Mind Your Language

I've been writing about artificial intelligence for around eighteen months now, which in AI terms feels both like a long time and almost no time at all. The pace of change has been remarkable. Things that seemed experimental when I first started exploring them are now becoming everyday tools. Capabilities that sounded like science fiction are appearing in products people are using at work and at home.

During that time I've watched the conversation around AI change almost as quickly as the technology itself. Initially, much of the discussion focused on what these systems could do. Now the conversation is increasingly moving towards what people think these systems are. And that’s where I think we need to be a little more careful.

I've been thinking, in particular, about the language we use around artificial intelligence because we've started using words that sound harmless but affect how we think about the technology itself. Over the last few years we’ve become increasingly comfortable describing artificial intelligence in very human terms. It isn't just journalists or marketers doing it anymore. Politicians do it. Technology companies do it. Commentators do it. People who should know better casually talk about AI thinking, understanding, learning, reasoning and even showing empathy. These phrases appear everywhere now, from technology conferences and interviews to LinkedIn posts and corporate marketing material Every time I hear those words, I wince a little. Not because I'm trying to be pedantic, but because they’re simply not true. The problem is that many of these words carry meanings that AI simply has not earned.

There’s no doubt that the technology has moved on at an astonishing pace. The latest generation of AI systems can write software, generate images and videos, analyse large volumes of information, summarise complex documents and hold conversations that feel remarkably natural. Some systems can now use other software, search for information, complete tasks and work through problems over extended periods. The technology has advanced at an astonishing rate, and it is genuinely impressive. Yet all those improvements have only encouraged people to attribute human qualities to machines that still don't possess them.

The technology has advanced at an astonishing rate, and it is genuinely impressive. Yet all those improvements have only encouraged people to attribute human qualities to machines that still don't possess them.

I think part of the problem is that humans have always been very good at seeing ourselves in things that are not human. We call it anthropomorphising. We talk to cars when they refuse to start. We apologise to computers when they freeze, even though they’ve never once accepted responsibility for anything. I imagine that my Satnav has a slightly disappointed tone when I ignore her instructions, and I still find myself feeling vaguely guilty, despite knowing perfectly well that there is no little voice inside the device wondering why I made a bad decision (I’ve even assigned her a gender which makes things worse).

With AI, however, the illusion is much stronger because the technology communicates using the one thing we associate most strongly with human intelligence: language. But capability is not the same thing as understanding.

Large language models do not understand what they are saying. They do not think in the human sense. They don't have opinions, beliefs, intentions or experiences. They don't wonder whether they have made a mistake or worry that they may have offended someone. They don't wake up with ideas or change their minds after reflecting on yesterday's conversation. Underneath all the polished conversation and remarkably convincing prose is still a statistical process. The model analyses patterns in unimaginably large quantities of data and predicts the output most likely to satisfy the prompt it has been given. That process has become extraordinarily sophisticated, but sophistication is not the same thing as understanding. That may sound like it reduces AI to something simple, but it really doesn't. Predicting language at this scale is an extraordinary achievement. The mistake is not in being impressed by what these systems can do. The mistake is instead assuming that producing human like output means there is a human like mind behind it.

This is where the confusion begins. Because the responses often feel so human, people naturally assume something human must be happening inside the machine. If you ask it to explain grief, it can produce something moving. If you ask for relationship advice, it often sounds thoughtful and compassionate. If you tell it you're having the worst day of your life, it will almost certainly reply with warmth and reassurance. It can be remarkably convincing. It may even be better written than something a tired or distracted human might say. Many of us have experienced conversations where a machine has been more patient than a person. That says something about the design of the system, but it also says something uncomfortable about us.

Many of us have experienced conversations where a machine has been more patient than a person. That says something about the design of the system, but it also says something uncomfortable about us.

An AI has been trained on millions of examples of human communication. It has seen how people express sympathy, encouragement and reassurance. It knows which words and phrases usually appear in those situations. It can reproduce the language of empathy extremely effectively, but it does not experience empathy. Convincing isn't the same as genuine. The system isn't listening to you. It isn't feeling concern. It isn't choosing to be kind. It is generating the response that best matches millions upon millions of examples of supportive language it has encountered during training. In many respects it is rather like an actor delivering a powerful performance. The audience may be moved to tears, but nobody believes the actor has actually experienced every tragedy portrayed on stage. An actor can deliver a convincing performance of grief without having experienced the exact circumstances being portrayed. Nobody watches a film and assumes the actor has actually lived through every event in the script. AI is different only because the performance happens in real time and responds directly to us.

Recent research has found that people often rate AI generated responses as more empathetic than those written by doctors or other professionals. That has produced countless headlines suggesting AI may be more compassionate than humans. I think those headlines miss the point entirely. The models haven't developed empathy. They've simply become exceptionally good at reproducing language that people associate with empathy. Those are two very different things. This is why some people have started saying things like, "AI understands me better than my colleagues," or "ChatGPT listens more than my friends." Perhaps sometimes those statements reveal something about our relationships with other people, but they do not prove that the machine cares. A system can produce the right words without having any awareness of why those words matter. Once we start talking as though the machine genuinely understands us, it becomes much easier to trust it in situations where trust should still belong to people.

A system can produce the right words without having any awareness of why those words matter. Once we start talking as though the machine genuinely understands us, it becomes much easier to trust it in situations where trust should still belong to people.

In healthcare, systems are already helping clinicians interpret scans, summarise patient records and identify possible diagnoses. Used appropriately, those are valuable tools. AI can help clinicians identify patterns in scans, summarise medical records and search through research much faster than any individual could. These are valuable capabilities. But they don't understand frightened patients sitting in consulting rooms. They can’t recognise hesitation, confusion or the subtle signs that tell an experienced clinician something isn't quite right. It can’t understand the lived experience behind the data. Clinical judgement has always involved far more than processing

Education has similar challenges. AI can explain difficult concepts, generate exercises and provide feedback at any time of the day. That’s useful, especially for people who struggle to access support. However, teaching isn’t simply about transferring information. Good teachers notice when confidence is falling, when someone needs encouragement or when a different explanation is required. They respond to the person, not just the question and change their approach accordingly. Machines can imitate personalised feedback, but they cannot build the relationships that make education effective.

The same issues appear in recruitment, lending, policing and criminal justice. AI systems are often described as objective because they are based on mathematics. But mathematics is only part of the story. The data used to train these systems comes from human decisions, and human decisions are rarely perfect. We bring our biases, assumptions, mistakes and prejudices into the information we create. AI doesn’t remove those problems automatically. Sometimes it simply repeats them more efficiently.

The arrival of AI agents makes this discussion even more relevant. Increasingly, AI is not simply answering questions but carrying out tasks. It can book meetings, write emails, search websites, analyse spreadsheets and interact with other software with minimal supervision. That feels like another step towards genuine intelligence, yet the underlying principle remains exactly the same. The system is still predicting actions based on patterns. It is not aware of the significance of those actions or the consequences they may have. My calculator is far better at arithmetic than I am, particularly when I am trying to calculate something quickly in my head and confidently produce the wrong answer. That doesn’t mean my calculator understands numbers. It means it is very good at calculation.

My calculator is far better at arithmetic than I am, particularly when I am trying to calculate something quickly in my head and confidently produce the wrong answer. That doesn’t mean my calculator understands numbers. It means it is very good at calculation.

Perhaps my greatest concern isn't that AI will become human. It is that we will increasingly behave as though it already has. We have a long history of attributing human characteristics to machines. We talk to Satnavs, apologise to vacuum cleaners and complain that the printer is deliberately refusing to work. Normally that's harmless. AI is different because people increasingly rely on it for advice, reassurance and decisions that genuinely affect their lives.

When things go wrong, and inevitably some things will, who takes responsibility? If we say the AI made the decision, accountability quietly disappears. Machines are not legally or morally responsible for anything they produce. Responsibility still lies with the organisations that build these systems, deploy them and choose how they are used. Loose language makes it far easier to forget that.

None of this is an argument against artificial intelligence. Quite the opposite. I use it regularly. I talk about it, write articles about it, teach people about it and I believe it has enormous potential when used appropriately. It can save time, improve productivity and help people tackle problems that would previously have taken days rather than minutes. It’s one of the most significant technological developments of our lifetime. But it also occasionally produces complete nonsense with absolute confidence, which is probably the most human thing about it. Anyone who has sat through a meeting where someone confidently explains something they clearly do not understand will recognise the experience.

So AI is one of the most remarkable technologies of our time. It will change the way many of us work. It will create opportunities and solve problems that previously seemed difficult. But we should be honest about what it is. AI is extraordinarily capable. But it is not a colleague. It is not a counsellor. It is not a conscience. It doesn't think, understand or care, no matter how convincing the conversation becomes.

The technology doesn't need us to exaggerate its abilities. It's impressive enough already. What it does need is for us to describe it accurately and mind our language, because the words we choose today will shape the decisions we make tomorrow.