"Reality" Bites

Sometime during 2025, my Apple Watch started trying to kill me.

I work pretty hard to stay fit. In fact, I am kind of an overachiever at it. Two days a week, I do a weightlifting and core workout for about an hour and a half. On the other five days, I do intense cardiovascular exercise of some kind, ideally for an hour, though I sometimes let myself do a little less. To make sure my cardio workout is sufficiently intense, I measure my heart rate using a breast strap on the days when I am not running. If I run, I keep an eye on my pace. I was pretty satisfied for years with the sort of feedback these metrics were giving me.

Then, in early 2023, I bought an Apple Watch. A few months after I got it, my son pointed out to me that it tracks something called VO2 max—that is, the maximum volume of oxygen that your cardiovascular system can process per kilogram of your weight per unit time. He took a look at the number Apple Watch was telling me based on my exercise since I’d acquired the watch: it was 50. “That’s really excellent!” he said. “Oh,” I said, pleased with myself, “Is my VO2 max as good as yours?” He laughed. “I mean, it’s excellent for someone your age.”

I liked the additional positive feedback about my excellent physical condition that VO2 max was telling me. So I started paying attention to it, along with the other metrics I was paying attention to. When I had some surgery done in the fall that forced me to take a break from cardiovascular exercise for a couple of weeks, I noticed that my VO2 max had dipped considerably. That was frustrating; however, receiving this information steeled me to make a comeback when exercise was once again permitted. And, sure enough, a few months later, my fitness had recovered to its original level. Big win!

So when, as I turned the corner into 2025, my VO2 max gradually dropped to 46, I didn’t despair. It was winter, and my fitness could be expected to fluctuate. When it stayed at that level until November, I felt a bit despondent, but not overly so. After all, I was getting older. It was only when, after the first week of November, that my VO2 max took an unprecedented nosedive to end the year at 40 that I began to panic. It didn’t help, of course, that the Apple Watch health app decided at this moment to notify me: “Your fitness level has declined over the past six weeks!” (The exclamation point here is mine, not the app’s.)

What had happened to me? I was still doing five days of cardio per week, and my running times were as fast as they had been I was turning in VO2 max measures in the high 40s. Was Apple Watch noticing some trend in my fitness that these other indications failed to capture? And... was I perhaps experiencing some shortness of breath? Was that an irregular heartbeat I noticed on the elliptical? Should I see a pulmonologist or a cardiologist?

Thankfully, I had the good sense to shake this off. I realized that the most cogent explanation of what I had observed was that my watch was getting old and was therefore no longer measuring my heart rate accurately. In other words, it wasn’t me. It was my watch. Telling myself this version of events—in which the watch had made a predictive error rather than accurately reflecting the state of my health—quite literally saved me.

The Apple Watch’s VO2 max reading is a type of predictive AI: the watch uses measurements of your heart rate and the intensity of your workout (for example, running pace) to estimate your oxygen capacity, using an algorithm. It’s not measuring that capacity per se. Prediction algorithms can be helpful in that they can extrapolate based on limited data and provide an assessment that otherwise wouldn’t have been possible. For example, predictive AI can provide an assessment of whether a job candidate should be advanced to the next stage in the vetting process. The trouble is it can get that assessment terribly, catastrophically wrong. And because AI algorithms are notoriously opaque, it can be quite hard for a person to determine whether a given assessment has gone wrong—for example, because their Apple Watch is getting old.

What is troubling about these wrong predictions is that they create a reality of their own. When we believe what these systems tell us about ourselves and about others, there is a risk that our subsequent actions might reinforce the assessments, making them more and more real. For example, we might not hire a person because the AI predicts they won’t fit a job based on a minimal, and potentially erroneous, evaluation of data. Over time, such decisions accumulate, such that the person’s joblessness becomes a reality that shapes their chances of getting a job going forward.

One way to describe what predictive AI is doing—and what ultimately is the source of our rightful concern about what is doing—is that it acts on abstractions. AI does not make use of all of the possible data that constitute the real world: there is no way it could, because one never has a complete data set of everything that is needed to make an evaluation and, in any event, the quantity of information involved would be impossible to process. Rather, AI forms a condensed picture, a picture that is necessarily incomplete—a projection of reality. Data are lost to create the abstraction and then filled back in based on the AI’s predictions. The result is inevitably some amount of distortion relative to what is real. Only concrete reality, with all the layers of its impossible-to-process detail, is truly free of distortion.

AI can get better. It can be trained to make use of more information. Its algorithms can be improved. But the distortions that come from abstraction can never be fully eliminated. Indeed, predictive AI becomes increasingly dangerous to the extent that we rely upon its assessments at a rate that increases faster than the rate at which its assessments improve. If AI brings us a dystopian future, it is most likely to be one in which we allow it to dictate stories about reality that we then take as correct by default. This is an opinion shared by many AI experts, including Arvind Narayanan and Sayash Kapoor, the authors of AI Snake Oil. Ray-gun-wielding robots may make for a flashier sci-fi thriller, but a future in which robots without ray guns convince us to believe their alternative version of reality is scary enough.

AI is novel, but what it’s doing really isn’t. The process of interpreting some subset of the information available and making a potentially faulty assessment is precisely what the human brain has always done. You’re doing it right now.

In a fascinating lecture about the predictive brain, the cognitive philosopher Andy Clark explains what recent research by psychologists and neuroscientists has taught us about how the brain creates what we perceive as reality. Our brains have a limited capacity for information processing and so must balance using their scarce resources to take input from our sensory organs against using those same resources to make predictions based on that input. If we err on the side of taking in too much information, we get overwhelmed by inessential detail and consequently have trouble making heads or tails of what we are observing. If instead we err on the side of making judgments based on too little sensory input, we hallucinate—that is, we can believe that we have observed something that wasn’t actually there.

Because our mental bandwidth is limited, we are doomed to mispredict, one way or another. We walk a tightrope, as it were, between two alternate forms of error.

The truly fascinating point that Professor Clark makes based on the research is this: there is no right answer as to how we strike this balance. There is not a right way to operate a predictive brain. There are only pros and cons with respect to how far we decide to go in creating a reality based on our interpretive abstractions.

Recognizing that what we perceive as real is merely a function of what our imperfect, tightrope-walking brains are reporting back to us, we have no choice but to ask ourselves: is there really any such thing as reality? The idea that there might not be can seem scary, especially at first. If indeed it’s the case, what do we do?

It’s tempting to think that what we should do is seek comfort by having someone else tell us what reality is. As children we rely on our parents to do this. As we grow older, many of us turn to authority figures—doctors, clergy, politicians, podcasters—or to others in our lives—spouses or friends—to take that role over from our parents. By this means, we can if we wish avoid for our entire lives the terrifying prospect of inventing our own reality.

The alternative is to embrace inventing our own reality. While doing so can certainly have hazards, it can also be empowering. It can be a tool for healing, growth, and the optimal experience of life.

What we’re talking about here may not sound like economics, but it is. Behavioral economists—along with their counterparts in the fields of psychology and finance—have studied and written about how our reality is shaped by our thoughts and actions. In a 2011 article1 in the Quarterly Journal of Economics, Roland Bénabou and Jean Tirole propose a model of identity in which the actions a person takes now lead them to develop future beliefs about their identity. To see the intuition of this, consider how when we contemplate a morally questionable action we may think, “But what kind of a person would that make me?” The actions we choose quite literally shift the reality of who we are.

In his work on narratives,2 Robert Shiller explores how a popular retelling of a major economic event, such as the Great Depression, can create a reality that influences what happens in the economy going forward. People can struggle with momentous events. We need a story that we can tell ourselves which makes sense of these and which guides our actions in the new world that follows in their wake. Shiller argues that compelling economic narratives spread through a process akin to contagion, passing from one person to another. As a narrative spreads, the economic actions taken by individuals are transformed, as those individuals conform to the reality of the narrative. When people change their actions en masse, global macroeconomic impacts can result.

What can we make of this? It is fascinating to consider that how we think about our actions and the stories we tell ourselves about what happens in our lives can affect what actually is true. Indeed, we have the potential, through our thoughts and narratives, to gain control over our lives rather than ceding it to other people, the media, and AI. We can tell our own story and in so doing change our world and the world around us.

It is this idea that is behind initiatives that seek to achieve broad societal impact through the power of individually-relayed narratives, such as The OpEd Project. Meanwhile, the power of narratives in our personal lives is reflected in numerous story lines from the movies and TV.

One such story is told by the 2002 movie Punch-Drunk Love. Its protagonist Barry Egan (played by Adam Sandler) lives a life that seems poised on the brink of spinning out of control. Random events—his witnessing a distressing car accident and discovering an abandoned harmonium—combine with his own erratic decisions to create an unstable situation in which Barry appears most likely to be headed for madness and embarrassing ruin. And yet, there are amazing possibilities latent in these events. Barry can sense this.

Narrative is a major player in this drama. Barry’s sisters paint him as ineffectual and as a sexual misfit. And for a while we find ourselves wondering whether their narrative might be the correct one. In the end, however, Barry invents a narrative for his life in which he rejects madness, anger and insecurity and embraces love. How that narrative becomes reality is the arc the movie traces—the narrative of the movie itself.

You too can decide. The story is yours to tell, if you dare.