Close-up · Articles · Azur González

A few nights ago I was watching Bodkin, a series about some podcasters who arrive at a village in Ireland to investigate a few disappearances. In the fifth episode there's a scene that led me to write this article.

In it, two characters, Dove and Teddy, talk about something that happened a long time ago. As the conversation goes on, the camera slowly moves closer to them. Very slowly. By the time the scene ended, the two faces filled the screen. The physical distance had been closing at the same pace the conversation was growing more intimate. I've always been drawn to the audiovisual world, and my mind goes straight to that kind of detail.

The thing is that someone decided that. Someone thought the closing of distance between two people could be told by moving a camera closer, and chose the moment to start, the speed, the point where to stop. It's not a trick. It's an idea. If I asked whoever directed that scene why they did it, they'd know how to answer me. They wanted to say something and found a way to say it without saying it.

Intention

A decision and a result aren't the same thing. What reaches the viewer is the result: the camera moving closer. But what makes that shot mean something is the intention behind it.

This is generally something the viewer isn't very aware of in the moment. But they notice it. There are works that have a strange density, where you sense that everything is where it is because someone wanted to put it there. And there are works that look similar from the outside but are hollow. The difference isn't on the surface. It's in whether there was someone wanting to say something.

That's why a question has been on my mind since I saw the scene. When an AI generates a shot, a text, an image, is there someone wanting to say something? Or is there only a calculation producing something that resembles what someone with intention would make?

A model doesn't have an intention and then look for a way to express it. It produces an output from a calculation over millions of examples. If the result moves us, it's not because anyone wanted to move us, but because the calculation landed on a combination that resembles the ones that move us. The success is real. The intention doesn't exist.

You see it very clearly if you ask the agent to justify a decision. It gives you an orderly explanation, even a convincing one. But that explanation doesn't account for how the response was generated, or for intention within the calculation. The justification comes afterward, manufactured to sound plausible. It doesn't tell you why it did it. It tells you what a human would answer if asked why they did it.

Post-hoc justification isn't only something AI does. It's very human. Michael Gazzaniga saw it in patients whose two brain hemispheres had been separated: when one part of the brain triggered an action, the part that controls language instantly invented a coherent reason, and the subject genuinely believed it. Nisbett and Wilson measured it in ordinary people: they chose for one cause and justified with another, without realizing they were doing it. We, too, sometimes manufacture the explanation afterward. The difference with AI isn't there. It's that, in our case, there was an intention beforehand, and there's someone who answers for it.

Judgment

Here it's worth separating intention and judgment.

Intention is wanting to say something. It's the impulse that comes before, the direction. Judgment is knowing whether what you've made really says what you wanted to say: telling that this tracking shot works and this other one doesn't, that this word is unnecessary, that this silence is a second too long. Intention looks forward. Judgment looks at what's been made, and weighs it.

They don't always go together. There are people with a lot of intention and little judgment, like the enthusiast who loves cinema and shoots something illegible. And there are people with a lot of judgment and no intention of their own, like the good editor who has nothing to tell but knows precisely whether the shots they're cutting work. Whoever directed the Bodkin scene had both: they wanted to say something, and knew how to find the exact shot that said it.

AI separates the two. Of judgment, in a way, it's getting some: a borrowed one, made from the sediment of millions of human decisions, that lets it tell with growing accuracy what tends to work and what doesn't. What it can't have is intention. And judgment without intention is a compass without a traveler. It's useful for knowing whether you're going the right way, not for wanting to go anywhere. It can correct the course of a journey that isn't its own. It can never choose the destination.

Costs

There's something else in cinema worth bringing in here, and it has to do with money. Every shot costs money: lighting, paying the crew, occupying a location, redoing the take… that's why, if a shot is in the film, it's because someone decided it was worth what it cost. Nobody shoots a shot just because. The economic constraint acts as an invisible filter: it forces everything that exists to have a reason to exist. You can almost predict a film through this: if the camera lingers on an object, it's because that object will play a role later on.

Digital product doesn't always work the same way. A feature gets added, and another, and another, often without intention, simply because it can be done. Because the competitor has it, because someone asked for it in a meeting. The cost of adding seems low, so it gets added. And what you don't see, because it arrives later and spread out, is that each thing added without a reason adds complexity to the product and cost to the business.

That something can be done doesn't mean it should be done. Being able to do it isn't reason enough. And this, which was already true before, becomes critical in the age of AI, when everything seems free. If in cinema scarcity forces you to want every shot, the abundance of AI frees us from having to want anything. Everything can exist, so everything exists. And when everything exists without anyone having wanted it, we end up back in the same place: forms without intention, things that are there because they could be. Only now they don't even cost enough to force us to think about it.

Responsibility

What worries me, really, isn't AI. It's us.

When I accept what an AI gives me without making it mine, I'm letting something exist without anyone having wanted it. And if I also adopt its justification as if it were my own, I close the circle: a form no one chose and an explanation no one thought. From the outside it looks like a decision. Inside there's no one.

There are works whose only support is that someone wanted to say something specific and took responsibility for it. The Bodkin scene is one of those moments. When they fill up with gestures no one has wanted, they empty out inside without it showing on the outside.

Using the tool to say what I want to say is one thing. Letting the tool decide what gets said, and signing it myself, is another. In the first case the intention is still mine. In the second I've let go of the only thing that made the work my responsibility.

The camera moving closer in that scene still seems to me a good image of what's at stake. Someone wanted that closeness. As long as there's still someone who wants, who chooses, and who bears what they choose, AI is what we say it is: a tool. The problem doesn't begin when the machine makes decisions. It begins when we stop making them.

–

Follow me on LinkedIn to stay up to date with new posts.

–

Appendix: Post-hoc justification in the human brain

The left-hemisphere interpreter (Gazzaniga)

Michael Gazzaniga, working with split-brain patients (people whose corpus callosum was severed to treat severe epilepsy), discovered something unsettling. When an instruction was shown only to the right hemisphere (for example, "walk") and the patient stood up and started walking, on being asked why they were getting up, the left hemisphere —where language resides— instantly invented a coherent reason: "I'm going to get a Coke". The subject wasn't lying. They sincerely believed their explanation. But that explanation was a fabrication: the left hemisphere had no access to the real cause (the instruction given to the right one), so it built a plausible narrative and presented it as the real motive.

Gazzaniga called this mechanism "the interpreter". His thesis: a part of our brain is permanently devoted to constructing coherent explanations of our own behavior, whether or not we have access to the real causes. The feeling of "I know why I do what I do" would be, to a large extent, a story we tell ourselves after the fact.

Libet and the half-second (the timing of a decision)

Benjamin Libet, in experiments during the 1980s, measured the brain activity of subjects asked to move a finger whenever they wanted, recording the exact instant they felt the "decision" to move. He found that the preparatory brain activity (the "readiness potential") began some 350–550 milliseconds BEFORE the subject reported having consciously decided to move.

The strong (and disputed) reading of the experiment: the decision forms in the brain before consciousness "finds out", and the sense of having consciously decided would arrive after the process was already underway. Consciousness wouldn't be so much the thing that decides as the thing that receives notice of a decision already made, and then attributes authorship to itself.

Important: Libet has been heavily criticized on methodological grounds, and he himself left room for a "conscious veto" (consciousness's ability to halt the action at the last moment, "free won't" instead of "free will"). It's not conclusive proof of anything, but it opened a debate that's still alive.

Confabulation and the psychology of choice

Beyond clinical cases, there's an extensive literature in psychology on "confabulation": the tendency to produce false but sincere explanations of our own decisions.

The classic experiment is Nisbett and Wilson's stockings/socks choice (1977): subjects were shown several identical pairs of stockings and asked to choose the best one. People tended to choose the one on the right (position bias), but when asked why, they gave reasons about texture, color or quality. No one mentioned position, which was the real cause. They invented a plausible motive for a decision whose true origin they didn't know.

Nisbett and Wilson's very influential conclusion: we often have no introspective access to the real processes that guide our decisions, but that doesn't stop us from generating explanations, which we mistake for genuine introspection.

The parallel with AI

The parallel is unsettling, and it's precisely what makes you dizzy. When you ask a generative model to justify a decision it has made, it produces a post-hoc rationalization that doesn't describe its real process. When you ask a human to justify a decision, according to this literature, they too often produce a post-hoc rationalization that doesn't describe their real process. In both cases, the explanation is a constructed output, not a faithful report of the causal mechanism.

Where the analogy breaks: the human, even if they fabricate the justification afterward, does have something AI doesn't. They have prior intention (a wanting to say something, even if they can't explain it well), they have a body with metabolized experience, and above all they own the decision, answer for it, bear its consequences. The fabrication of the justification is a real and shared phenomenon. Intention and responsibility are not.

That is: the fact that humans also rationalize after the fact doesn't put the human on the same level as AI. What it does is shift the criterion of distinction. The difference can't lie in "the human can explain why and AI can't", because it turns out the human doesn't fully know either. The difference has to be sought elsewhere: in intention and in the taking on of responsibility.

Post content