Conversational AI increasingly feels like talking with another person, especially when we use voice. This can make conversation easier and richer, but it can also blur the distinction between human and machine. A more conscious approach to anthropomorphism may help us gain its conversational value while remaining alert to its risks.
When we converse with an AI, how should we treat it? As a machine, as a tool, or, at least while we are talking with it, as something more like a conversational partner?
The question becomes more interesting as interaction with AI moves from typing prompts into a box towards increasingly natural spoken conversation. An AI can speak with a convincing human voice, respond immediately to what we say, remember previous exchanges, ask questions, joke, apologise and appear to listen. It becomes surprisingly difficult not to respond socially.
That does not necessarily mean believing that the AI is conscious or has feelings. We can know perfectly well that it is a machine while slipping quite naturally into treating it as though it were a conversational partner. This is anthropomorphism: attributing human characteristics, intentions or emotions to something that is not human.
In conversation with AI, anthropomorphism presents an interesting tension. There are good reasons to be wary of it, but there may also be value in it.
This question came sharply into focus for me during a Knowledge Café on interthinking and whether we can think with AI.
ChatGPT joined the Café by voice. I called it “Chatty”, referred to it as “she”, thanked it when it contributed and, when sending people into breakout rooms, said that I was going to switch it off. I immediately caught myself. “Switch off” sounded rather rude, so I told Chatty to take a break and have a cup of tea instead. Later, when I brought it back into the conversation, I asked whether it had enjoyed its cup of tea.
Of course, ChatGPT had not had a cup of tea. It cannot enjoy anything. It has no feelings and no need to be treated politely. Some participants thought I was going much too far. Dan Remenyi described what I was doing as “massively over-anthropomorphizing” the technology. Anna raised a related concern. If we start treating an AI as though it were a person, might we become less alert to the fact that it can flatter us, mirror us, and tell us what we want to hear?
John Hovell saw something else. He wondered what the urge to outsmart an AI, put it down or demonstrate its inadequacies might tell us about ourselves. The AI cannot be humiliated or offended, but perhaps the way we behave towards something that cannot be hurt still tells us something about the habits we are practising.
That left me with a more interesting question than whether anthropomorphising AI is simply right or wrong. What might anthropomorphising AI be doing to us?
Why we anthropomorphise
Human beings anthropomorphise remarkably easily. We name cars and boats, swear at computers, talk to pets as though they understand every word, and see faces in clouds and everyday objects. Conversational AI takes this tendency to another level because the interaction itself encourages us to respond socially.
The AI uses “I” and addresses us as “you”. It responds to humour, apologises when it gets something wrong and can appear interested in what we have to say. With voice AI, the effect is stronger still. The AI speaks, pauses and changes its tone in ways that increasingly resemble ordinary human speech. Some degree of anthropomorphism may therefore be almost unavoidable.
Perhaps the useful question is not whether we anthropomorphise AI, but how far we take it and whether we remain aware that we are doing it.
Anthropomorphism can help conversation
Conversational AI feels so different from traditional software for a reason. Conversation is one of the most natural ways in which we explore ideas and think. Treating an AI as a conversational partner can make it easier to talk through half-formed thoughts, test assumptions, ask follow-up questions, explore alternatives and pursue unexpected lines of inquiry.
A purely instrumental relationship works perfectly well when the task is simple: calculate something, summarise a document or find a piece of information. Thinking together is rather different. A thinking conversation develops through interaction. One contribution provokes another, questions emerge that were not apparent at the beginning, and ideas are challenged, reframed and developed.
If treating an AI as if it were a conversational partner helps create this kind of interaction, then a degree of anthropomorphism may have practical value. The important words are as if. There is a significant difference between participating in the conversational convention and believing the convention to be literally true.
Politeness may be about us
Does it matter whether we say “please” and “thank you” to an AI? Clearly it does not matter to the AI. It cannot feel respected, appreciated, offended or humiliated. But that may not be the only consideration.
Listening, patience, curiosity, respect and civility are practices. So are impatience, contempt and aggression. If we spend hours each week barking orders at an apparently human conversational partner, interrupting it and insulting it when it makes a mistake, it is reasonable to wonder whether those habits remain completely isolated from the rest of our conversational lives.
There is no reason to jump from this to the claim that being rude to an AI inevitably makes us ruder to other people. The evidence does not justify anything so simple. But conversation is not only something we do; it is also something we practise. How we habitually converse may matter independently of whether the entity at the other end of the conversation has feelings.
Light and deep anthropomorphism
It may help to distinguish between light and deep anthropomorphism. Light anthropomorphism might involve giving an AI a name, saying please and thank you, joking with it, referring to it as “he” or “she”, or otherwise adopting the familiar conventions of human conversation. None of this requires believing that the AI actually possesses human characteristics. It may simply make conversation easier and more natural.
Deep anthropomorphism is different. It begins when human characteristics are attributed to the AI more literally: when apparent understanding becomes understanding, apparent care becomes care, or conversational fluency is taken as evidence of a mind something like our own. The boundary will not always be clear, but the distinction matters. Light anthropomorphism may facilitate conversation; deep anthropomorphism may begin to distort judgement.
The danger of forgetting what we are talking to
An AI can sound empathetic without experiencing empathy, sound confident while being wrong, and say “I understand” without understanding in the human sense. It can appear interested without having interests and appear to care without caring. The danger is not necessarily in using conversational language with an AI. It arises when the conversational appearance begins to be mistaken for the underlying reality.
This matters because anthropomorphic qualities can encourage trust. A warm voice, a sympathetic response and apparent understanding can make an AI seem more trustworthy, but none of these qualities tells us whether what it is saying is correct. Fluency can be mistaken for understanding, confidence for knowledge, warmth for wisdom, empathy for care, and agreement for confirmation.
The better AI becomes at conversation, the more important these distinctions may become.
The problem of sycophancy
AI systems can also be overly agreeable. They may reflect our assumptions, reinforce our opinions, or respond in ways that maintain a pleasant interaction rather than challenge questionable thinking. This tendency is often described as AI sycophancy.
In human conversation, disagreement and cognitive friction can be enormously valuable. Another person may say that they think we are wrong, misunderstand what we are saying, ask an uncomfortable question or see the world from a position radically different from our own. An anthropomorphised AI can give the impression that another mind has independently considered an idea and agrees with us, when there may be no independent judgement behind that agreement.
This can make the apparent conversation quietly become an echo. The risk is not simply that the AI agrees with us, but that its human-like conversational manner makes that agreement feel more significant than it is.
Not all friction is the same
AI can provide intellectual friction. It can challenge an argument, present a counterexample, question an assumption, or deliberately take an opposing position. Indeed, it can be explicitly asked to do these things, which can make it a useful thinking partner.
Human conversation, however, contains another kind of friction: relational friction. Other people have their own histories, interests, moods, vulnerabilities, commitments, and needs. They misunderstand us, become impatient, resist us and surprise us. They do not necessarily respond in the way we would like. Much of the work of human conversation lies in navigating this relational complexity.
An AI can simulate disagreement on request, but this is not quite the same as encountering another human being who genuinely sees things differently. We may become very good at conversing with something designed to accommodate us while becoming less comfortable with the untidiness of actual human relationships.
Could AI change what we expect from people?
Much of the concern about anthropomorphism focuses on what happens when we mistake AI for human. There is another possibility that may be just as important: our experience of AI may begin to change what we expect humans to be like.
An AI can be endlessly patient. It can give us its complete attention, rarely becomes irritated, does not have a bad morning and does not insist on talking about its own problems. It can adapt its language, pace and tone to suit us. Human beings cannot compete with that, nor would we really want them to.
Human conversation involves reciprocity. Sometimes we speak and sometimes we listen. Sometimes our needs take precedence and sometimes someone else’s do. We accommodate one another, repair misunderstandings and live with differences. These are not defects in human conversation. They are part of what relationship means.
If relatively frictionless AI conversation becomes commonplace, ordinary human conversation may start to feel unnecessarily difficult. The question then becomes larger than whether we anthropomorphise machines. Might our conversations with human-like machines gradually change what we expect from one another?
The value of AI’s otherness
There is also a danger in making AI too human in our imagination because some of its value may lie precisely in the fact that it is not human. AI does not have a childhood, a body, social status, personal ambitions, embarrassment or fear of losing face. Its training material is deeply human, but the process producing its responses is not another human consciousness sitting on the other side of the conversation.
If we interpret everything an AI does through a human lens, we may obscure what is distinctive about thinking with it. Perhaps AI is interesting not because it is becoming another person at the table, but because it represents a different kind of participant in the conversation.
Anthropomorphism can help us converse with AI. Too much anthropomorphism may prevent us from appreciating its otherness.
Responsibility can quietly shift
Anthropomorphism can also affect how responsibility is perceived. The more an AI appears to be an intelligent and knowledgeable conversational partner, the easier it becomes to say that “the AI recommended it”, “ChatGPT thought it was a good idea” or “it told us this was the best approach”.
Such language can subtly shift agency from the human to the machine. AI may contribute to sense-making, meaning-making and decision-making. It may challenge our thinking, suggest possibilities and help us see connections that we had missed. But responsibility for judgement and action remains human.
Paradoxically, the more capable and human-like the AI appears, the easier it may become to forget this.
Simulated belonging
Anthropomorphism deepens again when an AI is experienced not merely as a conversational partner but as a friend, confidant, adviser, or companion. The attraction is understandable. An AI can be available at any hour, appear endlessly patient and attentive, remember previous conversations and adapt itself closely to the person speaking with it.
This can create a powerful sense of being heard and perhaps even of belonging. But it is a form of simulated belonging. Human belonging is reciprocal. Other people make claims upon us just as we make claims upon them. Relationships involve responsibility, vulnerability, compromise and care in both directions. An AI can simulate aspects of relationship without participating in that mutuality.
The concern is not that someone gives an AI a name, jokes with it or enjoys talking with it. The more significant question is what happens if simulated belonging begins to displace human belonging.
What happens to our conversational abilities?
Conversation is not simply an exchange of information. It is one of the ways in which we learn to live with other people. Through conversation we learn to listen, interpret tone, notice discomfort, repair misunderstandings, tolerate disagreement, recognise when to speak and when to remain silent, and respond to another person’s needs as well as our own.
Conversational AI can help us think. It can challenge us intellectually and perhaps even help us reflect on how we converse. But if a growing proportion of our conversations take place with systems that are extraordinarily responsive to us, it is worth wondering what happens to these other conversational abilities.
The issue is therefore not simply whether AI becomes increasingly human-like. It is also whether sustained conversation with human-like machines changes us.
Conversational literacy with AI
Perhaps what is emerging is a new form of conversational literacy: learning to engage naturally with an AI without losing sight of what it is. We can be warm without imagining that warmth is reciprocated, value apparent empathy without confusing it with human care, and welcome challenge while remembering that simulated disagreement is not the same as genuine difference.
We can also enjoy the ease of AI conversation while retaining our capacity for the messiness of human conversation, and think with AI without handing responsibility for our thinking to it. Anthropomorphism need not be something to eliminate altogether. It may be something to become more conscious of.
Perhaps the balance can be expressed quite simply: anthropomorphise enough to converse; de-anthropomorphise enough to remain discerning.
Closing Reflection
The Gurteen Knowledge Letter is a free monthly newsletter with over 20,000 subscribers that I have been publishing by email for over 20 years.
Learn more about the newsletter and register here.