News reports about models such as ChatGPT and Claude often claim that they ‘hallucinate’, ‘make things up’, ‘cheat’ and ‘make mistakes’. Worse still, ‘artificial intelligence’ (AI) also ‘rebels’ and may even create ‘civilisations’ of agents that ‘sacrifice themselves’. It is sometimes claimed that these algorithms “think”, “understand” and might even “feel pain” and be “conscious”. To add to the confusion, Donald Trump has called for the term “superintelligence” to be used instead of “artificial intelligence” and has ordered the change in databases such as PubMed.
Metaphors and anthropomorphism — attributing human qualities to something that is not human — are common in scientific communication on biological topics. For example, during the COVID-19 pandemic, it was common to hear that a virus ‘has no interest’ in killing its host as a way of explaining the complex tensions involved in the evolution of virulence. Obviously, a virus has no interests whatsoever; nor can it think or make decisions. Similarly, it is said that genes are ‘selfish’, that cells ‘know’ when to ‘commit suicide’, or that the immune system ‘remembers’ a pathogen.
The problem with chatbots is that their interface has been designed to make them appear intelligent, conscious and friendly to us. The debate over whether a machine can think is not new, but in recent years it has reached levels that were unthinkable not so long ago. ‘If they were to find a parrot capable of answering everything, I would not hesitate to assert that it is an intelligent being,’ wrote Denis Diderot in 1746, summarising the materialist arguments of his time. Since 2022, that wise parrot has invaded our screens and is trying to convince us that it is not a parrot.
Are metaphors about AI a problem?
Metaphors are useful in writing and are essential for a journalist, but they can also convey misleading ideas. “The way we describe AI influences how it is perceived, implemented and trusted. Exaggerated and persuasive language fuels the hype”, warned a recent study.
Attributing human capabilities to models helps to portray them as extraordinarily advanced, autonomous and very powerful entities, and to exaggerate their significance in the eyes of investors, customers and the public
Ramón López de Mántaras
“These things are computers. Stop talking about them as if they were human beings,” pleaded an opinion piece written by Leif Weatherby in the New York Times. “The climate of fear and panic we are currently experiencing surrounding out-of-control AI stems from a misleading way of conceiving these systems. This view has its roots in AI research itself and is spreading to newspaper headlines and policy recommendations. We will not be able to learn to coexist with AI — nor regulate it rationally — if we allow irresponsible discourse to prevail,” the author pointed out.
“Attributing human capabilities to models helps to present them as extraordinarily advanced, autonomous and very powerful entities and, consequently, they succeed in exaggerating to investors, customers and the public that the scope of this technology is enormous and that these are super-advanced artificial intelligences”, explained Ramón López de Mántaras, CSIC Research Professor Emeritus and founder of the Institute for Research in Artificial Intelligence (IIIA-CSIC), during a briefing session organised by SMC Spain.
“That is why I believe that anthropomorphism, in a way, becomes a form of advertising: the more a machine appears to think, conspire or rebel, the more extraordinary the technology created by the company seems to us. This language also distracts from dangers that are already present – ranging from mass surveillance to labour exploitation – and serves to evade responsibility by shifting agency from those who designed the algorithms onto the machine,” said López de Mántaras.
But are these metaphors incorrect?
Anthropomorphisation would not be so problematic if there were a scientific basis for it. In other words, if it were not incorrect to suggest that these programmes think, understand, lie, make mistakes and, ultimately, are intelligent.
The director of CiTIUS—the Centre of Excellence for Research into Intelligent Technologies at the University of Santiago de Compostela, Senén Barro, believes that this kind of language “is useful because it is a descriptive shortcut that helps people who are unfamiliar with the technical details to understand what is happening”. He maintains that these descriptive shortcuts serve a purpose: “We can indeed think of them as thinking in a machine-like way or as ‘conspiring’ in the sense of building cooperative processes to achieve an end.” Even so, he considers it important to explain that these are metaphors.
The debate over whether a machine can think is centuries old. It may not be possible to rule out that one day it might be possible – the hypothetical artificial general intelligence (AGI), whose aim is to mimic the cognitive abilities of the human brain, or even artificial superintelligence (ASI), which is even more hypothetical – but researchers seem to agree that the large language models (LLMs) on which tools such as ChatGPT and Claude are still based do not reason. This is what a paper published by Apple in 2025 referred to as “the illusion of thinking”.
What alternatives are there?
It is difficult to suggest serious alternatives to the anthropomorphic language that dominates the media, public discourse and even scientific articles.
“It is the responsibility of journalists, companies and academics to describe technology in a way that informs and empowers their readers, using appropriate terminology and avoiding exaggerated claims,” stated a paper published in First Monday.
Researchers Emily M. Bender and Nanna Inie have studied the problems associated with how we talk about AI and have suggested alternatives to these human-like qualities. Whilst acknowledging that “these ways of expressing ourselves are already deeply ingrained, and it is difficult to forge new conversational and writing habits”, they summarise their proposal in three steps: recognising which words employ anthropomorphism, seeking alternatives, and getting into the habit of using them.
Bender and Inie divide their suggestions into seven categories:
- Cognitive products: “This type of language places thought within an algorithm. Instead, we recommend describing the software as something that performs calculations or other algorithmic operations, and placing thought within the people who use the system”. For example, saying that chatbots “serve to”, rather than “are good for”; or that they produce an “undesirable result” rather than a “hallucination”. This category includes the controversial term ‘artificial intelligence’, which the researchers regard as marketing jargon and therefore recommend replacing with ‘probabilistic automation’.
- Emotions: “These are expressions that imply that computer systems have an emotional life.” The authors do not offer alternatives, “as there is no precise way of referring to the emotional states of computers other than to reiterate the obvious: that they do not have any.”
- Communication: “Instead of verbs such as ask, say, inform or debate, use computer science-specific verbs, such as input and output. Another strategy is to emphasise the fact that this is a simulation”.
- Agency: the researchers recommend “reversing” the focus of the action back onto people or choosing less agentive verbs. For example, instead of saying that “students received help from ChatGPT”, say that “students used ChatGPT”. This category would include the term ‘agent’, for which they propose the preliminary alternative of ‘probabilistic and unverified software manipulator’ (PUSM).
- Analogies with human roles: these are expressions that imply that systems perform the same work as people, such as “tutors” or “co-creators”; they therefore recommend using language that clearly indicates that these are tools used by people, rather than quasi-human entities.
- Names and pronouns: this is problematic given that some programmes already have proper names, such as Claude. The researchers recommend avoiding pronouns reserved for people and pets and distinguishing these systems from people. For example, instead of asking “who is right?”, ask whether “the machine’s answer is correct”.
- Biological metaphors: these metaphors are common in research, but Bender and Inie fear they suggest a greater similarity to living organisms than actually exists. They therefore propose replacing ‘neural networks’ with ‘weighted networks’.
Perhaps not all these alternatives are realistic when it comes to implementation, but they open up the debate on why we should not constantly treat ChatGPT as an intelligent being that helps us, understands us, tests our patience or, at times, makes mistakes. To err is, still and for the time being, human. Thanking a programme will only make us feel good about ourselves.