The Responsibility to Think

Artificial intelligence has a major confidence problem. That's not because it lacks confidence, but because it has far too much of it. Ask it a question and, within seconds, it will produce something neat, fluent and strangely self-assured. It will sound like a person who has read every book in the library, memorised the index cards, and never once stood in a room wondering what they went in there for.

That’s part of the attraction, of course. But it’s also part of the problem. A sentence can be beautifully constructed and still be wrong. A paragraph can sound authoritative and still be resting on foundations made of sand. In a world where machines can produce plausible answers at industrial speed, critical thinking stops being a pleasant intellectual accessory and becomes something closer to personal protective equipment. Helmet and safety goggles on. Check the evidence before walking into the machinery. It's our responsibility to think, and it doesn’t disappear just because the machine has answered quickly.

For years, the workplace rewarded people who could find information, organise information, remember information and then present it in a way that looked reasonably convincing on a slide deck. (I say “reasonably” because some slide decks have clearly been assembled with little thought, usually involving multiple fonts, three shades of corporate blue and an irrelevant stock photograph of people pointing at a whiteboard.) AI changes all of that. It can summarise, draft, rephrase, compare, organise and polish. It can take a scruffy thought and dress it in a jacket. So the value of the human contribution starts to move. It’s no longer enough to ask, “Can I produce an answer?” The better question is, “Can I tell whether this answer deserves to be trusted?”

It’s no longer enough to ask, “Can I produce an answer?” The better question is, “Can I tell whether this answer deserves to be trusted?”

That’s where critical thinking earns its keep. It’s not about sounding clever in meetings, or beginning every contribution with “If I may play devil’s advocate.” Critical thinking is more useful than that. It’s the habit of asking what something is based on, what has been left out, whether the source is reliable, whether the conclusion follows, and whether we are quietly believing it because it happens to agree with us.

There’s also a big difference between being sceptical and being cynical. Cynicism is lazy. It folds its arms, raises one eyebrow and decides that everything is rubbish before the kettle has boiled. Scepticism does something harder. It stays interested and asks relevant questions. It’s willing to be persuaded, but it refuses to be mugged by a confident paragraph wearing a tie.

This is exactly the mindset we need with AI. Not worship. Not panic. Not the sort of wide-eyed enthusiasm that suggests every problem in civilisation can be solved if only we ask the chatbot to “think step by step”. It can’t. Sometimes it’ll help. Sometimes it’ll hallucinate with the serenity of a monk. Sometimes it’ll give you the right answer for the wrong reason, which is arguably more irritating.

The point isn’t to distrust it automatically, but to keep your hands on the steering wheel. AI can be a useful passenger. It can read the map, suggest a route and occasionally spot something you missed. But you wouldn’t let it drive you into a lake simply because it announced, in a calm voice, that the lake was probably a shortcut. It is our responsibility to think in the quiet, human act of noticing the water before the wheels get wet.

I rather like the idea of critical thinking as a filter. Not a brick wall, or a shredder. A filter lets useful material through while catching the bits that need a closer look. Is this source real? Is this claim supported? Is this statistic being used properly? Has the answer quietly wandered away from the question like a toddler in a garden centre? That last one happens more often than you may think. AI can produce an answer that’s adjacent to the issue rather than actually on it. It can sound intelligent while missing the point. So can humans, of course, and some do it professionally. But with AI the danger is scale. One weak assumption can now be wrapped in polished language, copied into a report, sent round a team and absorbed into a decision before anyone’s had time to ask where it came from.

A filter lets useful material through while catching the bits that need a closer look. Is this source real? Is this claim supported? Is this statistic being used properly?

So I would start with three questions. Is it reliable? Is it relevant? Is it valid? Reliable means I can trace it to something I would not be embarrassed to cite in public. Relevant means it actually answers the question at hand, not a slightly shinier question it found lying nearby. Valid means it tells enough of the story to support the conclusion. A single fact can be true and still be misleading if it has been dragged out of context. It is our responsibility to think that should begin with these unglamorous checks, which is why it’s so tempting to skip them and so dangerous when we do.

I learned this long before AI became the shiny new toy. Years ago, when I was working with health data in a local public health department, I remember looking at a set of breast cancer figures that appeared to tell a very clear story. But it’s that obvious conclusion that often makes an analyst suspicious in the same way a silent child makes a parent suspicious. On the surface, the numbers pointed in one direction. But when we started asking dull, unfashionable questions about definitions, collection methods, missing data and local context, the story became far less clear but probably more useful as we avoided needlessly spending lots of money on a health promotion campaign. That experience has stayed with me. I realised early on that data doesn’t explain itself – as an analyst that was my responsibility. An AI answer doesn’t explain itself either. Both can look confident while quietly relying on outdated or incorrect assumptions.

One of the common assumptions about AI is that because it involves mathematics, it must somehow be neutral. This is a charming thought, rather like assuming a photocopier can’t reproduce nonsense because it has clearly labelled buttons. AI is trained on human material, and human material contains all the usual human clutter and bias: prejudice, fashion, bad evidence, overconfidence, historical blind spots, cultural assumptions and the occasional sentence that should’ve been stopped at source.

AI is trained on human material, and human material contains all the usual human clutter and bias: prejudice, fashion, bad evidence, overconfidence, historical blind spots, cultural assumptions

But the machine isn’t the only thing that needs watching. We do too. If I ask AI to help me support an argument I already like, it’ll probably oblige. It’ll fetch my slippers, light the fire and tell me my opinion has excellent structure (that’s figurative at the moment…). That can feel very comforting. It can also be disastrous. Critical thinking means asking the awkward follow-up: what would prove me wrong? Can you give me the counter-argument?

This is where the real work begins. Not in getting the answer you wanted, but in testing whether your answer survives contact with reality. If it collapses at the first sign of contrary evidence, it was never an answer. It was just an opinion with delusions of grandeur.

AI also rewards precision. Ask a vague question and you often get a bowl of beige porridge. Ask a sharper question and the answer begins to improve. Ask it to identify assumptions, challenge its own conclusion, compare competing explanations, separate facts from opinions, or show what evidence would change the answer, and suddenly it becomes much more useful.

This matters because the aim isn’t to outsource thinking. That’s the trap we fall into if we’re not careful. If we treat AI as some sort of oracle, we become passive consumers of fluent output. If we treat it as a thinking partner, we remain responsible for the work. It can suggest. It can provoke. It can organise. It can even irritate us into seeing something more clearly. But the judgement still belongs to us. The responsibility to think can’t be delegated, no matter how politely the software offers to help.

We shouldn’t ask AI, “Is this right?” and leave it there. Instead we should ask, “What are the weakest parts of this argument?” “What evidence would an informed critic expect to see?” “Which assumptions are doing the heavy lifting?” “Where might this fail in practice?” “What would a sceptical reader object to?” These are much better questions.

There are settings where this really matters. In research, a weak source can distort a conclusion. In healthcare, a missing context can affect patient safety. In public policy, an elegant summary can hide the people who are most likely to be affected by a decision. In consultancy, training or analysis, a confident answer that hasn’t been properly tested can become tomorrow’s embarrassing correction. The responsibility to think is not an abstract virtue in those settings. It’s part of the duty of care.

That’s why critical thinking belongs with the other human skills we keep being told will matter in the AI age. Curiosity asks what else might be true. Courage allows us to challenge the convenient answer. Communication helps us explain uncertainty without sounding as though we have misplaced the plot. Compassion reminds us that decisions land on real people, not spreadsheets. Critical thinking gives all of this a spine.

It’s also a useful defence against the great modern disease of sounding informed. We’ve never had more access to information, and yet we’re constantly at risk of mistaking access for understanding. Reading a summary, or following a cheat sheet, isn’t the same as understanding the subject. Asking AI for a briefing isn’t the same as doing the thinking. It’s a start. Sometimes it’s a very good start. But it’s still only a start.

My own rule for checking AI content would be simple. Treat every important AI answer as a draft, not a decision. Read it once for usefulness, and then again for any assumptions. Once you've done that ask what would have to be true for the answer to be safe, fair and sensible. If the answer relates to data, check the data. If it relates to evidence, check the sources. If it relates to people, check whether the people have somehow disappeared from the reasoning, which is something systems are rather good at doing when nobody is watching. That small, human pause is where the responsibility to think becomes practical.

ask what would have to be true for the answer to be safe, fair and sensible. If the answer relates to data, check the data. If it relates to evidence, check the sources. If it relates to people, check whether the people have somehow disappeared from the reasoning

None of this means we should be frightened of AI. I use these tools because they’re useful. They can accelerate thinking, widen a search, suggest structure and help us get unstuck. But useful isn’t the same as infallible. A ladder is useful. So is a chainsaw. I wouldn’t recommend using either without paying attention. (see my article Babies with Power Tools)

The future won’t belong to the people who can generate the most words in the shortest time. The future will belong to people who can ask better questions, notice weak evidence, challenge lazy assumptions, and recognise when a polished answer is trying to smuggle nonsense through customs.

That, to me, is the real human advantage. It’s not that we can out-perform AI, because in most cases we can’t. We also can’t pretend AI it isn’t changing things, because it obviously is. Our human advantage is that we can, and should, care about whether an answer is true, whether it’s fair, whether it’s useful, and what might happen if we act on it. Critical thinking isn’t a gloomy brake on progress. It’s what stops progress from driving confidently into a hedge.

AI may give us speed. It may give us structure. It may even give us the occasional moment of irritating brilliance. But the responsibility to think remains firmly ours. That, really, is the point. Not because AI is useless, and not because humans are wonderfully rational creatures who never make a mess of things. But it’s because thought carries consequences, and consequences require responsibility.

The AI models can help us produce answers. But they can’t take on the human responsibility to think and decide whether those answers are good enough to trust.