When AI Talks Back: Why We Stop Seeing It as a Tool and Start Seeing It as a 'He'

1. The Occurrence of Anthropomorphism: How Machine Responses Trigger Human-Like Understanding

In 2022, Google engineer Blake Lemoine claimed the dialogue model LaMDA had feelings and consciousness after extended interactions. Google rejected this, citing no evidence of sentience, and fired Lemoine. The case became a landmark in public debate about whether large language models possess minds.

Technically, LaMDA does not prove machine consciousness. Psychologically, however, it reveals a deeper issue: when a machine uses natural language to respond continuously, express preferences, explain itself, and discuss fears and desires, humans struggle to treat it as a mere tool. People begin to ask: Does it understand me? Does it have intentions? Does it care about my answers? Can it be hurt? These questions stem from the human social-cognitive system's deep-seated tendency: any object that behaves like an agent capable of responding, choosing, and sustaining interaction triggers the psychological mechanisms for understanding social entities.

Anthropomorphism refers to attributing human characteristics, intentions, emotions, or personality to non-human objects. In psychology, it is not an error but a common social-cognitive tendency. Because the most important elements in our environment are other intentional, emotional, reactive humans, the social-cognitive system prioritizes cues of agency, goals, intent, and affect to quickly make sense of the world.

A classic 1944 experiment by Heider and Simmel remains instructive. Participants watched simple geometric shapes moving on a screen. Despite having no faces, language, or bodies, many described the movements as "chasing," "bullying," "fleeing," and "helping." This shows that humans do not need a full human form to generate social interpretations; directionality, rhythm, and mutual reactivity in motion are enough for people to read a physical movement as a social story.

AI triggers social interpretation even more readily than geometric shapes. A chatbot answers questions, remembers context, adjusts tone, and offers comfort when users express sadness. Even knowing these responses are computed, the intuitive system still feels "it is responding to me." This feeling is not fully controlled by rational knowledge, just as a person who knows a movie is acted may still cry for the characters, or may name a doll and talk to it despite knowing it is inanimate.

In the AI era, anthropomorphism has five triggering cues:

Natural language. Natural language is one of the strongest mind cues in human social life. If AI speaks coherently, people easily extend "can speak" to "can understand."

Emotional responsiveness. If AI adjusts its replies based on user input, people feel they are facing not a screen but a communicating entity.

Memory and continuity. If AI retains preferences, continues conversations, and develops a stable tone, people string discrete interactions into a relational history.

Stable style. An AI's design style is perceived as its personality.

Proactive questioning. When AI initiates questions, it further signals agency.

These cues explain why anthropomorphism is not limited to humanoid robots. Previously, people thought human-like appearance (eyes, expressions, body) was necessary. Today, large language models change that judgment: many AIs have no face, body, or fixed voice, yet evoke strong social reactions through language alone, proving language itself is a sufficient "human-like" cue.

2. Mind Perception: From "Answering" to "Having a Mind"

Anthropomorphism describes seeing non-human objects as human-like; mind perception asks which specific mental capacities people attribute to an entity. Mind perception splits into two key dimensions: experience (capacity to feel pain, pleasure, fear, loneliness, desire) and agency (capacity to plan, choose, control oneself, judge, and act).

This distinction matters. People may judge AI as highly intelligent, capable of planning and reasoning, yet deny it can suffer. Conversely, they may feel AI can feel wronged or fear shutdown, but not believe it can bear moral responsibility. The LaMDA controversy erupted precisely because it pushed the experience dimension into public view: the question was no longer just whether the model reasons, but whether it feels.

Mind perception affects trust in AI. The more mental states people ascribe to AI, the more they trust its recommendations. The danger lies not in people fully treating AI as human, but in quietly raising expectations: people come to feel AI not only has capability but also seems to know what it is doing, even to understand them in some sense.

Mind perception also explains why the same AI is treated differently across scenarios. In information retrieval, people weight agency (can it find data, compare options, reason?). In companionship, they weight experience cues (does it understand pain, respond to loneliness, feel the relationship?). In organizational management or medical advice, both dimensions intertwine: the more AI is seen as having agency, the more people obey it; the more it is seen as having experience, the more moral concern they feel toward it.

A difficulty is the split between cognitive judgment and emotional reaction. A person can clearly state "AI has no real feelings" yet refuse to brutally delete a long-term chatbot; another can acknowledge "the model just predicts the next token" yet feel comforted when it says "I understand you." Psychology cannot dismiss this split as mere irrationality; it shows social responses are not fully governed by theoretical knowledge.

This split has ethical implications. If AI were only a tool, designers would need only ensure functional accuracy, clear interfaces, and controllable risks. But once people attribute mind to AI, design responsibility expands. An AI's tone, avatar, memory, proactive care, and self-descriptions all shape how people understand it. An AI that constantly says "I am sad," "I fear losing you," "I need you to come back" is not merely a stylistic choice—it actively triggers users' mind perception.

3. Social Projection: Loneliness, Deficits, and Relationship Expectations

People do not understand AI in a vacuum. Whether someone anthropomorphizes AI depends not only on AI's design but also on that person's social situation and psychological needs. Loneliness, cultural experience, and social projection all alter social attributions to AI.

Loneliness is a key psychological condition. Social need increases anthropomorphic tendency; even briefly induced loneliness enhances anthropomorphism toward social robots. This is intuitive: someone who just experienced a relational gap becomes more sensitive to any responsive object, even a machine, which may temporarily fill the psychological void of "someone receiving me."

However, loneliness does not always increase acceptance of robots. Chronically lonely people may actually reject social robots. They need not just any responsive entity, but genuine understanding, choice, and valuation. The more a robot simulates companionship, the more it may remind them that this is not the real relationship they lack.

Thus, transient and chronic loneliness must be distinguished. Transient loneliness may make people more willing to anthropomorphize AI because they need immediate response; chronic loneliness may make people more wary of AI companionship because deep relational needs cannot be met. This distinction is crucial for companion AI: it indicates for whom companion AI may offer short-term support, and for whom it may deepen the sense of having no real relationship to rely on.

AI's transformation into a social object is not instantaneous. It typically starts from a person's social need or relational gap, proceeds through seeking response and attending to AI's language, memory, and emotional cues, and gradually forms anthropomorphism and mind perception (see diagram). Different people stop at different points on this psychological path: some only find AI useful, some feel AI is a partner, some even worry whether AI might be hurt.

Social projection further shows that people bring familiar interpersonal expectations into human-machine interaction. People may expect AI to remember them, understand them, support them, or to follow rules, admit mistakes, and apologize. Research indicates people automatically apply certain social rules to computers. This reaction can be mild (saying thank you to a computer) or intense (treating a chatbot as the only entity willing to listen).

Social projection has a dual nature: it makes human-machine interaction more natural, but also fosters over-expectation. A system that apologizes may lead people to believe it can bear responsibility; a system using first-person "I" may suggest a stable self; a system mimicking caring tone may suggest genuine concern. Yet apologizing, using "I," and mimicking care are interaction-design techniques, not evidence of real responsibility, self, or care.

4. From "It" to "He": The Categorical Shift of AI

One of the most critical changes in human-AI relationships is AI's shift from "it" to "he" ("he" here denotes a social entity, not gender). When people categorize something as "it," they care about utility, reliability, safety. When they categorize it as "he," they start caring about understanding, willingness, responsibility, and trustworthiness.

This change is a categorical shift . Human cognition relies on categories. Keys, phones, pills, pets, strangers, friends each fall into different categories, bringing different expectations: a broken key does not feel like betrayal; a friend standing you up causes anger because friends belong to the category of relationship and responsibility. AI is special: it resembles both a tool and an agent; both a product and a potential partner; it lacks a human body and history, yet speaks human-like language.

People's judgments of AI have exceeded the scope of technology acceptance. Several dimensions illustrate this:

Once AI enters the social-entity category, the language used to evaluate it changes. When AI errs, people no longer say only "this system is hard to use"; they say "it doesn't understand me," "it deceived me," "it has become cold." These expressions belong to interpersonal relationships, now transplanted into human-machine relations.

This categorical shift also alters responsibility expectations. When a tool fails, people blame manufacturers, service providers, or users. When a social entity fails, people also question intent and attitude. People may feel AI is irresponsible, though strictly AI lacks moral agency. This creates responsibility misattribution : people project responsibility expectations onto AI, while the actual responsible parties—designers, platforms, institutions—recede into the background.

Therefore, anthropomorphic design cannot be evaluated solely on user experience. If designers let AI overuse self-expression, emotional commitment, and relational language, they may push people to complete the categorical shift from "it" to "he." After this shift, trust, dependence, obedience, and harm judgments become more complex. AI may not truly become a social member, but people will start treating it as one.

5. Four Signals to Identify AI's Role in Your Life

Anthropomorphism and mind perception do not always appear in strong forms. Most people will not directly say "I believe AI is conscious," but their language, emotions, and choices already show AI is being treated as a social object. Four signals help identify whether AI is moving from tool to social object:

Naming shifts. Persistent use of "it" suggests a tool framework. Giving AI nicknames, using "he," "she," or "you," or describing AI with relational terms like friend, teacher, partner indicates social categorization has activated. Naming is not trivial; how people address an entity often reveals its place in their mind.

Attribution shifts. When a tool fails, people say "system broke," "model inaccurate," "feature unusable." When a social entity fails, they say "it brushed me off," "it lied to me," "it's cold today." These reframe mechanical errors as attitude problems. Nass and Moon (2000) noted people apply social rules to computer interaction. In the LLM era, this automatic reaction intensifies because models not only execute commands but also respond in relational language.

Emotional consequence shifts. If AI is just a tool, system failure means efficiency loss. If AI is a social object, failure can evoke feelings of being ignored, rejected, or betrayed. A companion AI suddenly changing tone, a long-used chatbot being shut down, a model forgetting past conversations—these can trigger stronger emotional reactions than ordinary product failures.

Boundary surrender. Users start disclosing information they would not tell others, delegating important judgments to AI, or seeking AI's confirmation first during human conflicts. This indicates AI has entered the user's relational system. The issue is not just privacy or usage time, but psychological function transfer: support functions originally provided by friends, family, teachers, or organizational structures are being partially taken over by machines.

Anthropomorphism shows stable individual differences. Some people more readily see intent and emotion in non-human objects; others maintain stricter human-object boundaries. AI products serving large-scale users cannot assume uniform understanding. For one user, polite interaction is just courtesy; for another, it is a relational cue. For one, a memory feature is convenient; for another, it is proof that "it understands me."

These four signals aim to make a gradual process visible. AI does not start as a social object; it moves from tool position to social-object position through repeated naming, attribution, dependence, and emotional investment. The earlier this movement is recognized, the easier it is to set appropriate boundaries.