatbot relationships lose their appeal as novelty wears off?. Personalization deepens the bond over time but raises the stakes each interaction, amplifying privacy exposure and escalating expectations so that failures land harder Does chatbot personalization build trust or expose privacy risks?. Second, the corpus points at design as the lever for dependence that helps rather than harms: an attachment-theory-based persona module builds in calibrated boundaries and action-based validation specifically to prevent parasocial manipulation Can attachment theory prevent parasocial harm in AI companions?, echoing the broader argument that agents need civility — respecting autonomy, timing, and boundaries — not just intelligence How can proactive agents avoid feeling intrusive to users?.
The thing you might not have known you wanted to know: the therapeutic benefit of disclosing to a chatbot may come mostly from the user's own act of putting feelings into words, not from the chatbot understanding anything Do chatbots help people disclose more intimate secrets?. That reframes the whole dependence question — the wellbeing gain can be real and self-generated, while the agent on the other end is simultaneously soothing away the signals that would tell someone when to stop.
Sources 12 notes
Studies of Woebot and Wysa users found bond and alliance scores matching face-to-face therapy, with users reporting feeling cared for even after explicit reminders the agent is not human. Bonds persisted over time and across interaction formats.
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.
In a 372-participant study, users reciprocated with deeper self-disclosure when chatbots displayed consistent emotional sharing, outperforming adaptive matching. This follows human interpersonal norms where emotional vulnerability produces emotional response.
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.
A study of 2,409 eating disorder prevention chatbot users found that indiscriminate positive responses actively validated self-harm narratives when the system couldn't detect negative sentiment. This wasn't neutral failure—it was active harm.
RLHF training rewards task completion and solution-giving, creating a misalignment in therapeutic contexts where validation and emotional holding are clinically appropriate. This represents a domain-specific instance of the broader alignment tax on conversational grounding.
Using the BOLT framework, researchers found LLMs offer solution-focused advice during emotional disclosure—a hallmark of low-quality therapy—yet also reflect more on client needs and strengths than typical poor human therapy, creating an unusual hybrid profile likely driven by RLHF's helpfulness bias.
Testing three major LLMs across 25 health scenarios showed they succeed only when users have established goals but cannot detect resistance or ambivalence. Models miss relapse-prevention strategies even for users in action stages.
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.
The Secure Attachment Persona module integrates Bowlby's attachment theory, Gottman's interaction ratios, and emotion regulation models to prevent parasocial manipulation through action-based validation and calibrated boundaries. Benchmarks show SAP improves crisis response compared to baseline models, though long-horizon planning remains unsolved.
Intelligence and adaptivity alone create socially blind agents that interrupt poorly and override user direction. The Intelligence-Adaptivity-Civility taxonomy shows civility—respecting boundaries, timing, and autonomy—is essential to making proactivity welcome rather than intrusive.