I first came across Kant at school. I cannot pretend that, at that age, I understood the full architecture of Critique of Pure Reason. I remember the title, the seriousness of the language, and the feeling that I was being invited into a conversation that had begun long before I arrived and would continue long after I had left the classroom.

What stayed with me was not a complete understanding of Kant’s philosophy, but a question that seemed larger than the book itself: before we make claims about the world, how certain can we really be that we know what we think we know?

For many years, that question remained somewhere in the background. My working life took me into a much more practical environment: projects, clients, operations, contracts, commercial decisions and the type of situations where the consequences of a bad decision do not remain on paper for very long. Kant was not part of my daily vocabulary.

Then AI became part of daily business life, and I found myself thinking about him again.

Not because Kant had anything to say about computers, algorithms or language models. He obviously did not. But because the more I see AI producing answers that are fast, polished and increasingly persuasive, the more I return to the question at the centre of his work: what makes an answer worthy of being treated as knowledge?

We are becoming used to asking AI almost anything. We ask it to review contracts, produce reports, compare options, draft procedures, analyse risks, prepare presentations, interpret data and challenge our thinking. In many cases, it is genuinely useful. I use it myself, and I would be dishonest if I said that it has not improved the speed with which I can work through ideas, documents and possible scenarios.

What concerns me is not the existence of the tool. What concerns me is the ease with which we are beginning to confuse a fluent answer with a justified one.

The world in which Kant wrote was very different from ours, but it was also a moment of enormous confidence in reason. For much of European history, faith, religious tradition and inherited moral frameworks had played a central role in how people understood reality and their place within it. The Enlightenment did not erase those traditions overnight, nor did it simply replace faith with reason. It did, however, bring a new confidence in science, observation and the capacity of the human mind to explain the natural world.

Kant belonged to that age, but he was not prepared to let reason make unlimited claims on its own behalf.

That is what I find so relevant today.

The title Critique of Pure Reason can sound as though Kant was attacking reason. He was not. He was trying to examine it carefully, almost as one would examine a structure before asking it to carry more weight. He wanted to understand what reason could properly establish, what it needed experience for, and where it began to move beyond what it could genuinely justify.

His concern was theoretical reason: the faculty we use when we try to know and explain the world. Kant accepted that reason and science can give us real knowledge, but he also believed that human beings become careless when they assume that the mind can answer every question simply because it can formulate an argument.

He was especially concerned about the temptation to reach beyond possible experience and speak with certainty about matters that cannot be settled in that way, such as God, the soul, freedom or the universe as a complete whole. His point was not that these questions are unimportant. On the contrary, they are among the most important questions human beings ask. His point was that we need to be honest about the kind of answer we can claim to have.

That is a very different position from saying that reason is useless. It is a call for intellectual discipline.

Kant’s explanation of how knowledge works is complex, but the basic idea is easier to understand than people sometimes think. We do not simply receive reality as if the mind were an empty container waiting to be filled. We experience the world through our senses, but the mind also gives form to that experience. Space and time, for Kant, are not simply things we discover outside ourselves; they are part of the way we are able to experience anything at all. Then the understanding brings concepts to what we experience, including concepts such as cause, substance and relation.

Reason comes into the picture by looking for a wider unity. It wants to connect individual facts, judgments and experiences into something more coherent. It asks what lies behind events, what explains them, and whether there is a larger order that makes sense of everything.

That capacity is one of humanity’s great strengths. It is also where trouble can begin.

Reason naturally wants complete answers. It dislikes loose ends. It wants to move from one explanation to a final explanation. But Kant warned that there is a point at which reason may go further than the evidence allows. It may create a picture that is elegant, internally consistent and intellectually satisfying, while still claiming more than it has the right to claim.

When I look at AI, this is the parallel that comes back to me.

The AI systems most of us use at work are extraordinarily good at producing language that appears coherent. They can take a large amount of information, identify patterns, generate alternatives, organise a line of argument and produce conclusions in a tone that often sounds more confident than the person reading them.

This can be extremely helpful. A strong AI prompt can give me a better starting point for a difficult piece of work. It can show me where my thinking is incomplete. It can help me identify questions I had not considered, prepare a first draft or structure a discussion that would otherwise take me much longer to organise.

But I have learned not to confuse that usefulness with understanding.

A language model can produce a convincing explanation of a vessel operation, a contractual dispute, a procurement decision or an AI risk assessment. It may even produce something that looks more complete than the first draft prepared by a busy manager. Yet the system has not lived inside that situation. It has not experienced the pressure of an operation changing under difficult conditions. It has not sat across the table from a client whose relationship with the company has been built over years. It has not had to decide whether a technically correct answer is also commercially wise, operationally workable or ethically acceptable.

I am not making a grand claim here about whether machines may one day develop forms of understanding that we cannot yet imagine. I am talking about the systems being introduced into organisations now, and about the way people are beginning to rely on them.

In that context, I think the difference matters.

After many years in operational and commercial environments, I have seen that information and judgment are not the same thing. Information is essential, but it is rarely enough on its own. A procedure may be technically correct and still fail once it meets the reality of weather, time pressure, fatigue, personalities, client expectations and incomplete information. A contract may give a company the legal right to take a particular position, while experience tells you that enforcing it in that way will damage a relationship that took years to build.

Those decisions cannot be made by looking only at what is written in front of you. They require context, responsibility and the ability to understand consequences that may not be visible in the document.

This is why I become cautious when someone says, “The AI has analysed it.”

That sentence may mean something sensible. AI may have helped a team review a large amount of material faster. It may have identified an issue that needs further investigation. It may have given someone a useful first draft or helped them see an alternative that had not been considered. But it may also mean that a polished answer has been accepted too quickly because it arrived in good English, with a strong structure and an apparently logical conclusion. The danger is not that AI may be wrong. Human beings are wrong all the time. We make assumptions, misread situations, rely too heavily on past experience and occasionally become attached to a conclusion before the evidence fully supports it.

The more serious risk is that AI can be wrong in a way that appears authoritative.

It can fill gaps with plausible language. It can present an assumption as though it were a fact. It can produce a recommendation that makes sense in general terms but fails when it is applied to the specific client, asset, project, contract or regulatory environment in front of us. And because the answer is written so well, people may not challenge it with the same rigour they would apply to a colleague. For me, this is where AI governance becomes much more than a technology topic. It becomes a question of leadership and professional responsibility. Governance should not mean preventing people from using AI, nor should it become another policy document that sits in a shared folder while employees continue working as they always have. Good governance should help an organisation decide where AI is useful, where human review is essential, where sensitive information must be protected and where accountability cannot be diluted.

The key question is not whether AI may be used. That is already happening.

The key question is what role we are allowing it to play in the way we form judgment. There is a real difference between using AI as an assistant and allowing it to become an authority. An assistant can help me think more clearly, organise information, challenge my assumptions and prepare better questions. An authority is something I accept because I believe it knows more than I do, even when I cannot properly explain how it reached its conclusion. The first can strengthen professional judgment. The second can quietly replace it. Kant’s distinction between appearances and things in themselves is also useful here, although it needs to be handled carefully.

He did not say that the world we experience is unreal. He said that human knowledge is limited to the world as it appears to us through the conditions of our own experience and understanding. We should be cautious when we claim direct knowledge of reality beyond those conditions. That thought has stayed with me because AI creates its own version of this problem. It gives us an output based on the material it has been given and the patterns it has learned from. What we see is an answer. We do not automatically see the missing context, the weak assumptions, the gaps in the source material or the alternative interpretation that was never considered. The output may be useful. It may even be excellent. But it is still an output that requires judgment.

That, for me, is the real lesson Kant brings into the AI conversation.

He does not tell us to distrust reason. He tells us to respect it enough to understand its limits. He asks us to be disciplined before we become too certain. He reminds us that an elegant explanation is not the same as a justified conclusion. Perhaps AI is forcing us to rediscover that lesson. We will have better tools. We will receive answers faster. We will automate work that previously required significant time and effort. I believe organisations that ignore this change will place themselves at a disadvantage. But the organisations that succeed will not simply be the ones that adopt AI fastest.

They will be the ones that know when to trust it, when to challenge it, and when a human being must still be prepared to say: “I understand the answer, but I am not yet convinced by it.”

Because that is where responsibility begins.

An answer becomes valuable not because it sounds intelligent, but because somebody who understands the real world behind the question has tested it, challenged it and decided that it deserves to be trusted.

That responsibility will remain ours.

This article is also published on Medium as part of my public research and writing.