Your Face Tomorrow The puzzle of AI facial recognition

[soso]

Your Face Tomorrow

The puzzle of AI facial recognition

by Michael W. Clune

Listen to an audio version of this article.

https://harpers.org/archive/2025/08/your-face-tomorrow-michael-clune-ai-facial-recognition/?position=3&category=fascinating_stories&scheduled_corpus_item_id=8bef7904-c1df-41be-ad7a-fdc756ac18f2&url=https%3A%2F%2Fharpers.org%2Farchive%2F2025%2F08%2Fyour-face-tomorrow-michael-clune-ai-facial-recognition%2F

I first realized there was a new use for my face when I got my passport renewed in November 2023. I went to the local CVS to have my photo taken. The harried woman behind the counter groaned and led me to the corner of the store where they took the pictures.

She told me to stand in front of a white screen. It was the only usable surface of the store that was not covered by words about consumer medical products, images of consumer medical products, or the bright colors associated with consumer medical products. The screen was the blank white of a medical crisis.

She told me to take off my glasses. I did. It was blurry, but I could just make out the electronic camera’s tiny lens embedded in the black device before me.

I waited.

“Stop that,” she said.

“Stop what?”

“The . . . that thing you’re doing with your face.”

“My face?”

“Smiling,” she said irritably. “Stop smiling. You’re not supposed to smile.”

You’re not supposed to smile? I hadn’t remembered that being the rule the last time I got my passport renewed. Had something changed?

I relaxed the muscles of my face. It felt strange to be standing there in the middle of the store with no expression‎ at all. What they want with my face has nothing to do with my expression‎, I thought. This isn’t about what I can do with my face. This is about what they can do with it.

A word from something I’d read recently swam into my consciousness. Faceprint: “A digital scan or photograph of a human face, used for identifying individuals from the unique characteristics of facial structure.”

“Okay,” she said.

When I put my glasses back on, I looked into the reflective glass of the camera. It looked like the black, expression‎less eye of an insect. What does an ant see when it looks at your face? I didn’t know. I thought about ants because a bug’s perspective seemed like the most alien thing I could imagine.

I didn’t understand then that the machine behind the eye of that camera is much more alien than an ant. Scientists know a lot of things about insect optical processing. But no one knows what artificial intelligence sees when it looks at a picture of your face.

The large language models that programmers train to identify faces are black boxes. Even the engineers don’t know how or in what form your face appears to the system. All they know is that AI likes your face to be brightly lit. And that it prefers for you not to smile.

This initial brush with an alien way of seeing my face made me attend a little more consciously to the normal, human way of using my face. A couple weeks later, I went to my department’s annual holiday party, at the home of the chair. As my wife and I approached the front door, my face began to change.

Up until that moment, driving over to the chair’s house, I hadn’t been paying much attention to it. Lauren would say something funny and I’d laugh. I squinted a little when we turned off the brightly lit road.

My face was like a pool of water—someone would toss in a pebble and ripples would spread. Or it would hold the faint impression of a blown leaf, floating on the surface. Natural impressions, organic expressions.

But standing there, waiting for the department chair’s door to open, I put a very specific smile on my face. I did it manually, so to speak. I had a pretty good idea of what it looked like, from mirrors and photographs of myself. Happy. Open. Excited—not too excited. Definitely not too excited.

I briefly wished that all department chairs’ houses had cameras set in their walls, so I could compare my current smile-at-the-holiday-party-door with the smile I had put on when I was an associate professor, and before that an assistant professor, and before that a graduate student. I speculated that a skilled anthropologist could derive my precise status in the department from the nature of my smile. The “smileprints,” placed next to one another, would describe an arc—from the smile of someone who wanted to please to the smile of someone with a gracious willingness to be pleased.

As I moved through the party, my facial expressions modulated. Polite interest as someone described the repairs being done to their home. Quiet smile of acknowledgment at a perfunctory joke. Deferential and concerned attentiveness to an elderly retired faculty member’s description of a medical procedure. Furrowed brow of performed thought after someone asked my opinion of a recent essay.

And then, just once, just as my face began to tire from this incessant demand to express, I emitted an aggressive, open shout of laughter at my own joke, followed by a gleaming, tigerish grin around the little circle of colleagues whose own faces described an arc of smiles ranging from defensive to liberated.

Liberated? Liberated from the pressure—the constant pressure that accumulates on one’s face in social situations, a pressure felt in one’s facial muscles. There are approximately forty-three of them: some ten thousand possible configurations. Each configuration has a meaning, each one is judged and eval‎uated and acclaimed or condemned or let pass by the ten thousand possible configurations of the face of the person watching you, who is also being watched by you, and by others, one of whom may be artificial.

I knew it was time to leave the party when my face had frozen into a rictus grin. This condition exposed me. It’s a symptom of terminal facial overload, a sign that my forty-three muscles were fighting against relaxation.

The people to whom I was listening could tell that my face was trembling on the verge of a total absence of expression‎. I could sense this not in their mouths, but in the muscles around their eyes. Such a face would be like a giant unhewn boulder dropped into the middle of the chair’s tastefully decorated living room. People will do just about anything to avoid witnessing such a face in a human social situation.

This is the face AI likes best. Sometimes I think I might prefer it, too.

Like many other forms of AI, facial-recognition software received a shot in the arm from the maturation of artificial neural networks such as the large language models powering systems like ChatGPT. The process used by most current methods contains several steps. First, the software translates a photo or video of your face into a set of measurements—the distance between your eyes, between your nose and your lips, and so on. Then this “faceprint” is fed into an artificial neural network. This system uses statistical methods to match your faceprint to others in its database. Programmers “train” the system by rewarding it for correct matches and penalizing it for misses. The model generates results in the same way that LLMs produce language—by processing a large database in order to obtain statistical relationships among entities. Eventually, and with a sufficiently large database, the system can reliably match one image of a face with another. Given adequate lighting and a relatively clear picture, software can match your passport photo with your appearance in a party photo on a friend’s Facebook page, for example, 99 percent of the time.

No one knows exactly how the system obtains its matches. There is an entire field in AI known as “mechanistic interpretability” that is attempting to understand how LLMs move from a given input (your passport photo) to a given output (identifying your face in the Facebook post). The existence of this field—and its limited success thus far, despite the significant resources devoted to it—is one index of the alien quality of the “thinking” that goes on in artificial “minds.”

The opacity of these systems has created controversy, with various advocacy groups arguing that citizens possess a “right to explanation” in cases wherein a black-box process leads to an adverse outcome. If a bank denies your loan application using an algorithm, or if you discover at the airport that you’ve been placed on a terrorist watch list by an AI system, you should be able to find out why. The European Union, as part of the General Data Protection Regulation of 2016, provides such a right to explanation, though no similar legislation exists yet in the United States.

Even with rapid improvements in the technology, the danger of being wrongly identified as a criminal remains. A 2024 New York Times story identified three individuals who had been subject to such mistaken identification by the Detroit Police Department, which has used facial-recognition technology since 2017. Various factors contribute to mistakes of this kind. For example, while there’s been progress in obtaining correct matches with blurry or shadowed or smiling faces, systems can still struggle with such images; thus, perhaps, the CVS worker’s demand that I not smile. But the primary constraint on effectiveness is the size of the database. You need tens of thousands, and preferably millions, of images of faces to train the system to optimal precision.

If occasional inaccuracy is one reason to worry about AI-powered facial recognition, the larger concern is its effectiveness. Decades of unease over the rise of electronic surveillance have primed us to freak out. The nonspecialist writing on the subject operates in a genre one might call paranoid realism; the recent bestseller Your Face Belongs to Us, by the journalist Kashmir Hill, for example, tells the story of the company Clearview AI and dwells exhaustively on the dystopian implications of the technology.

Its possible use cases are certainly freak-out-worthy. Consider: A stalker with access to the software could take a picture of you and then find out where you live, where you work, who your friends are, and where you get your groceries. Surveillance cameras mounted in public streets could record you walking into a strip club or a pot dispensary, and on this basis, a credit agency or employer could deny you a job or a loan. And the government might deploy the technology to discover everything anyone has ever done, and punish them for it.

On the other hand, Hill describes the enthusiasm for the technology among law-enforcement agencies, which point to its efficacy in helping them track down wanted criminals. Cops can use facial recognition to identify child pornographers and abusers from images circulating on the dark web. They believe, plausibly, that the use of these systems could prevent terrorist attacks. But in the context of Hill’s well-researched book, the theoretical positive uses seem overwhelmed by a swirl of anecdotes, speculations, and predictions that suggest this technology is on the verge of plunging us into a more efficient version of Orwell’s Oceania. Certainly that was my experience reading the book.

When I try to examine the threat of facial-recognition software dispassionately, I see the problem boiling down to two basic questions. The first is: Do we trust the government agencies that have access to such systems? If we do, then the benefits of preventing terrorism or child abuse might well outweigh the potential abuses and inaccuracies. If we don’t, then it seems our first task should be to bring these agencies under democratic accountability. After all, if the government is out to get us, it has enough tools to do so even without AI.

The second question is: How much do we value our privacy? There’s been renewed discussion in recent years of the putative “right to privacy” that, in a celebrated 1890 article, Samuel Warren and Louis Brandeis argued was derivable from existing U.S. law. But if it exists—and this remains debatable—such a right comes, even in its original formulation, with a host of exceptions. A right to privacy can in many cases come into conflict with the better-established First Amendment right to free expression‎. Most of the photos held in vast corporate databases (think of your Facebook photos) or government databases (think of your passport photo) were gathered with the explicit or implicit permission of the subjects. Experts have long been familiar with the “privacy paradox,” whereby we say we value our privacy while uploading our photos to public forums, opting into requests to monitor and share our online behavior, and so forth.

In short, when I think carefully about the threat posed by facial recognition, my instinctive, paranoid response gives way to more ambiguous reflections. Maybe—like everything else—this thing will be good in some contexts and for some uses, and bad in some others. Perhaps this isn’t a black-and-white case, but one that involves trade-offs between different values, like privacy and safety.

And now I begin to wonder about that paranoid response, the feeling of horror that first came over me when reading Hill’s book, and subsequently as I searched the web for similar stories about facial recognition. What is the source of this paranoia? Perhaps, I began to suspect, my instinctive urge to reject and ban this software is something like a defense mechanism.

A defense against what? When we consider the way AI makes use of faces, we inevitably contrast it with the way humans make use of faces. The flip side of the paranoid revulsion at AI is an idealized, even romantic sense of the comforting familiarity of human face-to-face interactions. As with so many situations in which we confront new technology, our tendency is to project our fears onto the new thing and to cling to the old, natural way.

But sometimes the old, natural way is the problem. Professors like myself hate ChatGPT and similar platforms because our students turn in artificially generated, robotic papers. But if we ordinarily gave vapid, shallow papers the D’s or F’s they deserved, this problem wouldn’t exist. The fact that such papers routinely get A’s or B’s shows that we have come to expect and to train humans to write robotic papers. Similarly, when I worry I can’t distinguish a colleague’s genuine sentiments from the vaporous generalities Gmail’s AI suggests, what am I really worrying about? Is it that the machine is so good? Or that my interactions with my colleague are so empty?

Once we step back from the paranoid reaction, the problem presented by AI facial recognition assumes different contours. In posing anew the question of facial control, the technology provides us with an opportunity to think about how such control works in both its artificial and natural forms.

If we’re serious about privacy, we should examine the problem seriously—not neglecting the ordinary, traditional, everyday, nontechnical cases in which our privacy is at stake. In her book, Hill expresses one of my favorite arguments in favor of privacy when she writes, “Anonymity provides powerful protection for those who don’t conform to the status quo.” I am among those who believe that conformity is the enemy of the creative, social, and conceptual breakthroughs that enrich and transform human life. When my face is known, I shed the protective invisibility of anonymity. The protean, multiform energies within me become measurable, locatable, predictable, controllable. My face is the hole through which the status quo enters me, disciplines me. It has always been this way.

But now I have two faces. Two doors that swing open to two different forms of control. The first door is my face on my passport. The second door is my face at the department holiday party.

Let’s look a little closer at what’s behind door number one.

In early November 2024, I visited Yu Yin’s lab at Case Western Reserve University, where I taught at the time. In a modern research university, people rarely have a clear idea of what their colleagues in other departments are doing. I learned of Yin’s work on facial recognition during a reception for the board of trustees, where I found myself speaking with the head of her department. This intelligent, engaging man worked on a different branch of AI. Over the course of our conversation, I came to think of him as a character in an eighteenth-century novel, the Optimist. He all but suggested this name himself.

“I am an entrepreneur,” he told me at once. “I have to be an optimist.”

Upon learning that I was thinking of writing an article about AI, he expressed unalloyed enthusiasm for the revolution. I decided to test his optimism.

“Some people,” I said tentatively, “think that facial recognition poses a threat to privacy.”

He waved his hand dismissively.

“No one really cares about privacy,” he said, citing research on the privacy paradox.

I decided to try something a little bolder, remembering a theme of the AI conference in New York where I’d been invited to give a keynote lecture that October.

“Some people,” I suggested, “think that the development of AI will soon hit a wall. There was a recent study in Nature showing once LLMs begin to be trained on AI-generated text, they start spewing nonsense.”

He smiled.

“These problems are fixable. Do you understand the speed of progress in AI research? Let me give you an example. On Monday, my graduate students and I post a paper online about a problem. Thousands of people all over the world read it. They work on the problem. On Thursday they publish a new paper citing our paper.”

By this point, an elderly trustee had joined us.

“Everything is getting better all the time,” said the Optimist.

The trustee nodded in approval at these confident pronouncements. I wondered if my interlocutor’s optimism had any limit.

At the end of our little exchange, he told me of Yin’s work on facial recognition and suggested I contact her. Thus it was that I found myself in her lab in the School of Engineering on a cold, clear November morning. She’d promised she would make a faceprint of me and that I would be able to see the whole process. This excited me, because getting direct experience of the technology had proved rather difficult. I began my visit by mentioning this.

“Yes,” she agreed. “This software is not available for consumers. My personal view is that applications that identify a person’s face should not be licensed for consumer use.”

“Because of the danger of stalkers using it, that kind of thing?”

She nodded. She told me she had been born in China and had lived there through her undergraduate years, moving to the United States for graduate training.

Part of the work being done in her lab, she explained, involved taking photographs of individuals and then animating them. With a recording of the individual’s voice, one could use AI vocal programs to make their face say various things, complete with the requisite expressions and mouth movements.

One possible application might be for something like Zoom calls, where you could deploy an avatar of yourself, and speak into the microphone with your AI avatar speaking your words perfectly, your face looking absolutely natural, no one the wiser. They weren’t yet able to get the animation to work in real time, but they were making progress.

I envisioned someone getting a picture of my face, and then using the tech to make a Zoom call to my elderly mother, asking her for money. Yin smiled, and acknowledged that this technology might be abused. But it could also have benefits—for gaming or film, for example.

At this point, one of her graduate students arrived, and Yin informed me that it was time for my faceprint. Suddenly I became a little nervous. Somehow I had imagined a vaguely medical scenario. I would be led, I imagined, from this small white-paneled room, full of more or less recognizable computing equipment, into a different space. Perhaps I’d be invited to lie down, on the kind of bed they have in doctors’ offices. And perhaps a special camera would be lowered over my face, and I’d be told not to move while an enormous glass lens dilated and constricted above me.

Standing there in Yin’s office, I realized that what I’d been imagining was a larger and more intrusive version of the X-ray machine dentists use, with a camera pointed not toward some part of my jaw, but toward my whole face—a camera that was correspondingly larger. This strange anxiety-fantasy vanished instantly as Yin’s graduate student gestured toward a table holding an open laptop. I looked at it. It showed a video feed of myself and the grad student staring into the laptop’s webcam.

“What is this?” I asked.

“This is it,” he said.

“What?”

“It is capturing your face.”

“When? Now?”

“Yes,” he said. He looked at Yin. They both smiled.

I examined the screen more carefully. There was my face. But it was different. A blue highlight appeared over my left eye, like an eyebrow scrawled in marker on the screen. A green highlight appeared over my right eye. And a wavering white stripe, like something my five-year-old daughter might draw—hesitant, wobbly—circled the shape of my face. Some numbers flickered on the left side of the screen.

“This is it,” I repeated.

The effect was vaguely cartoonish. It reminded me of a certain genre of TV ad; I thought dimly of soda commercials, or maybe running shoes, ads in which bright pastel colors scribble over dancing consumers.

But, remembering what I’d read about the technology, I realized that this cartoon effect emerged from something like the polar opposite of a marketer’s mind. Those pastel colors were the expression‎ of a deeply alien process, a truly inhuman perspective on the human face.

What I was looking at as I stared at my concentrating face on the screen, I realized, was a trace of the way the machine saw me. The cartoon highlights, the wavering white outline, the numbers—these were the visible sign of the machine’s digestion of my human features into a code, a set of coordinates, a faceprint.

“What are those numbers?” I asked.

Two slowly ascending digits flickered in the upper left corner of the screen, like a reverse countdown: 81. 83. 86.

“That is the match,” he said.

He explained that the numbers showed the degree to which the current image of my face matched other instances of my face in the database.

I watched the number climb into the nineties. The machine was learning to recognize me.

The scene in Yin’s office represents the third time in my life that my face had been captured for a use beyond its standard social function. We can distinguish between these two ways of using my face by distinguishing between an interest in expression‎ and an interest in non-expression‎.

The regime of expression‎ includes photographic captures of my face ranging from family or elementary school class photos to television appearances. In each of these cases, the “users” of my face desire expression‎. They want to see my face emitting a true or false (this distinction doesn’t matter in most cases) rendering of a prosocial inner state. My family wants to see me smiling happily in the school photo. The TV interviewer wants to see me engaged in serious thought or conversation.

But in Yin’s office, during my third encounter with the non-expressive regime, none of this mattered. The point here wasn’t for me to perform some inner state, but simply for the software to identify me. My second encounter with this regime had been at the CVS when my passport photo was taken. It is probably true that there have been other times—when my previous passport photos were taken, or when I posed for my driver’s-license photo—that were also instances of interest in my non-expressive face. But if so, I was confused about this at the time, and treated each photographing as an expressive scenario, smiling broadly.

But the first time I encountered this regime, I didn’t smile, and I was not confused. This was when my mug shot was taken in a Chicago jail in 2002, after I had been arrested for felony possession of narcotics.

“Stand over there,” a cop commanded. “And look over here. Give me your glasses.”

“My glasses?”

He pulled me toward him and snatched my glasses off my face. He then half closed them, and brought the two stems—which made a kind of miniature vise—to bear on my throat. He pressed, hard.

“You see?” he said. “You see this is a weapon?”

Then he released me, glasses-less, and I stumbled backward. I heard the click of the camera, and then he took me to my cell.

Whatever expression‎ that camera captured (and I’ve never seen the mug shot; when my felony was expunged a couple years later, they sent me my fingerprints in the mail, but not the mug shot), whatever expression‎ I wore—dazed, blinking sightless at broad planes of gray and white jail-color—was not intentional. Perhaps it was in a sense the most authentic image of my face ever captured. At any rate, the justice system was certainly not interested in making me smile. This photo was for identification purposes only.

What links these three different scenes of facial capture—in Yin’s lab, in the CVS, in the West Side of Chicago jail? These different systems—the state, local, and federal police agencies; the U.S. State Department; the actual and possible end users of facial-recognition technology—are not oriented toward what is inside me. They’re not trying to capture my thoughts or feelings or words. They are oriented toward my actions.

Each system is designed to match a given action—a drug offense, a violation of customs or immigration law, an appearance at a party that someone is recording for an Instagram video—to a given name. My name.

This is the truth embedded in the term “faceprint.” It’s like a thumbprint: a pattern unique to one individual, the trace someone might leave in every space he traverses. Thinking of an image of your face as somehow analogous to your thumbprint leads to serious mental vertigo. Take a look at your thumb. Now look at your face in the mirror. Imagine your neck terminating in a giant thumb.

To understand how one’s face can become as neutral, as objective, as expression‎less as one’s thumb, is to grasp the key difference between the regime of facial recognition and the regime of everyday expressive facial control.

The example of my mug shot perhaps gives an unduly negative view of the possibilities of facial recognition. The effort to view a person from a distance, to observe their actions from an external perspective, doesn’t have to be creepy. One might even think that this effort is essential to being a good person.

Adam Smith thought so. In his Theory of Moral Sentiments, he wrote that we cannot judge whether our actions are morally good or bad from within. The only way to make such a judgment is to “remove ourselves, as it were, from our own natural station,” and to “endeavour to examine our own conduct as we imagine any other fair and impartial spectator would examine it.” Smith continues:

I divide myself, as it were, into two persons; and that I, the examiner and judge, represent a different character from that other I, the person whose conduct is examined into and judged of. The first is the spectator, whose sentiments with regard to my own conduct I endeavor to enter into, by placing myself in his situation, and by considering how it would appear to me, when seen from that particular point of view.

To be a good person, one has to be able to judge the moral status of one’s actions. And to do this requires alienation. I must disembed myself from what I think and feel, pry my mind out of my face, and observe myself as a simple name and body performing certain actions.

I do this all the time without thinking. When my five-year-old runs into my office as I’m playing a computer game and wants me to read her a book, I feel irritated, set-upon. Probably my facial expression‎ conveys these feelings.

But I also see the situation from a different perspective. I imagine a perspective outside the situation, watching me. And I imagine that this observing entity would be pleased if I stopped what I’m doing. It thinks that it would be good if I stood up, took the book the child proffers, and spent ten minutes reading it to her.

I want this observer to think well of me. So I perform the action that will produce this positive judgment. Afterward, my daughter leaves my office with a happy smile. And I am also smiling as I return to my game, feeling like a good father.

Who was this alien observer, whose gaze made me into a (slightly) better person, whose gaze (slightly) reduced my incorrigible self-centeredness? In one sense, of course, it was me. But it was a version of me identifying for the moment with an outside point of view that registers and judges my actions. The outsideness is crucial. This perspective doesn’t know or care what I’m thinking or feeling. It isn’t compelled by—doesn’t even recognize—the expression‎ of irritation on my face. It doesn’t attend to my thoughts about how I’m so close to beating this one level I’ve been trying to conquer for a week. It trains its cold, alien gaze only on my actions.

The kind of surveillance Smith describes is a form of control. Such control isn’t intrinsically tyrannical or oppressive. Its value depends on the aims it serves. In the example with my daughter, the aim is moral goodness. Smith argues, to my mind compellingly, that there isn’t an easy way to be a good person without this kind of surveillance by another, even if that other is, in the end, only you. Of course, it’s easy—maybe even too easy—to imagine scenarios in which the observer’s aim is a bad one. Nineteen Eighty-Four. China. Google.

But is it possible, I wonder, to imagine a world in which social control operates primarily, and optimistically, through a non-expressive model? Imagine a benign government, watching you through street-mounted surveillance cameras, tracking you through your friend’s Facebook posts. Maybe it’s not going to do anything in particular with these images. Perhaps it might intervene if it looks like you’re about to harm someone. But in general it just watches. It wants you to be a good person. The imaginary regime works, in fact, like Smith’s impartial observer: the constant alien and alienated eye that alone makes it possible for a person to truly do good.

This thought experiment will seem fanciful to some, terrifying to others. But I want to use it as a contrast with another regime, a regime with nothing fanciful or imaginary about it. This regime may have temporarily relaxed its hold on your face as you read this. But soon—very soon—it will be manipulating those forty-three muscles again.

The advent of facial-recognition technology arouses unsettling feelings in part because—fancifully or terrifyingly—it opens the prospect of an alternative to the way we’ve always used faces. The emergence of an alternative gives a new perspective on the old thing, and not always a flattering one. Think of the invention of indoor plumbing.

So let’s take a closer look at the old thing.

Return to the scene with my daughter. I’m sitting in my office, playing a computer game, when she comes in holding a book, demanding I read it to her. Now, instead of my actions being tracked by Adam Smith’s impartial observer, let’s imagine a real, live person sitting there watching me. Perhaps it’s a relative. Perhaps a neighbor who’s dropped by. Perhaps a friend.

Like the alien, outside gaze of Smith’s observer (or AI facial-recognition software), the mere presence of this person will exert control over me. But the control wielded by the human is deeper, more intrusive. This real person is not just interested in my actions. They don’t see my face as in any way like a thumbprint, a simple means of identifying my actions. They watch my expressions.

They see the child run up to me. If this third person wasn’t there, my face would instinctively reflect my actual emotional state of mild irritation. But with this person watching, I can’t afford to betray my feelings by this expression‎. The kid wants Daddy to read her a book. What kind of father could be irritated by so charming and salutary a request? So I turn to the child with a smile. Of course I’m happy to drop what I’m doing and read her a story.

That smile is a razor, cutting backward into my brain. It rips my irritation to shreds. It might even destroy the memory of it. My neck and shoulder muscles tighten with the internal effort of eradicating and erasing my first, natural feeling. The face I turn to my child with is the face of a loving, smiling parent, unbothered—excited even—to be interrupted.

In both scenarios—the one in which I adopt the perspective of the impartial observer, and the one in which I am watched by an actual human observer—being watched changes my behavior for the better. But consider the difference.

In the first example, as I realize the action I should take, my frown slowly turns upside down. I expel my irritation with a sigh, as I sit the child on my lap and start to read to her. By the end, both she and I are smiling. I have become the happy, good father. But I’ve arrived at this morally desirable end point through a natural, relatively slow process in which the demands of my own self-centered feelings and inclinations were challenged, and then defeated, by a sense of how an impartial observer would judge me.

In the second example, I become the happy, good father instantaneously. But this good man is an artifact. He is artificial. He has been created by the human sitting in the chair across the room, who, by triggering the manipulation of my facial muscles, nullifies the expression‎ of my natural feeling. In changing the way I appear, this person also causes a deeper change.

Hegel says that “the self perceives itself at the same time that it is perceived by others. . . . Self-consciousness exists . . . by the fact that it exists for another self-consciousness.” I become myself by identifying with the object you see. And what you—you other human beings—mainly see is my face. As the psychologist Silvan Tomkins writes, “the self lives in the face”:

Both transmission and reception of communicated information take place at the face. The mouth talks, the eyes perceive; and the movements of the facial musculature are uniquely related to one’s experienced affects and to the affects transmitted to others.

All of us, pretty much all the time, want others to see us as a good object. Since our sense of ourself depends so highly on the attitudes of others, we instantly, preemptively, and constantly work the forty-three muscles of our face to produce the expected response—the response that will please or impress others. This dynamic is so pervasive and unremitting that it can be hard to bring it to consciousness.

An offhand comment that William S. Burroughs makes in Naked Lunch illuminates the way others shape us from an unexpected angle. A character in the novel says you can learn more about someone by talking to them than by listening to them. How could this be? Because the part of our mind that learns about someone’s attitudes by listening to their words is inferior in its cognitive power to the part of our mind that grasps their attitude from a thousand tiny cues—their clothing, their posture, the precise modulations of their own facial muscles—and then responds to this information by working our own tone of voice, word choice, and facial expressions to produce the impression they expect.

When I enter a room and find my wife talking on the phone, I can almost always identify the person with whom she is speaking, even though I can’t hear anything the other person says. I simply listen to how my wife talks and I know. She has a special tone of voice, and special facial expressions, when her auditor is her mother. A different facial and vocal suite when she’s speaking to her friend John. Still another when she’s speaking to my daughter’s teacher at school.

I’m the same way. Am I conscious that I’m a slightly different person when I’m speaking to my friend Dave versus when I’m speaking to my friend Jason? Not normally. But now, when I’m thinking about this problem, when I’m listening to myself speak, I hear the difference.

We’ve examined the paranoia of the non-expressive regime of facial recognition. But there is also a paranoia of the expressive regime. “We can consider ourselves as ‘slaves,’ ” writes Sartre, “insofar as we appear to the Other.” A gaze tuned only to my actions—a gaze that sees my face as a kind of thumbprint—exerts control over what I do. But it leaves the space within me free. It doesn’t control what I think or how I feel. The machine recognizes facial structure but not facial expression‎. Such a gaze allows me the possibility of naturally coming to identify with the actions it encourages or pressures me to take—as when my feelings about reading to my child change from an initial irritation to eventual joy.

Even in a surveillance regime controlled by users who don’t want me to be a good person, but merely compliant—a tool of their own power—my interior remains relatively free. I can do one thing while thinking and feeling and perhaps planning another.

Of course, it’s also possible to do this in an expressive regime. I can smile while hating you. But the pressure is greater. People are adept at detecting the little blips in facial manipulation that indicates someone’s faking it. Over time, it becomes easier to simply feel what your face expresses.

“The self perceives itself at the same time that it is perceived by others.” Who am I? A real person, watched by artificial eyes? Or an artificial person, watched by real eyes?

I close my own eyes and see my face on the screen in Yin’s office. The pastel colors smudge like cartoon eyebrows. My expression‎—baffled, then interested.

I feel Yin’s eyes on me. I smile.

The digits on the left side of the screen ascend.

From the

August 2025 issue

Download PDF

From the Archive

Timeless stories from our 175-year archive handpicked to speak to the news of the day.

Email address

Michael W. Clune

is a contributing editor of Harper’s Magazine. His novel Pan was published in July by Penguin Press.