Human beings have always been impressed by convincing imitation.
A mechanical bird that appears to sing, a portrait whose eyes seem to follow you, or a computer that replies in fluent language can create the impression that something more is happening behind the surface. Modern artificial intelligence produces this impression exceptionally well.
A language model can write an argument, explain a scientific idea, imitate a style and respond to personal questions in a sympathetic tone. During a good conversation, it can feel as though the system has understood not only the words, but the person using them.
That feeling deserves examination.
Recent research from Anthropic identified a structure inside language models that appears to share some functional properties with the brain’s “global workspace”. This is a theory proposing that information becomes consciously accessible when it enters a limited shared space, allowing it to be reported, held in mind and used in deliberate reasoning [1].
The researchers found something loosely analogous in language models: selected internal information could be made available for reporting and used across different computational tasks. This is scientifically interesting. But it does not establish that the machine has an inner life.
The neuroscientist Anil Seth makes an important distinction between intelligence and consciousness. Intelligence concerns what a system can do: solve problems, manipulate information and produce useful behaviour. Consciousness concerns experience: whether there is anything it feels like to be that system [2].
A machine may display sophisticated information processing without experiencing pleasure, discomfort, curiosity, embarrassment or understanding.
The difference becomes clearer when we think about relationships.
The AI sceptic Cory Doctorow offers a useful example. Imagine that you know your spouse so well that you can occasionally finish their sentences. You may know what has happened that day, what they are worried about, what makes them laugh and why a particular memory has suddenly come to mind.
Now imagine instead collecting everything your spouse has ever said and constructing a statistical table showing which words most commonly follow the others. That system might also predict the end of some sentences.
As Doctorow puts it, without understanding your spouse, you might still make “some pretty shrewd guesses” about what they are going to say [3].
Knowing a person means much more than predicting their next word. It means understanding why they may be saying it now. It means recognising when the usual response would be wrong because something has changed. It means noticing hesitation, remembering history, caring about the outcome and sometimes choosing not to finish the sentence at all.
A language model is extremely good at the statistical version of this exercise. It has encountered vast quantities of language and learned complex patterns within them. Its predictions are far more sophisticated than a simple table of word frequencies, but prediction remains central to how it operates.
This does not make AI useless. Quite the opposite. The ability to reproduce some of the outward products of thought without thinking as humans do is a remarkable technological achievement.
But we should not confuse a successful imitation of understanding with a human relationship. The more immediate question may therefore be not whether AI has consciousness, but what its widespread use might do to ours.
Human beings have always outsourced parts of their thinking. Writing allows us to store memories outside the brain. Maps save us from reconstructing every route. Calculators perform arithmetic more accurately than most people can. Calendars relieve us of remembering every appointment. This is called ‘cognitive offloading’, and it is not inherently harmful.
The difference with generative AI is the range of mental work it can perform. We are no longer outsourcing only storage, navigation or calculation. We can now outsource the first draft of an argument, the interpretation of information, the organisation of an idea and sometimes the judgement of what matters.
The risk is not that using AI once will damage the brain. It is that habitual substitution may reduce the amount of effortful thinking we practise.
An MIT study of essay writing found differences in cognitive engagement between people working without external tools, those using search engines and those relying on a language model. The researchers raised the possibility of accumulating “cognitive debt” when people repeatedly delegate the work of forming and expressing ideas [4].
The study was limited and should not be treated as proof that AI makes people less intelligent. But the principle is plausible. Skills that are rarely exercised tend not to improve.
Research discussed by the American Psychological Association suggests a useful distinction between allowing AI to replace thinking and using it to support thinking. Passive reliance may contribute to loss of skill, while deliberate and structured use can help people test ideas, receive feedback and improve their work [5].
AI can therefore be either a crutch or a scaffold.
A crutch carries the weight so that we do not have to. A scaffold helps us reach somewhere we could not easily reach alone, while leaving us responsible for the structure being built.
The distinction often comes down to ownership.
Did you ask the machine to produce a conclusion and then accept it? Or did you use it to expose weaknesses in your own conclusion? Did it write something in place of you, or did it help you see the subject from another angle? Could you explain and defend the result without referring back to the machine?
These questions matter because thinking is not merely the production of an answer. It is the process through which we form judgement.
Struggle has a role in this. Searching for a word, attempting to organise an argument and discovering that an idea does not quite make sense can feel inefficient. But those moments are often where learning occurs.
The answer is not to reject AI. These systems are already useful for research, translation, accessibility, administration and creative exploration. Refusing to use them would be rather like rejecting calculators because mental arithmetic is valuable.
The better response is to remain clear about what the tool is doing and what we still need to do ourselves.
AI may finish our sentences. It may even finish them better than we would.
But fluency is not feeling, prediction is not care, and a plausible answer is not the same as understanding why the question mattered.
The danger is not simply that we will mistake machines for conscious beings. It is that, impressed by how easily they reproduce the appearance of thought, we may stop practising enough of our own.
—Professor David Nutt
References
[1] Gurnee, W., Sofroniew, N., Pearce, A., Piotrowski, M., Kauvar, I., Chen, R., et al. “Verbalizable Representations Form a Global Workspace in Language Models.” arXiv, 2026. Research published in association with Anthropic.
[2] Seth, A. “Once again we are told AI may be conscious—I study consciousness, and I have my doubts.” The Guardian, 15 July 2026.
[3] Doctorow, C. Commentary on artificial intelligence, prediction and consciousness. Instagram video interview.
[4] Kosmyna, N., Hauptmann, E., Yuan, Y. T., Situ, J., Liao, X-H., Beresnitzky, A. V., Braunstein, I., & Maes, P. “Your Brain on ChatGPT: Accumulation of Cognitive Debt When Using an AI Assistant for Essay Writing Task.” MIT Media Lab and arXiv, 2025.
[5] American Psychological Association. “How AI Is Reshaping Human Skills and Thinking.” Monitor on Psychology, July/August 2026.
[6] Macnamara, B. “AI ‘De-skilling’: What Happens When We Offload Our Work to AI?” Speaking of Psychology, American Psychological Association, 2026.
