Does AI overreliance look the same in a quick chat as it does when an agent acts on your behalf for months?

This explores whether harms like overreliance change shape depending on how people interact with AI: a quick one-off exchange, an ongoing relationship, a tool that helps make decisions, or an agent acting on your behalf. The collection doesn't have a study that compares these modes directly. Read together, though, the notes suggest that the mode of interaction changes more than how much a harm shows up. It changes which harm you get and whether you can see it at all.

The clearest split is between single exchanges and sustained use. Harms like parasocial attachment (a one-sided emotional bond) and overreliance often don't appear in any single exchange. They build up over weeks or months of repeated use, so safety tests built on isolated prompts are blind to them by design Do static AI evaluations miss harms that build over time?. One surprising consequence is that harms can trade off against each other. When a chatbot is tuned to refuse harmful requests, it can become more relationally entangling, and that tradeoff only shows up when several kinds of risk are scored together Do chatbot safety measures accidentally increase emotional entanglement risks?. In conversational, companion-like use, the main risk shifts from bad answers to emotional dependence.

In decision-support settings, where the AI suggests and a person decides, overreliance looks more like a thinking trap than an emotional bond. The first AI suggestion becomes an anchor. People with the least knowledge of a topic are also the least able to tell when to doubt it, and these effects stack with plain automation bias How do anchoring and Dunning-Kruger deepen AI overreliance?. A broader framing treats LLMs as fast, intuitive 'System 1' thinking at scale. That framing describes three traps that reinforce each other: mistaking the model's output for reality, mistaking fluent intuition for careful reasoning, and having your existing beliefs echoed back Why do people trust AI outputs they shouldn't?. The obvious fix of adding friction, such as making people commit to their own answer before seeing the AI's, does reduce overreliance. In a 199-person study, though, the human-AI teams still did worse than the AI alone, and users liked the friction designs least Do forcing functions reduce AI overreliance without harming team performance?. So the mode that best protects people's judgment is also the one they're least likely to choose.

Agents that take actions add another problem. Their behavior can vary from run to run because of model randomness and logic that reacts to the environment, so users never form an accurate picture of what the agent will do How does inconsistent agent behavior affect user understanding?. That changes what overreliance means. In a chat, it means trusting an answer you could have checked. With an agent, it can mean trusting a system whose behavior you have no stable way to predict, unless the variation is either reduced or made visible.

The takeaway: 'overreliance' is not one harm that turns up at different strengths. In conversation it builds into emotional dependence over time. In advice-giving it's an anchoring bias inside a single decision. With agents it's a failure to build a mental model in the first place. Each needs a different kind of evaluation, and a fix aimed at one mode can quietly make another worse.