If It Walks Like a Duck

A few days ago, I saw a LinkedIn post about somebody who'd bought what appeared to be a rather spectacular slice of chocolate cake. It sat in the display cabinet looking rich, dense and indulgent, with thick layers of chocolate sponge and icing suggesting it would require at least a mild sense of guilt to finish it. Unfortunately, when they cut into it, they discovered that much of the apparent bulk was empty space. The outer layers had been carefully shaped to create the impression of a generously sized slice, but there was remarkably little cake hidden beneath the surface. The disappointment wasn't that it tasted bad. The disappointment was that it wasn't what it appeared to be. The appearance had created an expectation that reality couldn't satisfy.

It's hardly a matter of national importance, and nobody's likely to launch a public inquiry into misleading confectionery. Yet the example stayed with me because it captures a much larger problem. We see something that looks substantial and naturally assume the substance exists beneath the surface. Most of the time that's a reasonable assumption. A slice of cake that looks full of cake is usually full of cake. The difficulty comes when appearance and reality part company. Once that happens, we're left relying on a judgement that may no longer be as reliable as we think.

Which brings me to an old expression:

If it walks like a duck, swims like a duck and quacks like a duck, then it must be a duck.

It’s one of those expressions that’s so familiar that we rarely stop to think about what it actually means, or indeed where it came from. We use it as though it were some sort of simple rule for establishing what something is, when in reality it’s a way of making an inference from what we can observe. We don’t necessarily know what something is, but if it has enough of the characteristics we associate with a particular thing, we make a judgement that it probably is that thing.

We don’t necessarily know what something is, but if it has enough of the characteristics we associate with a particular thing, we make a judgement that it probably is that thing.

Even the history of the expression turns out to be rather less certain than the expression itself suggests. It’s commonly attributed to the American poet James Whitcomb Riley, who died in 1916, and the version usually attributed to him is,

“When I see a bird that walks like a duck and swims like a duck and quacks like a duck, I call that bird a duck.”

The difficulty is that the quotation is impossible to locate in Riley’s published work, and there’s significant doubt that he actually coined it. A similar version is much more firmly documented in 1946, when Emil Mazey, secretary treasurer of the United Auto Workers, used the analogy while discussing whether somebody was a Communist.

“I can't prove you are a Communist. But when I see a bird that quacks like a duck, walks like a duck, has feathers and webbed feet and associates with ducks, I'm certainly going to assume that he is a duck."

A few years later, in 1950, the American ambassador to Guatemala, Richard Cunningham Patterson Jr., used another version when explaining his belief that the Guatemalan government was Communist.

“Suppose you see a bird walking around in a farm yard. This bird has no label that says ‘duck’. But the bird certainly looks like a duck. Also, he goes to the pond and you notice that he swims like a duck. Then he opens his beak and quacks like a duck.”

His point was that you didn't need a label saying "Communist" if the observable behaviour appeared to demonstrate it. Patterson's use helped popularise the expression in American political discourse.

The phrase, then, is really about inference rather than certainty. We see characteristics, recognise a pattern and draw a conclusion. In ordinary circumstances, that’s a perfectly sensible way to behave. If I walk past a pond and see a bird with feathers, webbed feet and a familiar quack swimming around on the water, I don’t need to establish its genetic history before deciding that it’s probably a duck. There’d be very little point in doing so because the evidence is good enough for the decision I need to make.

We see characteristics, recognise a pattern and draw a conclusion. In ordinary circumstances, that’s a perfectly sensible way to behave

For most of human history, that way of thinking has worked remarkably well because there was generally a connection between the thing we were observing and the characteristics we were using to identify it. The duck walked like a duck because it was a duck. It swam like a duck and quacked because it was a duck. A photograph showed something because something had stood in front of a camera, while a recording contained someone’s voice because that person had spoken. A witness described something because they had, presumably, seen it happen. None of these things was guaranteed to be true, and we’ve always had liars, forgers, actors, propaganda and people who simply get things wrong, but there was still a relationship between the evidence and the thing the evidence appeared to represent.

I think that relationship has now been broken in a way we’ve never experienced before. That’s why I keep coming back to the idea behind my earlier article, When Seeing Is No Longer Believing. When I wrote that, the concern was that things we’d traditionally treated as evidence, particularly photographs, recordings and video, could no longer automatically be trusted simply because we could see or hear them. At the time, I was thinking about the implications of increasingly convincing generated and manipulated material, but the more I consider it, the more I believe the bigger issue isn’t the technology itself. The technology is simply exposing a weakness in the way we’ve always thought.

For the first time, we’ve reached a point where something can have all the observable characteristics of a real thing without that real thing ever having existed. That’s quite a profound change, although I suspect most of us haven’t quite got our heads around it yet. I can show you a video of a duck walking towards a pond, swimming across it and quacking at another duck on the other side. You can watch the water move, see the feathers stirring in the breeze and hear the quack. You might quite reasonably conclude that there was a duck in front of a camera, except there might not have been. The entire thing could’ve been generated by a computer, and if it’s been generated well enough, there may be nothing in the video itself that tells you that.

I'm not entirely convinced that we've ever been as good at judging reality from appearances as we'd like to think. We often talk as though this is a brand new problem caused by AI, but history is littered with examples of people mistaking the convincing for the true. Victorian séances were enormously popular despite many of them relying on techniques that would struggle to fool a modern school pupil. Financial bubbles have repeatedly persuaded sensible people that prices can continue rising indefinitely. Entire industries have emerged around making things appear larger, more effective, more successful or more prestigious than they really are. Estate agents have been describing cupboards as bedrooms for longer than most AI companies have existed. Even food packaging has developed an art form around this. Open a bag of crisps and you'll often discover that a surprisingly large proportion of the contents consists of atmosphere. We seem endlessly surprised by these discoveries despite repeatedly encountering them. Perhaps the remarkable thing isn't that appearances can be deceptive. Perhaps it's how willing we remain to trust appearances when experience has spent centuries trying to warn us otherwise.

Perhaps the remarkable thing isn't that appearances can be deceptive. Perhaps it's how willing we remain to trust appearances when experience has spent centuries trying to warn us otherwise.

Anyway, back to ducks.

The duck test has failed, not because the duck has changed, but because the relationship between appearance and reality has changed. It would be easy to turn this into another article about artificial intelligence, because AI is obviously responsible for much of what’s changed. We can generate photographs of people who don’t exist, voices of people who never said the words we hear, videos of events that never happened and documents that appear to have been written by someone who never wrote them. We can also ask an AI system a question and receive an answer that sounds as though it came from somebody who knows the subject extremely well, even when the answer is wrong. If we make the story entirely about AI, however, I think we miss the more important point, because this is really about us.

AI has changed the environment in which we make judgements, but it hasn’t changed human nature. We still want shortcuts, make assumptions and look for patterns. We still tend to believe things that fit with what we already think, particularly when they’re presented confidently and attractively. We still respond to authority, status, appearance and repetition. In some respects, AI hasn’t created a new problem at all. It’s simply made our old ways of judging things much less reliable.

Think about what happens when you ask an AI system a question. You don’t see the process by which the answer has been produced, the uncertainty involved or the things it considered and rejected. What you get is a response, usually written in clear and confident language, and because it looks like the sort of thing an intelligent and knowledgeable person might write, it’s very easy to assume there must be intelligence and knowledge behind it. The difficulty is that a well-supported answer and a completely fabricated one can be presented with exactly the same fluency and confidence.

The difficulty is that a well-supported answer and a completely fabricated one can be presented with exactly the same fluency and confidence.

That’s something we haven’t had to deal with in quite this form before. If I ask a person a question and they give me an answer, I can use all sorts of other information to judge how much confidence I should place in it. I know who they are, I know something about their background, I can observe how they respond when I challenge them and I may have known them for years. None of those things makes them automatically right, but they provide context. With AI, much of that context disappears. We’re presented with the answer rather than the person, and because the answer is often so fluent, we can end up judging it on how well it’s written rather than on whether it’s actually supported.

That’s the new duck. It looks like knowledge, sounds like knowledge and behaves like knowledge, but that doesn’t mean it is knowledge. The appearance is convincing because it reproduces the characteristics we associate with an informed answer, yet those characteristics no longer guarantee that there’s any firm understanding, evidence or even truth behind it.

The same problem appears in rather less technical forms all over the place, which is why I don’t think we can solve it simply by teaching people how to spot AI-generated content. Even if we became extraordinarily good at recognising generated photographs and synthetic voices, we’d still have the underlying problem of believing things because they look convincing. LinkedIn provides a particularly rich supply of examples, although calling it rich may be a little generous.

I’ve lost count of the number of posts I’ve seen from people who appear to have discovered the secret to making vast amounts of money. There are claims of £100,000-a-month businesses, seven-figure consultancies, extraordinary career transformations and seemingly effortless success. There are people who apparently went from having no clients to being inundated with work, people who’ve discovered a simple formula for generating income and people who are more than happy to tell us how they did it, usually in exchange for something.

Some of these claims will be true. I’ve no reason to assume otherwise, and I certainly don’t want to fall into the trap of deciding that anybody who talks about their success must be lying. There is, however, a difference between not assuming that something is false and assuming that it’s true. That gap is important, yet it’s one we seem rather eager to step over whenever the claim is sufficiently attractive.

There is, however, a difference between not assuming that something is false and assuming that it’s true. That gap is important, yet it’s one we seem rather eager to step over whenever the claim is sufficiently attractive.

If somebody tells me they made £100,000 last month, I don’t know whether they did. If they show me a screenshot, I still don’t necessarily know. If they tell me how they achieved it, I know what they claim happened, but I don’t necessarily know that their explanation is correct. Even if the £100,000 figure is genuine, I don’t know whether it was revenue or profit, whether it happened once or happens regularly, how much was spent to generate it or whether the same circumstances could possibly apply to me.

The photograph of the duck isn’t the duck, just as the screenshot isn’t the business and the story isn’t necessarily the explanation. We seem to have become rather bad at distinguishing between evidence, interpretation and reality. We’re shown something that resembles proof and then supply the missing certainty ourselves, often without noticing that we’ve done it.

I sometimes think the modern world has become extraordinarily good at measuring the wrong things. Somewhere along the line, we've developed a habit of counting whatever happens to be visible and then quietly assuming that the visible thing must also be the important thing. Followers become a proxy for expertise. Views become a proxy for influence. Engagement becomes a proxy for value. Revenue becomes a proxy for success. Before long, the measure has replaced the thing it was supposedly measuring. It's rather like judging the quality of a book by weighing it. The number itself is perfectly real. The book genuinely does weigh a certain amount. The difficulty is that the measurement may have very little to do with the question we actually care about. Then again, perhaps that's unfair. At least a heavy book usually contains more paper. The relationship between LinkedIn engagement and wisdom sometimes appears considerably more tenuous.

Anyway, where was I?

Part of the problem is that we like certainty, particularly when the certainty is attractive. If somebody tells me they’ve found the secret to success and presents a convincing story, it’s much easier to accept it than to start asking awkward questions about sample size, selection, survivorship, costs, alternative explanations and all the other things that might spoil it. We do the same thing with statistics. A large percentage looks impressive, a graph that rises sharply looks important and a statistically significant result looks convincing. Sometimes all of those things are exactly what they appear to be. Sometimes they’re not, but the presentation encourages us to stop asking questions before we discover the difference.

That’s why the arrival of AI requires us to think differently rather than simply learn how to use another collection of tools. We need to become more conscious of the difference between observation and interpretation. We need to distinguish between something being possible, something being plausible and something being established. We also need to become more comfortable with uncertainty and, perhaps most difficult of all, better at recognising when we don’t actually know something.

We need to become more conscious of the difference between observation and interpretation. We need to distinguish between something being possible, something being plausible and something being established. We also need to become more comfortable with uncertainty and, perhaps most difficult of all, better at recognising when we don’t actually know something.

None of this means becoming suspicious of everything. The appropriate level of scepticism should depend partly on the consequences of being wrong. If a friend tells me they’ve had a miserable day, I don’t need documentary evidence, witness statements and a timeline of events before I listen to them. If somebody tells me that investing £10,000 with them will produce a guaranteed return, I’m probably going to ask rather more questions. That isn’t inconsistency. It’s applying judgement in proportion to the potential consequences.

We don’t always behave as though that distinction is obvious. We can spend an hour researching which television to buy and then accept a financial claim because somebody with a large following posted it. We can read three reviews before booking a hotel and then accept a statistic that supports our existing opinion without looking at where it came from. We can challenge an AI answer because we know AI sometimes hallucinates while accepting a human answer without checking because the person sounds authoritative. The technology may be new, but our inconsistency in deciding what deserves scrutiny certainly isn’t.

There are practical techniques that can help. One is lateral reading, where instead of spending all your time examining the material in front of you, you leave it and look elsewhere. Who is the author? What else have they written? Do other sources say the same thing? Can the original source be found? Is the quotation actually attributable to the person who supposedly said it? Does the impressive statistic still look impressive when you find the underlying numbers?

Another is the SIFT approach, which provides a simple way of remembering some of this. Stop before accepting the claim, investigate the source, find better coverage elsewhere and trace the claim back to its original context. The point isn’t to turn every Facebook post into a research project or spend three days tracing the origins of a picture of somebody’s lunch. It’s to develop the habit of stepping outside the information we’ve been given and asking whether there’s another way of looking at it.

One useful question is something we don’t ask nearly often enough: what would change my mind? If the answer is nothing, then I’m not really being sceptical. I’m defending a belief, and there’s a considerable difference between the two. That applies just as much to things I agree with as it does to things I don’t. If I see a statistic that supports an argument I already hold, I should probably be at least as interested in checking it as I would be if it contradicted me. Confirmation is comforting, but it isn’t evidence simply because it happens to confirm what I already thought.

what would change my mind? If the answer is nothing, then I’m not really being sceptical. I’m defending a belief, and there’s a considerable difference between the two.

There’s another useful question, particularly when we encounter extraordinary claims online: am I believing this because it's true, or because I want it to be true? That’s harder because it brings us into contact with ourselves. We aren’t neutral information-processing systems, however much we may occasionally like to imagine we are. We have hopes, fears, prejudices, experiences and preferences, and those things influence what we’re prepared to believe, what we’re inclined to reject and how much evidence we demand in either direction.

We therefore need to learn to be sceptical without becoming less human, and that may be the hardest part of all. There’s a temptation, when faced with synthetic information and increasingly convincing machines, to respond by becoming more machine-like ourselves. Check everything. Trust nothing. Demand evidence. Reduce every conversation to facts. Treat every assertion as a proposition requiring verification. Use one system to assess another system and then let an algorithm decide whether something is credible.

There’s some value in all of that, but taken too far it creates another problem. We could become very good at detecting whether something is genuine while becoming rather poor at dealing with genuine human beings. Human relationships don’t work like statistical tests. A person can tell you something that isn’t objectively verifiable and still deserve to be heard. Someone can be mistaken without being dishonest, remember an event differently without deliberately misleading you or say something emotionally rather than factually while still expressing something important.

If we approach every human interaction with the same suspicion we apply to an anonymous claim on the internet, we’ll lose something important. We need to be able to say, “I believe you,” even when we can’t prove what somebody has told us. We also need to be able to say, “I’m not sure,” without making the other person feel that we’ve rejected them. That’s where judgement comes in, and I’m not sure judgement can be automated.

Perhaps that’s one of the unintended dangers of AI. We spend so much time worrying that machines will become more like us that we don’t notice the possibility that we might start becoming more like machines. We could become more transactional, more suspicious and more focused on verification than understanding. We could become so concerned about being deceived that we forget trust is part of being human.

We spend so much time worrying that machines will become more like us that we don’t notice the possibility that we might start becoming more like machines.

In fact, we're already seeing signs of that. One of the more peculiar developments of recent years is the number of people who write something and then immediately ask an AI system whether it sounds human. I explored this in my article Get By With a Little Help from AI, where I reflected on the growing habit of asking machines to assess, validate or even improve our own communication. If somebody had suggested a decade ago that people would routinely ask artificial intelligence whether a piece of writing sounded sufficiently human, it would've sounded faintly ridiculous. Human beings have spent centuries writing letters, essays, books and articles without requiring a machine to certify their humanity. Yet here we are, increasingly seeking reassurance from an artificial system that the words we wrote ourselves resemble something a human might have written.

There is a curious irony in that. We worry about AI-generated text sounding too human while simultaneously asking AI to judge whether our own writing sounds human enough. In our efforts to distinguish ourselves from machines, we risk handing machines the authority to define what being human looks like. The technology may be useful, just as I argued in Get By With a Little Help from AI, but there is an important difference between using a tool to help us think and allowing the tool to become the judge of our humanity. If we're not careful, we may find ourselves measuring ourselves against the machine rather than the other way round.

There has to be a balance. We need evidence, but we also need empathy. We need scepticism, but we also need trust. We need to check things that matter, but we don’t need to interrogate every person we meet. We need to recognise that appearances can deceive without deciding that everything and everybody must therefore be deceptive.

We need to check things that matter, but we don’t need to interrogate every person we meet. We need to recognise that appearances can deceive without deciding that everything and everybody must therefore be deceptive.

That may be the real change AI is forcing upon us. For centuries, the world gave us a reasonably reliable relationship between appearance and reality. It wasn’t perfect, but it was good enough for the shortcuts we used to navigate our lives. We could see the duck and reasonably assume there was a duck. We could hear someone’s voice and reasonably assume someone had spoken. We could look at a photograph and reasonably assume that the camera had been pointed at something that existed. We can’t make those assumptions with the same confidence now.

That isn’t something AI can fix for us because it isn’t really a technological problem. It’s a problem of human judgement, and no amount of clever software will remove our responsibility to exercise it. The software may improve, detection tools may become more sophisticated and methods of verification may evolve, but none of those developments will relieve us of the need to decide what evidence means, how much confidence it deserves and what we should do when certainty isn’t available.

We don’t need to abandon trust, but we do need to become more thoughtful about how we grant it. The shortcuts that served us reasonably well for centuries now require a little more scrutiny, because appearances no longer provide the reassurance they once did. The question isn’t whether something looks convincing, but whether we have sufficient reason to believe that what we’re seeing reflects reality. In a world where authenticity can be manufactured and certainty can be simulated, judgement matters more than ever. Because something isn’t necessarily true simply because it looks as though it should be, even if it walks like a duck.