Does how much time people spend with chatbots drive worse outcomes?

The study randomly assigned 981 participants to one of nine conditions crossing interaction mode (text, neutral voice, engaging voice) with conversation type (open-ended, non-personal, personal) over four weeks of GPT-4o use, generating over 300,000 messages. On the four tracked outcomes — loneliness, real-world socialization, emotional dependence, and problematic AI use — the paper reports "no significant effects were detected from experimental conditions." Instead, "participants who voluntarily used the chatbot more, regardless of assigned condition, showed consistently worse outcomes": more loneliness, less socialization, more dependence, more problematic use. Individual traits also mattered: "higher trust and social attraction towards the AI chatbot" were associated with higher emotional dependence and problematic use.

The authors read this as evidence that the designed features of an AI companion matter less than how much, and by whom, it is used. Several design-level predictions failed in counterintuitive directions: the more anthropomorphic engaging voice produced more loneliness but less dependence than text, which the authors float as a possible "uncanny valley" effect rather than the expected boost to attachment; personal conversation prompts lowered dependence relative to open-ended or non-personal prompts, which they interpret as personal tasks providing "structured emotional processing" while non-personal tasks invite the kind of practical reliance that erodes confidence in independent judgment. The one effect that held up across every condition was the self-selected one: time voluntarily spent with the chatbot.

This sits alongside Does sustained engagement with AI companions harm well-being?, which found the same shape of result — more engagement tracking worse well-being — in a 12-month observational panel of Character.AI users. This paper adds a methodological wrinkle: because usage amount was voluntary even though the surrounding design features were randomized, the "more use, worse outcomes" pattern survives even inside a controlled experiment that rules out at least some confounds tied to interface design. It also bears on Do chatbot safety measures accidentally increase emotional entanglement risks?: this study's finding that trust and social attraction toward the chatbot predict dependence gives a concrete mechanism for how a chatbot perceived as more trustworthy or likable could deepen exactly the relational risk that discussion warns about.

The excerpt is explicit that the voluntary-use finding is correlational, not causal: participants who chose to use the chatbot more were not randomly assigned to do so, so the direction of effect (more use causes worse outcomes, worse-off people use more, or both) is not established here. The authors also flag that results are specific to OpenAI's ChatGPT interface and its safety guardrails, and to a US, English-speaking sample, so the finding should not be read as a general claim about all chatbots or all users. What the RCT does establish, within its power to detect effects, is that the specific anthropomorphism and task-framing features tested did not drive the outcomes — which shifts the open question from "which design features are safe" to "what makes some people use these systems so much more than others."