The Man Who Knew Too Much

Alfred Hitchcock's The Man Who Knew Too Much was made twice, first in 1934 and then, with James Stewart and Doris Day, in 1956. The story involves an ordinary couple who become caught up in events far beyond their experience after stumbling across information that puts them and their family in danger. Knowing something they probably wish they didn't know turns out to be rather less useful than they might have hoped. There's a difference between having information and understanding what to do with it, particularly when you don't fully understand the circumstances in which you've acquired it.

There's an obvious connection here with the world we live in today. We've got Google, social media, enormous quantities of data and, more recently, AI, all providing us with answers to questions we might previously have struggled to answer. We can look up medical symptoms, interpret statistics, challenge scientific opinion and produce a report that looks as though it's been written by someone who knows what they're talking about. The question is whether all this information has made us more knowledgeable, or simply more confident in what we think we know. After all, there's a difference between knowing something and understanding it, and an even bigger difference between understanding a little and assuming you know it all.

The question is whether all this information has made us more knowledgeable, or simply more confident in what we think we know. After all, there's a difference between knowing something and understanding it, and an even bigger difference between understanding a little and assuming you know it all.

Back in 2018, well before the proliferation of LLMs, I wrote an article called Everyone's an Expert, which explored our tendency to mistake access to information for genuine expertise. Eight years on, it seemed worth revisiting the subject. In this intervening period we've had a global pandemic, seen data become increasingly central to public debate and witnessed the arrival of AI, which has made it possible to generate an apparently authoritative answer to almost any question. The original problem hasn't gone away. We've simply acquired more sophisticated ways of convincing ourselves that we know what we're talking about.

So here's the original article but with a bit of a necessary update,

How many times has it happened to you? You get a strange twinge in the neck, a rash, or an unusually persistent headache. You try the ibuprofen, and maybe the tablets that you had left over from last time you had some unexplained pain, but nothing shifts it.

Suddenly, what was a mild annoyance becomes something more. We refer to our most trusted of practitioners - Google - and put in our symptoms. Before you know it, our exact symptoms come up. But it's not good news. Wouldn't you know it, I've got a rare form of Polio (I'm kind of aware that Polio isn't an issue in the UK anymore, but maybe I am unlucky). I double check, but to my horror, the addition of a rash means it’s probably fatal, and I'll be dead within the week. Tearfully I bid farewell to my wife and go to see the local GP for confirmation.

It’s a long wait in the waiting room. I try not to get too close to anyone in case I am contagious. I wonder whether they ever put a GP practice into lockdown, and how long it will take the governmental epidemiologists to get here. I wonder if I should warn the receptionist and stop anyone else visiting the surgery.

And then my name is called, and I stagger forlornly into the doctors office.

"What seems to be the problem?" she asks.

Not wanting to upset her, I don't mention Google. To stop her panicking (a screaming doctor does nothing for patient morale) I recount my symptoms one by one. I can feel the terror rise as I finish and wait for my worst fears to be confirmed. The moments tick by, but the doctor says nothing. Perhaps she has realised the truth and she is too frightened to speak. To confirm my suspicions, she starts tapping away at her keyboard - presumably emailing her colleagues that there is a very real crisis. Finally, she turns to me and utters the dreaded words

"There's not much I can do for you"

Well that's it then. I'll never get to finish the second series of the Handmaid's Tale. Why didn't I go on that holiday to Shanghai last year instead of to Butlin's? What use is saving money now? Is my life insurance up to date? What happens with my pension (if it’s worth anything at all).

"I'm afraid...you've been out in the sun a little too long. I suggest you drink plenty of fluids and take some paracetamol. There's no point in me giving you a prescription".

I stagger into the glorious sunlight and realise how silly I've been. I don't understand how I could make such an error. I did a GCE O'level in Human Biology, surely that's sufficient to be able to diagnose something simple like this - and I had the wealth of knowledge of Google behind me......

And then a revelation.... we have GPs, clinicians, nurses and others because they are specialists in their field. They have undergone years of training to think and act in a certain way. No matter how much I look over the notes I scrawled at 15, I will never be able to think like someone who has been appropriately trained. Google cannot give me knowledge or wisdom, it just provides a plethora of information.

we have GPs, clinicians, nurses and others because they are specialists in their field. They have undergone years of training to think and act in a certain way.

When I wrote the original article in 2018, Google was the obvious example. Today, I can ask AI to explain my symptoms, summarise medical research and compare possible diagnoses. It can present an answer in clear, accessible language and keep going for as long as I have questions. That's an impressive development, and there's plenty to be said for making specialist knowledge more accessible.

The problem is that an answer can be clear without being correct. AI can misunderstand a question, overlook important context, invent references or produce information that looks entirely plausible but isn't supported by the evidence. It can also be right often enough to encourage us to trust it when we shouldn't.

If I don't know enough about the subject to recognise an error, I'm not really in a position to assess the answer. Asking the AI to check its own work might simply reinforce the original mistake. Asking another system could produce a different answer, leaving me to decide which one to believe. I've moved from asking Google to asking a very articulate machine, but I still need the knowledge to judge what it tells me.

If I don't know enough about the subject to recognise an error, I'm not really in a position to assess the answer. Asking the AI to check its own work might simply reinforce the original mistake.

In Hitchcock's film, the information the characters acquire draws them into a situation they don't understand. In our case, the danger is often less dramatic, but the underlying problem is familiar. We can acquire information much more quickly than we can develop the experience needed to interpret it. Sometimes we know enough to ask a useful question. Sometimes we know just enough to reach the wrong conclusion with considerable confidence.

Then came COVID.

Suddenly we were all talking about epidemiology, infection rates, prevalence, excess deaths, R numbers, confidence intervals, modelling, vaccine effectiveness and statistical significance. Things that many people would previously have struggled to define became part of everyday conversation. Graphs appeared everywhere, and we all became surprisingly familiar with curves. We were shown forecasts, projections and estimates, sometimes accompanied by a level of certainty that wasn't justified by the evidence available at the time.

For a while, it seemed that everybody had acquired a working knowledge of epidemiology. Social media was full of people explaining what the statistics really meant, why the government had got it wrong and why the scientists were either completely correct or hopelessly incompetent. Sometimes the criticism was justified and raised important questions. At other times, the argument appeared to be based on a chart somebody had shared without reading the accompanying explanation, or a statistic stripped of the context needed to understand it.

There were genuine difficulties, and it would be unfair to pretend otherwise. COVID was a new and rapidly changing situation. Scientists didn't have all the answers at the beginning because much of the evidence simply didn't exist yet. Estimates changed as more information became available, models were revised and guidance changed as researchers learned more about transmission, immunity and the effects of different interventions. Some judgements made at the time didn't stand up particularly well when examined later.

None of that means scientific expertise is worthless. Science involves developing explanations, testing them against evidence and revising them when necessary. If our understanding of a new disease doesn't change as evidence accumulates, we have a much bigger problem than changing guidance. The trouble is that people often want certainty at precisely the point when certainty isn't available. Saying that we don't yet know the answer can sound like an admission of incompetence, even when it's the most honest response anyone can give.

The trouble is that people often want certainty at precisely the point when certainty isn't available. Saying that we don't yet know the answer can sound like an admission of incompetence, even when it's the most honest response anyone can give.

COVID didn't create the phenomenon of everyone being an expert, but it showed how easily uncertainty becomes an argument about trust. Experts can be wrong, and they should be challenged when their reasoning or evidence warrants it. The problem arises when the fact that an expert can be wrong is taken to mean that somebody without relevant expertise must be right.

And then, just as we were getting used to that, along came AI.

We can now ask an AI system to interpret a study, explain a statistical method, write code to analyse data or produce a report complete with recommendations. It can save time, help people learn and allow those without specialist training to explore subjects that would once have been difficult to approach. Experienced professionals can use it to test ideas and support their own work too.

The difficulty is that AI doesn't just help us find information; it can generate the information we want to present to other people. We can produce reports, articles, presentations and visualisations with relatively little effort. Some will be excellent, some useful with a little editing, and some wrong. A polished presentation tells us very little about the quality of the evidence behind it.

This matters because we use information to make decisions. A medical interpretation may influence someone's treatment. A financial analysis may influence an investment. An organisational dashboard may influence how resources are allocated or which services are expanded. Producing an answer quickly isn't enough. We need to know whether it's reliable and whether it addresses the question we actually need to answer.

Producing an answer quickly isn't enough. We need to know whether it's reliable and whether it addresses the question we actually need to answer.

The same problem exists with data, and this was the point I was making in the original article. In 2018 I wrote about people pulling out the bits of data that interested them and making an analytical diagnosis. Eight years later, we've got considerably more ways of doing it. We can build dashboards, download datasets, manipulate them in Excel, create visualisations in Power BI and ask AI to analyse the results. We can produce charts that look as though they've come from a professional analytical team, complete with explanatory notes and recommendations.

The barrier to producing an answer has become remarkably low. The barrier to producing a good answer hasn't.

I can take a spreadsheet and calculate an average. That doesn't necessarily mean I've understood what the average represents. If I'm looking at waiting times, for example, an average might give me a useful indication of performance, but it could also conceal a small group of patients who have been waiting far longer than everyone else. If I compare average waiting times between two hospitals, I need to consider whether they're dealing with similar patients, whether the figures have been collected in the same way and whether the periods being compared are appropriate.

I can create a beautiful chart, but that doesn't mean I've asked the right question. I can find a correlation, but that doesn't mean I've discovered causation. If activity increases after a new policy is introduced, that doesn't automatically mean the policy caused the increase. Perhaps activity was already rising, another change occurred at the same time, or the way activity was measured changed. The numbers won't necessarily tell me which explanation is correct, and a chart won't make the answer obvious just because it looks impressive.

The numbers won't necessarily tell me which explanation is correct, and a chart won't make the answer obvious just because it looks impressive.

Give an AI system a dataset and ask it to find interesting patterns, and it may well oblige. But if the data are incomplete, biased or poorly defined, it may produce an analysis that reflects those problems without identifying them. The resulting report could be very convincing while answering a question nobody needed to ask.

Everyone's an expert with data because we all did maths at school. We've all used Excel, seen a graph and got access to Google. Now we've got AI as well. We can produce a number, draw a chart and ask a machine to explain what it means. Yet the difficult part of analysis has never been simply carrying out the calculations. It's knowing which calculations are appropriate, understanding their assumptions and deciding whether the results provide a reasonable answer to the question.

Consider what happens when an organisation wants to understand why its performance has deteriorated. Someone produces a dashboard showing that the number of completed cases has fallen. The figures are correct, the chart is clear and the decline is real. An explanation is required, so someone suggests that staff productivity has fallen. Perhaps the organisation introduces a new performance target, restructures a team or changes the way work is allocated. All this might happen before anyone has established whether the decline is related to staffing, changes in demand, the complexity of cases, a recording problem or something else entirely.

The dashboard hasn't necessarily lied. The problem is that the interpretation has gone beyond what the data can tell us. We've moved from observing a change to explaining it, and then from explaining it to making a decision, without necessarily having the evidence needed for either step. Asking AI to produce recommendations doesn't resolve the problem. It may offer useful explanations to investigate, but those explanations still need to be checked.

We've moved from observing a change to explaining it, and then from explaining it to making a decision, without necessarily having the evidence needed for either step. Asking AI to produce recommendations doesn't resolve the problem.

This isn't an argument against using AI or dashboards. Managers need to understand the information they're using. Clinicians need to question the figures they're presented with. Researchers need to understand the limitations of their evidence. People sitting around boardroom tables need to be able to ask whether a number actually means what somebody says it means.

That's why we have analysts, statisticians, data scientists, mathematicians, epidemiologists and researchers. They've spent years learning how to think about evidence, uncertainty, bias, variation, sampling, measurement and causation. They understand that a calculation can be technically correct and still produce an answer that is misleading in practice. They've also learned that sometimes the data don't support a firm conclusion, more information is needed, or the honest answer is that we don't know.

We don't necessarily need fewer people who understand data because AI can do more of the calculations. We may need more people who understand what those calculations mean. If an AI system produces a statistical model, somebody still needs to decide whether the method was appropriate. If it identifies a pattern, somebody needs to consider whether the pattern is meaningful. Those aren't questions that can be answered simply by asking the machine to be more careful.

This is one of the reasons I'm still teaching statistics, and why I think this kind of training is particularly important in 2026. My Key Statistical Concepts session is designed to help people understand the ideas underpinning the analysis and interpretation of data. You don't have to become a statistician, any more than reading a book about medicine would turn me into a doctor, but understanding concepts such as averages, uncertainty, correlation and causation can help you question the figures you're presented with and recognise when a conclusion isn't justified by the evidence. That's particularly valuable when AI can produce a statistical analysis in seconds. The easier it becomes to generate an answer, the more important it is to understand whether that answer makes sense.

You don't have to become a statistician ... but understanding concepts such as averages, uncertainty, correlation and causation can help you question the figures you're presented with and recognise when a conclusion isn't justified by the evidence.

The point isn't to make everyone an expert in everything. It's to give people enough understanding to know what they can reasonably conclude from the information in front of them, and when they need specialist advice. Being willing to question an answer is a good thing. Assuming that our own interpretation must be right because we've found something on the internet, read a paper or asked an AI system to explain it is quite another.

Of course I can still Google my symptoms. I can ask AI about them. I can read medical research until three in the morning if I really want to. But none of that makes me a doctor (although I might be able to hold a more informed conversation with my GP than I could before.) In exactly the same way, I can open Excel, create a dashboard, download a dataset or ask AI to analyse it. Those are useful skills, but they don't automatically make me an analyst, any more than owning a stethoscope makes me a clinician.

And that's what connects the Hitchcock story with the point I was making back in 2018. Information can be valuable, but its value depends on how well we understand it and what we do with it. Knowing something doesn't necessarily make us an expert, and knowing a little can sometimes be more dangerous than knowing nothing, particularly when it gives us the confidence to make decisions we aren't equipped to make.

The world has changed considerably since I wrote Everyone's an Expert. We've lived through a pandemic and entered a period in which AI can generate an apparently authoritative answer to almost any question we care to ask. We're still very good at finding the answer we want, particularly when somebody, or something, is willing to give it to us. The difficult part is being prepared to question that answer, even when it agrees with what we already think.

We're still very good at finding the answer we want, particularly when somebody, or something, is willing to give it to us. The difficult part is being prepared to question that answer, even when it agrees with what we already think.

Maybe being an expert isn't about knowing everything. Perhaps it's knowing when you've reached the point where you need somebody who knows more than you do, and being sufficiently confident to admit it.

Hitchcock's characters found that knowing too much could put them in danger. In our world, the danger may be less dramatic, but the principle still applies. We can have access to more information than ever before and still misunderstand what it tells us.

Perhaps, in the end, the real problem isn't that we know too little, but that we know just enough to believe we've understood everything. And in a world where answers are so readily available, that may be the greatest danger of all: becoming the man who knew too much.