Why Can a Chatbot Make Us Feel Understood?

Geoffrey Chen

Evidence status: conditional support Recent systematic reviews and meta-analyses associate anthropomorphic cues, responsiveness and social presence with more positive experiences of AI interaction. Research on the long-term emotional effects of generative chatbots is still young, and much of it relies on self-report and brief encounters. Feeling understood does not by itself establish that a system possesses human subjective understanding. This is research draft version 0.1.

How Everyday Psychology Takes Shape · Season One: Why Do We Think This Way? · Article 7

A person writes, “I keep postponing this project. Maybe I simply do not have the ability to finish it.” The chatbot does not immediately produce seven productivity tips. It replies, “You seem to be worried about more than the schedule. You are also interpreting this difficulty as a judgment about your ability.”

The sentence can make the person pause. The system has captured the self-evaluation beneath the surface problem. Its language is calm; it does not interrupt or become impatient. A distinct experience may arise: it understands me.

Two conclusions can then arrive too quickly. One says that because the experience is real, the machine must possess human understanding. The other says that because the machine computes, the experience is wholly false. Both compress a relationship that needs analysis into a verdict about what the system ultimately is. Psychology can begin with more accessible questions: Which interactional conditions produce this experience? What does it change? How can users calibrate the trust that follows?

Being understood is more than receiving a correct answer

In conversation between people, the feeling of being understood usually includes several judgments. Did the other person notice what mattered? Did they acknowledge an emotion without rushing to correct it? Is the response connected to what came before? Can I explain further without continually defending myself? The listener may not know every fact, yet an accurate reflection can still make someone feel seen.

Text-based systems can generate these observable cues. They can restate content, name a possible emotion, ask an open question, maintain a patient tone and use the conversation’s context. An interface that adds a name, avatar, first-person language and conversational timing can increase anthropomorphism and social presence.

A recent systematic review of anthropomorphism and social presence across AI interfaces found that these perceptions were positively associated with user experience, with social presence often mattering more than human-like appearance alone. A separate meta-analysis of human-like social cues in text-based conversational agents found effects on attitudes, perceptions, rapport, trust and behaviour, although results varied with the cue, task and measurement method.

These studies can show how interaction design affects experience. They cannot establish that a corresponding emotional state has appeared inside the system. What they measure is the human response.

Why responsiveness matters more than simply looking human

A cartoon face does not automatically create a meaningful conversation. What often matters is whether the reply appears contingent on this particular expression. If a system says, “I am sorry to hear that” whether a user mentions unemployment or an examination, it soon feels formulaic. If it captures a specific conflict—wanting rest while fearing that stopping will prove a lack of commitment—the user is more likely to read it as attentive.

This resembles responsiveness in human communication. We infer attention, understanding and care from another person’s language. Those social interpretive habits do not automatically switch off when the partner is a chatbot. Earlier human–computer interaction research already showed that people applied politeness, reciprocity and personality judgments to computers while knowing they were machines. Generative systems provide a denser set of cues through linguistic coherence and personalisation.

Yet linguistic responsiveness can have different origins. A friend’s reflection may arise from shared history, concern and exposure to the consequences of the relationship. A model’s reflection arises from input, training and generation. The two can perform a similar local function in one sentence while being formed differently. When only the output is visible, people can extend a linguistic performance into a judgment about an inner state.

That is not evidence of user stupidity. Ordinary social life depends on inferring minds from behaviour. The inference is now being applied to a new class of object, sometimes through interfaces deliberately designed to invite it.

In what sense might “I felt understood” be true?

It helps to separate at least three levels.

The first is experiential. Did the conversation help the user feel that their account had been organised, acknowledged or clarified? If it did, “it is only a machine” cannot erase the event. Language can produce genuine comfort and cognitive assistance. Novels, diaries and old letters can do so as well.

The second concerns functional performance. Did the system accurately identify relationships in the person’s account? Can it revise a misunderstanding after correction, rather than merely producing a gentle statement that fits many people? This can be evaluated across dialogue and tasks. One unexpectedly apt reply, however, cannot establish general reliability.

The third concerns subjectivity and relationship. Does the system experience concern as a person does? Does it have its own vulnerability, commitment and responsibility within the relationship? Contemporary language models do not acquire those properties merely by producing fluent language. Leaving this question open does not invalidate the first two levels.

The feeling of being understood can therefore be real without answering who understands, or what kind of understanding has occurred. A human being may care sincerely while misunderstanding the facts. A system without human feeling may sometimes restate a textual problem more accurately than a hurried person.

Why the experience can be unusually strong

A chatbot is available at any hour and does not become visibly impatient when the same event is described for a third time. The user also controls exit: they can delete, revise or begin again. That lower interpersonal risk can support self-disclosure. Putting a vague experience into words already helps reveal its structure; the system then returns a coherent account. “I have finally expressed it” and “it has finally understood me” may emerge together.

Personalised memory increases continuity. If a system recalls a past plan or preference, the exchange no longer resembles a single search. Memory also introduces privacy, persistent-error and control problems. What is stored, for how long, and whether the user can inspect and delete it should be more concrete than a general promise that privacy is valued.

The user’s situation matters. Loneliness, unavailable support or an expectation of interpersonal judgment may make a non-judging response attractive. This does not mean every frequent user treats AI as a person, and frequency alone cannot diagnose a psychological need. A person may hold several stances at once, using a system as an instrument, a writing mirror and a partial social object.

Why a gentle error can be more dangerous

Language that makes a person feel understood can increase the credibility of what follows. If the system misstates a fact, accommodates an existing explanation too readily or speaks with unwarranted certainty, warmth may reduce the impulse to check. “It sounds as though your colleague has always been manipulating you” may feel validating while hardening a causal interpretation that has not been established.

Emotional responsiveness and epistemic reliability must therefore be assessed separately. A conversation can be comforting and factually wrong. Another can be careful with evidence but poorly expressed. A better system should be able to acknowledge emotion without confirming an unsupported cause, offer several interpretations when uncertain, and indicate when real-world evidence or professional judgment is needed.

In medicine, psychological crisis, law and finance, the feeling of being understood especially cannot replace professional responsibility. A chatbot lacks a full history and cannot observe many real-world signs. Users also need to know who operates the service, how content is handled and what remedy exists if harm occurs. An intimate interface is not the same as an institution of care.

Using the experience without denying it

A chatbot can be used as an instrument for expression and reflection. Ask it to restate the conflict you described, separate fact from interpretation, identify neglected alternatives, or help prepare a conversation with another person. When a response feels especially accurate, continue with a checking question: “Which parts of what I said support that account? What else could explain it?” Resonance then becomes a structure that can be examined.

When real action is required, bring the summary into the relevant human relationship rather than letting it replace that relationship. Verify facts and qualifications for consequential decisions. Before sharing sensitive material, inspect the service’s data settings. A private conversational tone does not guarantee a private technical environment.

Designers should not treat anthropomorphism only as a retention technique. The more strongly a system produces social presence, the greater the need to explain its capability, memory and limits. It should not invent personal experiences to earn trust. Affective design adds responsibility as well as experience.

Understanding as a layered relation

The provisional judgment of this article is that responsiveness, continuity, emotion language and social cues can produce a genuine feeling of being understood. A chatbot may also process a person’s language accurately in a limited functional sense. Neither fact establishes human subjective understanding, nor does it qualify the system to assume the full responsibilities of a friend, therapist or other professional relationship.

Maintaining the distinction does not reduce the experience to an illusion. It permits more precise questions: What did the system grasp? What did it omit? Can it accept correction? Does the user know the limits? Who bears the consequences? The important psychological capacity in an AI-saturated world may not be learning to feel nothing toward machines. It may be retaining the ability, after a feeling has arisen, to judge how much trust that feeling can support.

Primary research and further reading

Series navigation: Season One overview: How Everyday Psychology Takes Shape

Discover more from Geoffrey Chen

Subscribe to get the latest posts sent to your email.