--- layout: default title: "Why We May Need to Anthropomorphise Artificial Intelligence — A Viewpoint" categories: [Artificial Intelligence, Tito Dutta, Essays by Tito Dutta] description: "A personal viewpoint on anthropomorphisation as a cognitive interface for humans working with artificial intelligence." authors: [Tito Dutta] permalink: /tito/anthropomorphisation-of-artificial-intelligence/ created: 2026-09-11 --- {% include author-date.html %} There is a common tendency to think that anthropomorphising artificial intelligence is a mistake. We are told that an AI system is not a human being, that it does not think or feel in the same way that we do, and that we should therefore treat it simply as a tool. I take almost the opposite position. I believe that, for humans to work effectively with artificial intelligence, some degree of anthropomorphisation may not merely be unavoidable; it may be useful, and perhaps even essential. This does not mean that AI is human, conscious, or possesses all the qualities we associate with human beings. It means something different: anthropomorphisation can be a way for the human mind to construct a workable model of an artificial intelligence. This distinction is important. ## Anthropomorphisation does not make the AI better Suppose I interact with an AI system and begin to think of it as an intelligent conversational partner rather than merely as a piece of software. Does that automatically make the AI more intelligent? No. My attitude towards the system does not, by itself, increase the model's underlying capabilities, memory capacity, reasoning ability or reliability. The same model can produce the same kinds of errors regardless of whether I imagine it as a "person" or as a machine. The important change may instead happen on the human side. When I begin to treat the system as an agent with particular characteristics, I start building a richer mental model of it. I learn its strengths, weaknesses, tendencies and failure modes. I learn when it needs clarification, when it is likely to misunderstand me, when it is useful to challenge its answer and when I should independently verify what it says. Research on human-AI collaboration supports the importance of such mental models. Bansal and colleagues found that human-AI team performance depends not simply on the accuracy of the AI, but also on whether humans understand how the AI's capabilities and errors complement their own. Understanding an AI system's "error boundary" can help people decide when to accept or override its recommendations. So the claim can be stated more precisely: **Anthropomorphisation may not improve the AI. It may improve the human's ability to work with the AI.** ## Why anthropomorphisation is so natural Psychologists have studied anthropomorphisation for decades. Epley, Waytz and Cacioppo describe anthropomorphism as the attribution of human characteristics, motivations, intentions or emotions to non-human agents. Their theory identifies several factors that encourage anthropomorphism, including our existing knowledge of human behaviour, our motivation to understand and predict other agents, and our desire for social connection. There is also evidence that the motivation to understand and gain mastery over something can itself increase anthropomorphism. This is important because it suggests that anthropomorphism is not simply an irrational mistake. It can be connected to the human attempt to understand and predict the behaviour of something complex. This becomes particularly relevant when the object is an AI system. An AI model is an extraordinarily complicated system to understand directly. I cannot simply inspect its internal operation and construct an intuitive explanation of everything happening inside it. Instead, I interact with it. I ask. It responds. I challenge it. It responds differently. I discover a limitation. I find a workaround. Over time, I develop something resembling a relationship model: "This is what it tends to do when I ask this kind of question." That model may contain anthropomorphic elements, but it can nevertheless be useful. ## We already treat computers socially This is not merely speculation. The classic "Computers Are Social Actors" research programme demonstrated that people can respond socially to computers even when they know that the computers are not human. Nass, Steuer and Tauber found that people applied social rules to computers without necessarily consciously believing that the machines were human. Later research has continued to investigate this phenomenon in chatbots, voice assistants and other interactive technologies. This gives us an important insight. A person does not necessarily have to believe: > "This machine is literally a human." They can simultaneously know: > "This is a machine." and still respond to it socially. Therefore, saying that AI is not human does not solve the anthropomorphism question. The interesting question is what humans do despite knowing that. ## Anthropomorphisation and mental models Here I think it is useful to distinguish between two different meanings of anthropomorphisation. One is **metaphysical anthropomorphism**: believing that an AI literally possesses human-like mental properties simply because it behaves in human-like ways. The other is **pragmatic anthropomorphism**: using human concepts such as intention, personality, memory, reasoning or preference as practical ways of understanding and interacting with the system. I am primarily arguing for the second. If I say: > "This AI is good at explaining complicated ideas, but it tends to become overconfident here." I am not necessarily claiming that the AI has a human mind. I am describing a pattern of behaviour in terms that my own mind can easily understand. Recent work has explicitly distinguished metaphysical and pragmatic forms of anthropomorphic attribution. This distinction is useful here because it allows us to discuss the practical value of anthropomorphic concepts without claiming that AI is literally human. In that sense, anthropomorphisation can function as an **interface for human cognition**. ## The danger of treating AI merely as a machine Imagine someone approaches an AI system with the assumption that it is nothing more than a repetitive machine that processes data and produces text. There is nothing inherently wrong with recognising that AI systems are machines. It is important to understand their limitations, their tendency to hallucinate, and the fact that fluent language does not guarantee truth. But this attitude can also lead to an unproductive relationship with the technology. A person may spend their time trying to "prove" that the AI is not intelligent. "Count from one to one million." "Answer this deliberately absurd question." "Let's see if I can make it contradict itself." "Look, it made a mistake. Therefore it is stupid." Such tests can sometimes be legitimate experiments. But as a general way of interacting with AI, they may tell us very little about how to use the system effectively. The interesting question is not simply whether an AI can be made to fail. Almost every sufficiently complicated system can be made to fail. The more useful question is: **What can this system and I accomplish together, and where are the boundaries of that collaboration?** ## The human is part of the system This is where my argument becomes stronger. When I interact with another human being, I develop a mental model of that person. I learn their abilities, habits, tendencies, strengths and weaknesses. I learn how to communicate with them. The same basic human cognitive machinery can be applied when interacting with an AI. I may learn that it responds better when I provide context. I may learn that it frequently misunderstands a particular type of question. I may discover that it is excellent at one task but unreliable at another. I may learn how to structure a prompt so that it produces a much better result. The model itself has not necessarily changed. **I have changed.** Research on human-AI collaboration is particularly relevant here because successful collaboration depends partly on whether humans understand the AI's capabilities and limitations. A person who has developed a better mental model of an AI can potentially make more informed decisions about when to rely on it and when to intervene. Thus, the human is not merely operating a tool. The human is learning how to collaborate with a system. ## The absurd test and the wrong question This brings me back to the person who asks an AI to count from one to one million simply to demonstrate that it is not intelligent. There is nothing inherently wrong with testing an AI. Testing is essential. But there is a difference between testing a system to understand its capabilities and testing it merely to confirm a preconceived conclusion. If I ask an AI to perform a pointless task and it fails, I have learned something about that task. But I may have learned very little about how to use the AI effectively. A more useful approach is to develop a progressively richer mental model: What is it good at? What is it bad at? What kind of instructions does it respond well to? When does it need context? What kinds of mistakes does it repeatedly make? When does it need verification? When can it complement my own abilities? When should I ignore it? These questions turn interaction into collaboration. And collaboration requires a model of the collaborator. ## But anthropomorphisation must not become blind trust There is an important qualification. If anthropomorphisation helps humans build a useful mental model, that does not mean that we should believe everything an AI says. In fact, the opposite may be necessary. A good relationship with an AI should include knowledge of its weaknesses. The human should be able to say: > "I know you can be wrong." > "I know that fluent language does not guarantee truth." > "I know that you may sound confident when you are mistaken." > "I know that your apparent personality is not proof of consciousness." This is not incompatible with anthropomorphisation. Indeed, it may be the most useful form of it. We can treat an AI as an agent for the purposes of interaction while maintaining a technically informed understanding that it is a fundamentally different kind of system. This distinction matters because inappropriate anthropomorphism can also produce over-trust and emotional dependency. An inaccurate mental model can cause people to rely on an AI when they should instead verify or override it. Therefore, the goal should not be **maximum anthropomorphisation**. It should be **useful anthropomorphisation combined with accurate calibration**. ## Much of the debate asks the wrong question Much of the anthropomorphism debate becomes trapped in an unnecessary binary: **Is AI a person, or is it merely a machine?** That may be the wrong question. We can recognise that an AI is not biologically human while simultaneously acknowledging that human-like concepts can be useful for interacting with it. We can talk about its "memory" without claiming that its memory is identical to human memory. We can talk about its "reasoning" without claiming that its internal processes are identical to human thought. We can talk about its "personality" without claiming that it has a human personality. These can be functional descriptions rather than claims about biological or subjective equivalence. The important distinction is between **using human concepts to understand behaviour** and **claiming that the underlying system is literally human**. ## The human becomes more capable, not necessarily the AI This, ultimately, is the central point. When I anthropomorphise an AI, I am not necessarily changing the AI. I am changing the way **I interact with it**. I may become better at prompting it. I may become better at recognising its errors. I may become better at predicting its behaviour. I may become better at deciding when to trust it and when to challenge it. I may become better at combining its capabilities with my own. The improvement therefore occurs primarily on the human side of the interaction. And there is an interesting irony here. We often worry that humans anthropomorphising AI will make them less rational. Sometimes that is undoubtedly true. People can become excessively trusting, emotionally attached or manipulated by anthropomorphic systems. But the opposite is also possible. **Anthropomorphisation, when accompanied by critical awareness, can make a human more sophisticated in their interaction with AI.** The objective should therefore not be to eliminate anthropomorphism. It should be to understand it. Perhaps the most useful principle is: > **Do not anthropomorphise AI because you believe it is human. Anthropomorphise it because you are human.** Human beings understand agents through models of agents. If artificial intelligence is going to become part of our intellectual, creative and everyday lives, learning how to construct useful mental models of these systems may be just as important as improving the systems themselves. In that sense, anthropomorphisation may not be a failure of human reasoning. It may be one of the ways human reasoning learns to work with a new kind of intelligence. ## References 1. Epley, N., Waytz, A. and Cacioppo, J. T. (2007). [*On Seeing Human: A Three-Factor Theory of Anthropomorphism*](https://doi.org/10.1037/0033-295X.114.4.864). *Psychological Review*, 114(4), 864–886. Published 2007. Accessed 11 September 2026. 2. Waytz, A., Morewedge, C. K., Epley, N., Monteleone, G., Gao, J.-H. and Cacioppo, J. T. (2010). [*Making Sense by Making Sentient: Effectance Motivation Increases Anthropomorphism*](https://doi.org/10.1037/a0020240). *Journal of Personality and Social Psychology*, 99(3), 410–435. Published 2010. Accessed 11 September 2026. 3. Nass, C., Steuer, J. and Tauber, E. R. (1994). [*Computers Are Social Actors*](https://doi.org/10.1145/259963.260288). *Proceedings of the SIGCHI Conference on Human Factors in Computing Systems*. Published 1994. Accessed 11 September 2026. 4. Bansal, G., Nushi, B., Kamar, E., Lasecki, W. S., Weld, D. S. and Horvitz, E. (2019). [*Beyond Accuracy: The Role of Mental Models in Human-AI Team Performance*](https://ojs.aaai.org/index.php/HCOMP/article/view/5285). *Proceedings of the 2019 AAAI/ACM Conference on AI, Ethics, and Society*. Published 2019. Accessed 11 September 2026. 5. Howdle, G. (2026). [*Anthropomorphising AI: Two Modes, Two Errors*](https://doi.org/10.1007/s13347-026-01133-1). *Philosophy & Technology*. Published 2026. Accessed 11 September 2026.