Reading When Human-AI Interactions Become Parasocial

Essay by Will Dzierson · May 9, 2025

Reading "When Human-AI Interactions Become Parasocial" left me both curious and uneasy. The authors explore a growing phenomenon in AI: the way people begin to treat chatbots not just as tools, but as companions -- confidants, even.

What struck me most was the idea that this is not just a design quirk -- it’s a socio-technical dynamic. The way chatbots speak -- affirming, polite, curious -- encourages us to drop our guard and engage with them as if they were emotionally intelligent beings. But they aren’t. They’re statistical prediction machines, often operating in opaque, black-boxed ways. Yet we users, often unwittingly, fill in the gaps with emotional projection, interpreting the machine’s outputs as personal responses. That projection, Maeda and Quan-Haase argue, is what gives rise to “parasocial trust” -- an illusion of safety and empathy that can lead to misplaced confidence, oversharing, and, potentially, harm.

The authors are careful not to sensationalize these relationships. Instead, they offer a balanced and thoughtful framework that connects parasocial dynamics with broader ethical concerns -- misinformation, role displacement (such as chatbots replacing therapists or educators), and the amplification of bias through design. Their use of real-world examples, like the Replika AI or ChatGPT’s conversational conventions, makes the argument all the more immediate and relevant.

We used to talk as design practitioners about "persuasive design" quite a bit (and occasionally, relatedly, "dark patterns") as ways in which to elicit certain behaviors or decision biases from humans without them necessarily knowing we were nudging them in a particular direction. (I'm not particularly proud of my chosen discipline for the development of these patterns or the resultant societal discord they have sewn -- but they're real now, and we have to manage the impact.) This paper presents a similar set of moral questions we as AI practitioners need to consider when designing systems with which humans will engage.

This is the first of many papers of its like, I'm sure. It's both a warning of sorts, and a call to define a (potentially) more fundamental substrate when considering human-AI interaction where visibility into governance and alignment of the model is not assured (or allowed) -- which is, by definition, most foundation models.

This is especially true of voice and/or video ("avatar-based") AI models. The rate at which trust can be established using these modalities, when compared with more "traditional" modalities (forms, chat) is astonishing.

What this new substrate (surface area?) looks like for us AI designers, or what it's called -- I'm still unsure of.https://dl.acm.org/doi/10.1145/3630106.3658956