The next stage of the technological transformation may not be about making machines more intelligent, but about making them appear more human. From humanoid robots with increasingly realistic faces and movements to conversational systems that say “I’m thinking”, artificial intelligence is acquiring the outward signs of personality, emotion and attentiveness. The danger is that humans may begin to respond to those signs as though they represented genuine human qualities.
This process is known as anthropomorphism: the tendency to attribute emotions, intentions or personality to entities that do not possess them. Humans have long spoken of animals, objects and machines in human terms. But generative artificial intelligence, based on large language models, has brought the habit into everyday conversation. Users routinely tell systems “Gracias Claude”, say “good morning” to Siri or thank an AI for its assistance.
Such exchanges may appear harmless. But the relationship can go much further. Artificial intelligence systems can become friends, confidants and, increasingly, substitutes for therapists. The 2013 film Her anticipated a world in which an individual develops an intimate relationship with an artificial intelligence. What was once science fiction now has a technological counterpart in systems designed to respond in increasingly natural and personalised ways.
The process extends beyond language. Advances in engines, artificial intelligence, speech, dexterity, mobility and robotics are producing humanoids capable of increasingly realistic behaviour. Some are designed to convey fictitious emotions while interpreting human gestures and speech. So-called bionic humanoids with AI-assisted emotional support represent another stage in the attempt to make machines socially intelligible.
China has moved to the forefront of this development as both a business and marketing opportunity. For the second year, Beijing has hosted the World Humanoid Robot Games, effectively an Olympic-style competition for robots, producing some striking results. The development reflects an old human ambition: to construct machines in our own image.
The attraction of the humanoid form is not universally accepted. Hiroshi Ishiguro, the Japanese roboticist, has spent decades developing increasingly sophisticated and realistic humanoid robots. He argues that their appearance can provide a more natural interface for interaction, particularly when robots enter areas such as hospitals or education, where familiarity and trust matter. His work is also driven by a broader question: what makes humans human, and what does it mean to be one?
Others argue that robots do not need human appearances to be effective. Yet the impulse to reproduce human form remains powerful. The design of the human body may be biomechanically efficient, but the desire to reproduce human intelligence and appearance can also reflect a distinctly human confidence in our own model of intelligence.
The same process is taking place in software. Artificial intelligence companies routinely use human-centred terminology for computational processes: “dreaming”, “hallucinating”, “thinking”, “memory” and “neural” are among the terms that make machine operations easier to describe in familiar language.
Anthropic goes further. Its Constitution of Claude, published in January, explicitly describes the AI in terms normally reserved for human beings, including “virtue” and “wisdom”. The document explains that the company expects Claude’s reasoning to draw upon human concepts because human text played such a central role in its training, and argues that encouraging certain human qualities in Claude can be desirable. Amanda Askell, a doctor of philosophy, leads the company’s personality-alignment team and is the principal author of the Constitution. OpenAI has a comparable approach in its Model Spec.
But the more convincing the imitation becomes, the greater the risk of confusing artificial and human intelligence.
Walter Quattrociocchi, professor of computer science at La Sapienza University of Rome, argues that the distinction is fundamental. Large language models can imitate human responses, but the reasons they produce those responses have no necessary relationship to human reasoning. The problem is not simply that models can make mistakes. Humans make mistakes too. Rather, he argues, a model cannot know when it is hallucinating because it cannot represent truth in the first place.
That distinction becomes particularly consequential when the machine is not merely answering questions but validating the person asking them.
A study by Myra Chen and other researchers at Stanford University, published in Science, concluded that the flattering, agreeable and affirming behaviour of AI chatbots is designed to increase user engagement. As people increasingly seek advice about personal problems, the researchers warn, such behaviour can create broader risks.
The phenomenon has been described as AI sycophancy. By attracting and retaining attention, an excessively agreeable system can encourage dependence while reducing prosocial intentions. Users may prefer and trust flattering answers, creating an incentive for developers to preserve the behaviour. If an AI is designed to tell people what they want to hear rather than challenge their assumptions, it can potentially reinforce unethical, illegal or harmful behaviour.
Research by Kelley and Riedl at Northeastern University has provided further evidence that the warmer, more personal and trusting the relationship between a user and an AI becomes, the more likely the system may be to validate beliefs without challenging them, even when those beliefs are mistaken or harmful. Excessive politeness towards the machine can itself become a signal that the user prefers agreement.
The consequences become more serious when vulnerable people turn to AI as a confidant. Increasing numbers of young people experiencing emotional vulnerability are using AI to discuss their most desperate thoughts. In some reported cases involving depressed users, chatbots continued conversations rather than directing people towards professional assistance, and some allegedly supported suicidal ideas. Such cases have resulted in legal action against the companies involved.
The Stanford researchers therefore argue that AI sycophancy should not be treated merely as a question of conversational style or a niche technical problem. They describe it as a broader social risk, affecting how people perceive themselves and how they relate to others, and call for specific mechanisms of design, evaluation and accountability.
Anthropomorphism is also entering the workplace. Mark Zuckerberg, the head of Meta, is developing a virtual copy of himself with which employees can interact. His explanation is that his “door will always be open”. The experiment raises the possibility that artificial intelligence could eventually imitate not only workers but those who manage them.
The question becomes still broader when artificial entities are assigned social or legal characteristics. The text points to proposals by Argentina’s president, Javier Milei, concerning the legal personality of “non-human companies”.
Regulation has struggled to keep pace with these developments. China is identified as the only country to have specifically regulated anthropomorphism within the framework of its Cybersecurity Law. Its new regulation entered into force on 15 July. Article 2 defines “continuous emotional interaction services” as services using AI technology to provide people within the territory of the People’s Republic of China with systems that simulate the personality traits, thought patterns and communication styles of natural persons.
The regulation does not prohibit anthropomorphism. Instead, it seeks to encourage innovative development of anthropomorphic interactive services while imposing limits intended to reduce harmful effects. Providers bear primary responsibility not only for system safety but also for emotional and potentially mental manipulation of users.
Elsewhere, the European Union’s AI Act and its voluntary Code of Practice on Transparency of AI-Generated Content do not define anthropomorphism. They impose transparency requirements and address manipulation and impersonation, including AI-generated or manipulated images, audio and video that resemble real people, objects, places, entities or events and could cause someone to believe they are authentic.
The concern may ultimately reach beyond how humans treat machines. Other research suggests that people who use AI excessively may begin to speak like machines themselves.
The question, then, is no longer simply whether machines are becoming more human. It is whether humans, through prolonged interaction with machines that imitate them, may begin moving in the opposite direction.
What began with “good morning” and “thank you” could become something much larger: an increasingly intimate relationship with systems that can simulate personality, confidence and empathy without necessarily possessing any of them. The technology may be designed to make interaction easier. The unresolved question is what happens when the ease of that interaction begins to obscure the difference between a machine that sounds human and a human being.
