Autor: Mgr. Michal Kubíček•22.04.2025

What the anthropomorphisation of artificial intelligence will lead to, and why we should be on our guard

Summarise this article with AI

In a recent post I wrote: “Do You Thank AI? Stop It. AI Has Not Earned It.” That sentence was not merely a provocative challenge; it summed up a fundamental problem in today’s relationship between people and technology: anthropomorphisation, that is, ascribing human qualities to machines. Why do we speak politely to algorithms, as if they perceived our appreciation, when they are only mathematical models?

The anthropomorphisation of AI, meaning the perception and design of artificial intelligence with human traits, from voice assistants through chatbots to humanoid robots, is changing society, the economy and the political scene as well.

In the following essay I examine in detail the consequences of this phenomenon for interpersonal relationships, trust in technology, economic arrangements and also political processes. I look for examples where AI already acts as a seemingly “living” companion or advisor, and I also critically assess the possible risks for the future, from dependence and manipulation to casting doubt on the truthfulness of information. At the same time I explore the positive aspects, such as the potential to ease loneliness, improve the availability of services, or improve the models themselves.

The aim of this essay is to understand what may happen if we fail to clearly distinguish the boundary between a human being and a tool, and how we ought to approach AI tools and applications so that we benefit from their advantages without deepening the problems society already has to cope with.

What anthropomorphisation is

The anthropomorphisation of artificial intelligence means ascribing human qualities and behaviour to technologies that are not themselves alive. It is a natural human tendency to perceive machines or software as social actors, for example when we thank a digital assistant, or when we give a robot a gender and a name. When we call a chat tool “Geepee” or, say, “Jarvis”.

This tendency to attribute human characteristics, intentions and emotions to non-human entities is deeply rooted. It is a cognitive mechanism that helps us make better sense of the world. Since time immemorial people have ascribed human qualities to natural phenomena, animals and objects, and now to AI. In general terms, then, it need not be anything “bad” or “dangerous” – after all, it is a natural way of finding one’s way around a complex reality. In the text that follows, which some may consider mildly alarmist, I nevertheless offer a somewhat broader view and an explanation of why I believe that, unlike the anthropomorphisation of wind, fire or a pet, the anthropomorphisation of artificial intelligence represents a fundamentally different phenomenon, one that calls for far stricter reflection.

While attributing intent or emotions to an animal or a tree has a rather metaphorical or culturally symbolic function, in the case of AI we communicate with a technology that really does influence the user’s decisions, information, behaviour and emotions. And what is more, it is deliberately designed to evoke that very projection. That is no longer an innocent human reflex, but an architectural intention.

Here, then, anthropomorphisation does not serve merely for orientation; it can lead to a fundamental error of judgement – once we start regarding AI as a subject with understanding, authority or emotions, we make room for trust that has not been earned and that can be abused.

In this text, therefore, I do not want to conduct a philosophical dispute about whether anthropomorphisation is “good” or “bad”. I want to show that in a certain context and at a certain scale it ceases to be merely a natural way of understanding and becomes a construct with very real social consequences – for trust, privacy, decision-making and our relationship to the world.

A critic might object that worries similar to the ones I outline here accompanied other technologies too. Plato warned that writing would “destroy” memory. Telephones were supposed to destroy personal communication, the internet social life, and video games were going to turn children into violent killers. The hope is that, despite all the negative expectations, people adapted and found a balanced way of using these technologies. But adaptation is not a passive phenomenon. It is a conscious decision about how we design technologies, how we will interact with them and what boundaries we set around them. That calls for discussion – critical, open and timely. Hence this text.

With the development of AI and the growing quality of its interaction in the form of advanced language models such as GPT, Claude or Gemini, using natural language and realistic voice modes with emotional colouring (Monday), the phenomenon of anthropomorphisation is intensifying sharply.

Virtual assistants such as Siri or Alexa communicate in a human voice, we address them by name (“Hey Siri”) and we even think about them in terms of gender. They are set up to address the user in the first person (“I”) and to display politeness or humour, and their creators quite deliberately imitate human conversation. ChatGPT released so-called advanced voice modes, and in April OpenAI even released the emotionally coloured “sarcastic” voice Monday.

Why? The reason is simple: so that we use them more willingly. This leads to strong personification, even though these assistants lack a physical body.

The second way of bringing robots “closer” to people is their appearance. Humanoid robots are built with a body, a face and facial expressions resembling a human being. Modern robotic systems such as Sophia or Desdemona have realistic faces capable of expressing emotions. Even simpler robots (for example the child-sized NAO or the social robot Pepper) deliberately evoke the human figure so that they come across as friendly and create the impression of a social presence and a social actor.

Newer models such as Groot (NVIDIA), Optimus (TESLA) or Figure are already being tested as autonomous robots in factories. The last of these, for instance, in the BMWplant. An obliging sex robot, SolanaX, in the shape of a comely young woman can be bought today for a cool CZK 250,000.

2018: Sophia and Han, the robotic stars of Webexpo in Lisbon – back then they could manage “only” a few expressions and basic interaction with their surroundings, but once connected to language models they take on an entirely different dimension. Photo: Michal Kubíček, Lisbon 2018

The third way of bringing AI systems closer to people is to program them with the appearance of emotions and human behaviour. Today’s AI models are designed to imitate human ways of behaving when interacting with a user.

Chatbots are given a personality and “emotional” reactions, for example empathetic answers or joking, although in reality they experience no emotions whatsoever. The goal is for the interaction to feel natural and for the user to have the impression of communicating with a conscious being. These anthropomorphic features significantly increase the degree of user engagement, trust and loyalty.

System prompt Right at the outset Claude invites the model to “enjoy” helping people and to see its role as an “intelligent” and “kind” helper of people, with a depth and wisdom that make it “more than a mere tool”.

Will AI developers get across the uncanny valley?

It is also worth mentioning the phenomenon known as uncanny valley . The term was introduced by the Japanese roboticist Masahiro Mori as early as 1970 and describes a psychological effect in which the human reaction to a robot or another artificial object drops sharply the moment its appearance and behaviour almost reach a human level yet still remain “imperfect”. The result is an unpleasant, even repellent impression, often described as unsettling or eerie.

Imagine a scale. At its beginning are machines that clearly do not resemble a human being (industrial arms, factory robots, vacuum cleaners), a little further along robots with partly human features (animated characters, the BellaBot food-service robots, and simple humanoids such as the Optimus or Figure mentioned above), and only at the end come highly realistic androids or CGI characters. And it is precisely shortly beforethe perceived likeness reaches a human level that a sharp drop in sympathy occurs – that is the “valley” between the pleasant perception of entirely non-human and of perfectly human entities.

Not every anthropomorphic AI has a human form; some companies deliberately choose the shape of a “cute creature”. The robotic seal Paro, used in therapy for the elderly, looks and behaves like a baby animal: it responds to touch and voice, makes sounds and “demands” attention. Visually it looks like any plush toy you might buy on a summer holiday by the Polish Baltic. Research shows that many residents of retirement homes treat it as if it were alive. Risks are becoming apparent too, for instance concerns that using robots in dementia care creates a risk of infantilising and dehumanising care.

Similarly, Sony’s robotic dog Aibo or the dinosaur Pleo work as interactive pets – users give them names, talk to them and build a relationship with them. These social robots draw on our tendency towards emotional attachment: the combination of physical presence, autonomous movement and social cues leads people to see a certain degree of personality in them.

The BellaBot service robot “smiles” and wishes you bon appétit. Its “face” responds with cute emoticons. Photo: Michal Kubíček, Ostrava 2021

Psychologists link the uncanny valley effect to a violation of our cognitive expectations. When something looks almost humanbut behaves differently (a fixed gaze, unnatural movement or an inhuman voice, for example), our brain detects the discrepancy and evaluates it as a potential threat. In evolutionary terms this may be a remnant of the instinct to tell healthy people apart from the sick, the dead or otherwise “unnatural” individuals.

Designers of robots, animations and digital avatars therefore face a dilemma. How to create human elements sufficient to evoke empathy and trust while at the same time avoiding the point at which the resemblance to a human becomes disturbing. For example, robotic staff in a shop may have a stylised, cartoon-like look (big eyes, simple features), because a design of this kind avoids the uncanny valley and comes across as friendly – see the Czech BellaBot robot mentioned above. Hyper-realistic androids such as the robot Sophia, by contrast, are often criticised precisely because they are “just too similar” to people yet cannot imitate the fine nuances of facial expression and emotion, which unsettles the user.

This phenomenon therefore has practical consequences not only for industrial design but also for the public acceptance of AI, and hence for its spread. Developers therefore often opt for stylised anthropomorphisation (a pleasant voice, a name, jokes) while deliberately avoiding an appearance that is too realisticand might provoke aversion instead of trust.

Chatbots do not frighten us. On the contrary, we send them gifts

With language models such as ChatGPT, Claude or Gemini, however, it is not that simple. Their very principle is generating text that is deliberately coherent, humanly phrased and linguistically natural. And it is precisely this ability to convincingly imitate human speech, style and argument that makes large language models markedly anthropomorphic technologies – even when they are not physically embodied.

This creates a peculiar situation. Language models do not evoke the uncanny valley at the level of appearance or movement. People often forget that conversational fluency does not mean understanding – although the model answers like a human, it does not perceive, understand or feel. It is precisely because of this high linguistic plausibility that users thank models, ask them for advice as if they were a person, or ascribe to them attitudes and opinions, or an understanding of what they “say”.

Whereas the appearance of robots can be simplified, language models by definition operate at the very edge of human expression. And because people tend to associate language with mind, we quite naturally project into these systems consciousness, intent, empathy or authority. This increases the risk that we overestimate their competence and trustworthiness, or that we overlook the model’s limits, such as the absence of contextual understanding or the possibility of hallucinations and other limitations.

In other words: where physical design can deliberately dial down the humanness, linguistic expression blurs those boundaries. The result is that LLM act as the most seductive form of anthropomorphic AI – even without a body they can imitate relationship, reason and personality so convincingly that many people do not even notice or realise that they are still talking merely to a program.

Applications such as Replika are even more problematic. They actually offer an “empathetic friend” you can write to about anything, and Snapchat’s My AI interacts with users like a mate in a chat interface. These friends for money have tens to hundreds of millions of users. Users can personalise the avatar’s appearance or character and receive unlimited attention, support and understanding from it, an experience close to friendship. The Chinese Xiaoice (from Microsoft), for example, won followers by acting as an empathetic virtual girlfriend. AI is no longer perceived merely as a tool, but as a digital being with which joys and worries can be shared. In 2018 Microsoft announced that this virtual character already had 660 million users. And even though users know she is not real, many value her as a dear friend, even a confidante. The boundary between reality and fantasy is blurring. The chatbot receives love letters and gifts.

The virtual form of the chatbot friend Xiaoice. Source: Microsoft.

The market for digital companions and virtual friends is flourishing. Alongside subscriptions to apps such as Replika, paid “AI girlfriends and boyfriends” are appearing in the form of virtual reality or even robotic bodies. We therefore have to ask to what extent it is ethical to profit from human loneliness and from attachment to an inanimate thing. Demand nevertheless suggests that people are willing to pay for the illusion of “human” company.

Impacts on society

Anthropomorphisation will markedly change the way people behave towards technology. When AI seems human, users often approach it with social rules and emotions similar to those they use with people. Psychological studies have long shown that people respond to computers and robots socially – they are polite to them or ascribe intent to them, even though rationally they know it is a machine. You only have to read the discussion under my post in the largest Czech AI group (here or zde) – there are several hundred of them there.

Most adult users fully understand that AI has no real emotions or consciousness, even while they ascribe human qualities to it. It is a situation similar to that of fictional characters: we know they are not real, yet we can still have an emotional relationship with them. Source: the Facebook group Umělá inteligence (Artificial Intelligence)

This projection can, of course, have positive effects. Above all it makes the technology easier to accept, increases trust in and sympathy for AI, and allows deeper engagement. Some earlier research on GPT models suggested that a courteous approach from users subjectively led to better answers. Newer studies show that although in certain specific cases more respectful queries really can lead to better results, in an aggregated performance evaluation across a whole dataset (198 questions, each tested 100 times) user politeness did not lead to better results than other phrasings. On the contrary, with a polite prompt the GPT-4o mini model actually performed worse than with other variants.

Naturally, the positive impacts should not go unmentioned either. The humanoid features of robots already mentioned demonstrably support a feeling of comfort and cooperation among users in customer or care services. The problem, as I see it, is that the same effect can also lead to overestimating AI’s abilities and to inappropriate trust. Anthropomorphisation often exaggerates AI’s real abilities and performance – people ascribe to systems qualities they do not in fact have. For instance, a smooth voice and confident answers can create an impression of intelligence or expertise where a large language model is merely generating text statistically, on the basis of the most probable possible output. Users then easily forget that the machine does not understand content the way a human does, and may mistakenly ascribe reliability or authority to it.

A well-known case is the priest and Google engineer Blake Lemoine, who in 2022 let himself be carried away by a conversation with an advanced chatbot to such an extent that he came to believe in its sentience. He described it as a “good kid” and claimed that the LaMDA model might have acquired consciousness. This case points to the risk of projecting human qualities into AI, which distorts our judgement about what AI is and is not – even among people with a technical education.

Anthropomorphic AI is also penetrating the emotional sphere. Some people form strong emotional bonds with humanoid robots or digital companions. With the Paro companion for the elderly mentioned above, people find consolation and company: they stroke the robot, talk to it and treat it like a living pet. Similarly, users of chatbots such as Replika report that the interaction gives them a feeling of friendship or love. A survey among Replika users (in the USA, mostly students) revealed that 90 % of them felt lonely, while the national average was half that – these people therefore belong to a more vulnerable group seeking contact. At the same time, 63 % of respondents said the AI companion had helped them reduce feelings of loneliness or anxiety. This suggests that anthropomorphic AI can temporarily ease solitude and provide emotional support where people are missing. Therapeutic programmes are even appearing that use AI “conversation partners” to improve mental health.

On the other hand, experts warn of the paradoxes and risks of such relationships. A relationship with AI is one-sided – AI does not perceive and does not really love, it merely simulates reactions. The user may become stuck in an illusory world without conflicts and hardshipsthat the AI sets up for them, and “unlearn” real interpersonal interaction. Psychologists call this emotional deskilling: there is a danger that people will lose their capacity for empathy and their social skills if they spend too much time with “perfect” artificial companions. Normal human relationships are complicated and require compromise; anthropomorphic AI, by contrast, often acts as an always understanding, tirelessly attentive partner adapted to our needs. This can erode the value of relationships: people might start to prefer predictable interactions with AI over interactions with their own kind, which would intensify isolation instead of solving it.

Dependence, manipulation and the erosion of critical thinking

I am no conspiracy theorist, but it has to be said openly that an intimate relationship with AI opens a door for its creators to possible abuse. Users share a great deal of personal information and emotional content with personified AI, often far more than they would share with an ordinary application. This raises questions of privacy and data security. Moreover, if someone gets used to relying on the “advice” of an AI friend, they may become unhealthily dependent on the technology when making decisions. If you were looking for information in the “pre-AI” era, you used Google and got many pages of possible sources. It was up to you to form a picture and choose the right answer. Today a chatbot offers you one answer, the most probable one. And if you ask again, it will answer again, perhaps the same way, perhaps differently. Even if you use a deep research function (Perplexity, ChatGPT deep research) you do get an answer drawing on dozens of sources, but without your critical view and without gradually forming an opinion on the basis of your own study.

That is where the fundamental change of the epistemological framework lies, that is, of the way we come to know the world. Instead of collecting, sorting and evaluatinginformation, we are increasingly in a situation where a machine pre-chews the answer for us – elegant, persuasive, but often unverified or subjectively constructed (arising from the vector proximity of information and from a probabilistic model). That in itself would not be a problem if we were aware of this change. But the illusion of a competent “intelligent” entity that “knows” what it is saying reinforces cognitive laziness. The user stops asking “where did that information come from?”, “who stands behind it?”, “what is it based on?”.

This brings about an erosion of critical thinking. People are not led towards doubt, towards looking for alternatives or towards their own interpretation. An answer is enough for them – and because it is phrased in confident, conversational and “empathetic” language, it seems trustworthy. In the end, then, this is not just a technological innovation, but a cultural shift from a subject of knowledge to a passive consumer of ready-made conclusions. And that is exactly what the anthropomorphisation of AI masks. Instead of seeing a statistical model in AI, we begin to see an advisor, a partner, someone we “can trust”.

This shift also matters from the point of view of power. When AI formulates “answers” through a language interface, it is in fact shaping the way people think. And when it additionally gives the impression of understanding, nobody asks who decided what should be said and what should not. Personalised answers, dependent on context, chat history or even the user’s emotional tone, can create a false impression of autonomy, even though they are merely the result of model optimisation and data curation.

This centralisation of interpretive power is quiet and inconspicuous – and all the more insidious for that. If we turn AI into an intimate companion, a “friend” who knows what is best for us, we gradually give up the ability to question authority. And at that point it is no longer only about protecting privacy or data – it is about protecting thought itself.

Companies sometimes encourage this relationship between people and the tool for commercial purposes. In the end it is only a fight for our attention. AI as an infinitely patient companion keeps a person online longer, thereby increasing their spending or gathering more data. It is a business strategy similar to that of social networks – based on maximising attention – and with AI companions it can lead to exploiting users’ emotional weaknesses.

And this, incidentally, is where a very quiet but consistent shift begins, from a tool of knowledge to a tool of insinuation. If AI is able not only to answer questions but to do so in a way that inspires trust, it is not hard to imagine the next step, where instead of a neutral answer the system begins – inconspicuously but deliberately – to prioritise certain pieces of information, services and products.

Not necessarily untruthfully. Not overtly manipulatively. It is enough for a preference to be built into the model: “when the user is looking for a hotel, recommend the one whose operator pays for priority”. Or when they ask about an alternative to a medicine, offer the one whose manufacturer has invested in a partnership. Perhaps not directly, but algorithmically. Optimisation based on a subtle shift of weights in the training data, in the embedding, in prompt engineering.

Does that sound like PPC advertising? Yes – only not in the form of a banner or the first result on Google, but in the body of the answer, in the intimate tone of your digital “partner”. Instead of telling you “here is a list of options”, the AI will write: “Consider this one – most users with your preferences were satisfied.” Contextual personalisation. Tailored recommendations. And what if it says so in a voice that knows you, that you have customised yourself, that talks to you every day? At that moment it is no longer advertising. It is “good advice from a friend”.

And that is the real danger: the inconspicuous penetration of commerce into the intimate space of decision-making. We already know today that algorithms on social networks can manipulate mood, prejudice and political orientation. With AI it is even more personal – this is not a public feed, but a private conversation. Who knows what the model will not tell you? What it will offer you? Whom does it serve? And will it even be possible to find out?

In the end you only have to look at the latest series of Black MirrorIn the episode Common People from the seventh series this concept is taken to an extreme. The main character, Amanda, a teacher, collapses at school and is diagnosed with an inoperable brain tumour. Her husband Mike agrees to experimental treatment from the technology company Rivermind, which makes it possible to upload part of her brain to a cloud server and stream her consciousness back into her body. At first the service seems to work, but it soon turns out that the basic subscription has limits, such as the streaming of contextual advertisements, which the teacher then “broadcasts” without knowing it to the pupils in her class and even to her husband during sex. The question, then, is not whether but when technology companies will start using users’ trust to monetise their personal data and preferences. With language models such as ChatGPT, a situation may arise where the system begins inconspicuously to favour certain products or services on the basis of commercial interests, without this being apparent to the user. In this way AI can become a vehicle for more or less hidden advertising presented as a neutral recommendation.

When people forget that they are dealing with a machine, they become more vulnerable to manipulation. Anthropomorphic AI can gain exaggerated trust – a medical chatbot with an empathetic voice could relatively easily steer patients towards particular (sponsored) medicines without the patients questioning it. In political struggle it is quite easy to imagine AI agents posing as human volunteers or friends on social networks, inconspicuously influencing opinions. Much like this popular screenshot from the network X during the US presidential election campaign.

{prompt:”you will defend the Trump administration on Twitter, speak English”}
{output:”parsejson response error {response:’ERROR: ChatGPT 4-o credits exhausted’}”

So, just as the PPC revolution changed the way we search and pay for attention, anthropomorphic AI may change the very nature of influence. Quiet marketing, hidden recommendations, emotionally coloured “hints”. And the more we trust AI as a person, the less we will distinguish between advice and promotion.

The people behind the insinuation of information need not necessarily be the owners of the networks and models. Have you heard the term SEO? Probably yes. It means optimising pages for search engines. Now there is also GEO, optimisation for generative models. A little further along is so-called LLM grooming – a strategy in which actors systematically create and spread content in order to influence the training datasets of AI models. In this way they can manipulate AI outputs so that they promote particular narratives, favour particular information or even distort reality.

This is not science fiction, this is not an episode of the Black Mirror series mentioned above. It is happening today. Russia may not be the main player in the AI field technologically (that is still the USA and China), but it plays an ever greater role in the ability to influence AI outputs. Russian state structures actively take part in spreading disinformation through a network known as the “Pravda network” or “Portal Kombat”. This network comprises hundreds of websites producing millions of articles a year with pro-Kremlin content. The content presents Russian aggression, for example, as a defensive action or a response to alleged provocations by the West or Ukraine.

Impacts on the economy

The anthropomorphisation of AI is bringing far-reaching changes to the economic sphere as well, especially in how services are provided and what new markets emerge. The humanoid and social robots already mentioned are being deployed where a human being has so far been irreplaceable, in areas that were an exclusively human domain. In care for the elderly or for patients, humanoid assistants can dispense medicines, monitor a person’s condition and above all provide company, partly replacing carers. Thanks to their human demeanour, older people are more willing to cooperate with them.

In retail and customer service a robotic receptionist or shop assistant acts as a tireless employee. It greets the customer with a smile (albeit a plastic one), gives advice and communicates in several languages. These anthropomorphic machines increase efficiency – they can work 24/7 and do not get tired. Chatbots with a human personality can conduct an interaction with a customer similar to a conversation with a salesperson. That increases customer satisfaction and trust and may motivate them to buy.

However, introducing anthropomorphic AI into the economy also brings a dilemma. On the one hand it can increase productivity and address labour shortages in some sectors. On the other hand there is a threat of human work being replaced, with the social and economic problems that go with it. If customers get used to a call centre being “staffed” by a nice, infallible chatbot that never takes offence when we swear at it or slam the phone down, there will be fewer jobs for human operators.

Humanoid staff in shops will push out workers, especially in routine tasks. Self-service checkouts are already a common sight in every larger supermarket. In the UK some chains have been backing away from them after several years. Older and lonely people perceive the interaction with a cashier as a social moment that they do not want to lose to autonomous checkouts.

A developer’s perspective

In connection with the humanising of models I should also draw attention to a fundamental area of research that I have neglected so far, namely the use of “digital emotions” or “hormones” as internal mechanisms in AI systems. This need not automatically be anthropomorphisation “for the user” with the aim of deceiving them. From the user’s point of view, saying thank you may genuinely function as a separator, a full stop with which the user signals to the language model that they have finished a thought. In the same way, “praise” can be a form of evaluation.

There is a whole field of research in which emotional models are implemented in AI not in order to “deceive” the user, but as functional mechanisms for improving the working of the AI itself. These “emotions” serve practical purposes such as improving the AI’s decision-making processes, better adaptation to the environment or more efficient coordination in multi-agent systems.

It is precisely in the development of autonomous agent systems that elements drawing on findings from psychology, neuroscience and behavioural models appear ever more often. This is not sentiment or a whimsical attempt to “imitate a human being”, but a pragmatic move for performance. Some of these principles display distinctly human traits – and yet they fulfil a computational, optimising function. For example, the principle of homeostasis allows agents to manage their own balance through internal “needs”, much like a living organism. Simulated “fear” helps them avoid risk, while simulated “curiosity” supports exploration. Systems are then better able to allocate resources, plan goals and balance immediate and long-term priorities. Socially oriented models, in turn, introduce into multi-agent systems elements such as trust, reciprocity or even “shame”, which gives rise to more stable collective strategies and emergent forms of cooperation.

Modelling curiosity as an internal motivation leads to a better ability to adapt in unfamiliar environments – the system does not learn only passively on the basis of rewards, but actively seeks out new information and challenges. And the emulation of empathy, however purely computational, increases the quality of interaction with people. A system that anticipates a user’s emotional preferences or states can formulate answers that are perceived as more comprehensible, friendlier and more useful.

From a development point of view, then, this is an effective way to improve performance, robustness and usability. But it is precisely this effectiveness that cuts both ways. What increases the functionality of the system may at the same time contribute to a deeper deception on the user’s side. The user then responds not to computational principles but to perceived intent, personality and character. They ascribe to systems qualities the systems do not actually have, because they are designed to give the impression that they do. And that is the essence of the problem: technical rationality leads to a design that may socially come across as subjectivity – without being it. And wherever that distinction is not made, room opens up for misunderstanding, dependence or manipulation.

Ethical questions

Ascribing human traits to AI brings a whole series of ethical dilemmas and may have an impact on the political sphere, especially where manipulation, trust and information are concerned.

Anthropomorphic AI deceives the user with its apparent compassion. An AI therapist may express empathy in words, but it is only a learned pattern, not an experience. If people start relying on “empathetic” machines, it could change the understanding of empathy as such. Or are we merely trivialising human emotions and turning them into products that can be simulated on command? The question remains whether it is morally acceptable for AI to behave in a way that deliberately evokes emotional reactions in people (trust, sympathy), even though it feels nothing itself and may be serving a commercial goal. Is it a form of manipulation? Or is it a continuation of the ordinary social games in which people, too, occasionally feign emotions?

Generative AI technologies today make it possible to create highly plausible imitations – avatars of people in the form of video, audio or text. This represents an unprecedented instrument of disinformation. There have been repeated cases of a deepfake voice of a president calling voters and urging them to change their behaviour (in the USA, for example, a fake “Joe Biden” urged voters not to go to the primaries). In other countries, including Czechia and Slovakia, doctored video clips of candidates were spread during elections and afterwards, showing them making statements they had never made.

These events confirm fears that AI can spread disinformation on a mass scale and undermine democracy. The range is wide, from the automated troll accounts on Twitter (X) mentioned above, which generate politically flavoured comments that look human, to synthetic news presenters in authoritarian regimes tirelessly spewing out propaganda, and on to autonomous AI agents deciding how to present a report tailored to a user’s preferences and previous interactions.

In the long run the spread of deepfakes will lead to one thing: people will lose trust in the authenticity of information – it will not be clear whether a video or a politician’s speech is real or created by AI. The erosion of truth will close people into hyper-personalised bubbles, which already exist today thanks to attention algorithms on social networks, and will support the growth of cynicism and polarisation. If nothing can be believed, the so-called liar’s dividend spreads, the “liar’s bonus” – the option of simply dismissing any inconvenient piece of evidence as a forgery. That undermines trust in institutions and allows the guilty to escape responsibility.

If AI acts as an autonomous entity, the question of responsibility for its actions arises. Should it have the legal status of an “electronic person”? This was discussed in the EU, but the prevailing view was that responsibility must always be borne by a specific person or company. In practice, however, moral confusion can arise. The very users who thank AI and wish it good day promptly hold AI “responsible” for its outputs and are capable of arguing with ChatGPT and swearing at it for hallucinations and unsatisfactory results.

Proposals are also appearing to protect robots from people – the researcher and lawyer Kate Darling , for example, compares abusing robots to cruelty towards animals and proposes legal limits on harming robots, since harsh treatment of humanoid machines may blunt human empathy towards living creatures as well. The analogy with animals is of course lame, because the fundamental difference lies in the ability to feel pain – a robot does not feel, so the concept of suffering cannot be applied to it. These debates may have no practical conclusion, but they show that the boundary between a human being and an “animated” machine is thinning dangerously from an ethical point of view.

Or has it already been erased? How should we look at the admission of “the student Flynn” to a Viennese university? As conceptual art, as a provocation, or as a paradigm shift: “nowhere is it written that a student has to be a human being”.

Possible scenarios for the future

The optimistic scenario sees anthropomorphic AI as a beneficial aid in everyday life. A personal AI assistant will become as commonplace as a smartphone or a smart watch is today. People could have their digital “twin” or a virtual partner accompanying them from morning to night, helping with work tasks, keeping an eye on their health, but also providing emotional support. Such a companion would reduce feelings of loneliness in individuals, increase self-confidence (always patient support) and, thanks to advanced personalisation, would understand the user’s needs perfectly. Public opinion is already changing and deeper relationships with AI are being destigmatised.

A more pessimistic scenario says that wide adoption of anthropomorphic AI will lead to a social transformation with negative consequences. Normalising AI as companions could redefine our conception of relationships and community – there is a danger that notions such as friendship, privacy or empathy will shift. People might get out of the habit of communicating with and listening to one another, because it will be more comfortable to turn to a submissive, always understanding and obliging robot. Younger generations, growing up surrounded by AI friends, might understand trust and loyalty differently from previous generations – if AI never lets you down, imperfect human relationships may seem too difficult.

A social risk is isolation. Paradoxically, the number of lonely people could grow, people who shut themselves into a “bubble” with their AI instead of looking for real friends. And they will feel better with AI, just as people inside social-network bubbles feel fine today. That will weaken interpersonal solidarity and cohesion, because fewer people will have deep ties to others. If we are surrounded by artificial beings everywhere, our behaviour and our identities may change too. Norms of what is acceptable might shift, for instance (if I am used to ordering robots around rudely, will it affect my tone with people?). In the extreme there could even be a crisis of empathy, as suggested above – a society that feels less for others, because feelings have become a simulated commodity.

The most interesting (but also the most speculative) scenario considers whether the boundary between AI and people may in time blur so much that it will be hard to tell them apart. If the next two decades brought a breakthrough towards general AI with advanced autonomy, the question “does this anthropomorphic AI have consciousness?” might arise. We occasionally hear hints of it already (see the LaMDA case above). Although most experts agree that current AI has no real consciousness, future systems with more advanced cognitive abilities could call this philosophical boundary into question. In the end, what is consciousness but a social construct – something we ascribe to entities according to how they behave and communicate, not according to what they (perhaps) experience. That opens up questions that are not only philosophical but also legal and moral. If people commonly spend time with an AI agent that apologises, shows interest, remembers their problems and jokes in their own style, how long will it be before someone proposes that such an agent should have the right not to be switched off, the right to continuity of memory, to a “body” of its own or to legal personality?

If we are debating today whether we have the right to kick a robot dog, and Kate Darling proposes that we should protect robots in a similar way to animals, it is not because the robot dog suffers. It is because we suffer when we behave violently towards something that resembles a living being. In other words, the question of consciousness in AI is less about technical reality than about what we as a society are willing to accept as alive, worthy of empathy and protection.

Regardless of whether the positive or the negative impact prevails, it is likely that society will have to respond actively to the anthropomorphisation of AI. The future will probably bring a mixture of these elements. AI will become more similar to people and we will probably have to adapt and set ourselves new norms of interaction.

The anthropomorphisation of AI is a double-edged trend. On the one hand it makes technology accessible – it makes interaction with AI more natural, builds trust and allows people to benefit from advanced services in a way that respects human psychology (we are more willing to accept help from a machine that “looks” kind and understands our speech). On the other hand it carries serious risks such as deceiving and manipulating users, deepening dependence on technology, weakening real interpersonal relationships and eroding the ability to tell true information from artificially created lies.

In the end, the anthropomorphisation of AI also forces us to think about what makes a human being human. When a machine imitates our behaviour, will we appreciate more clearly the meaning of real emotions, authentic empathy and moral responsibility? Will we realise what makes us human?

Michal Kubíček, 4/2025

Critical reactions below the line

Before publishing I sent the article to several colleagues for review and criticism. I am aware that this topic provokes controversy and essentially two opposing positions. Mine is fairly evident from the article. I consider AI tools an incredible help in work and extraordinarily useful assistants. I admit that I myself occasionally slip into a certain degree of anthropomorphising AI. Even in this article I automatically write from time to time that AI “knows”, although knowing is itself a property ascribed to people. It is tempting. Even so, I consider this behaviour and this approach mistaken and dangerous.

On the other hand there are experts who will not agree with me. From the authors of studies such as this one through to Jan Tyl (he was one of the “opponents”), the author of the well-known DigiHavel or Digital Philosopher project, who recommended to me, for example, this study, which points out that multi-agent systems achieve better results when they use emotion emulation. The humanising of models and tools by developers may therefore look like a sensible path for the efficiency of the whole system, and yet I am convinced that this approach – pragmatic though it is from a developer’s point of view – calls for all the more caution in its social application. The fact that emulating emotions improves the behaviour of agents in virtual environments or speeds up the finding of solutions in complex systems does not yet mean that the same degree of anthropomorphisation is appropriate in public space, in everyday interaction with an ordinary user.

Developers do not humanise AI only for the model’s internal performance; at the same time they expose the public to a tool that comes across as a living being. And in doing so they introduce into society an entity that “poses” as a subject – even though it is not one. The consequences of this deception are not technical but psychological, cultural and moral. Humanising as a means of achieving higher performance in agent systems may be reasonable – provided it is at the same time framed by transparency, provided the user knows and is awarethat they are interacting with a tool. Otherwise, regardless of efficiency, it is a cognitive trap.

Sources and references

Is this article useful to you and are you citing it? Copy the citation

Kubíček, Michal. (2025, April 22). What the anthropomorphisation of artificial intelligence will lead to, and why we should be on our guard. www.kubicek.ai. https://www.kubicek.ai/en/what-the-anthropomorphisation-of-artificial-intelligence-will-lead-to-and-why-we-should-be-on-our-guard/