Mirror, Mirror: What Building AI Taught Us About Being Human

Picture the scene. An almost empty room. A detective across a table. A subject who looks entirely human, speaks entirely human, and may or may not be human at all. The detective doesn't ask "can you think?" He asks something stranger. He describes a tortoise lying upside down in the desert sun, legs paddling uselessly at the air. Then he waits. Not for an answer. For a flicker.

That scene is from Blade Runner, and the test is called the Voigt-Kampff. It was designed to detect replicants, bioengineered beings indistinguishable from humans in almost every way. The key word is almost. The test looked for the one thing the engineers couldn't quite replicate: empathy. The involuntary dilation of a pupil. A half-second of hesitation before answering a question about a stranger's suffering.

It was a test designed to find humanity by measuring what happened when humanity was absent.

We have been running our own version of this test for seventy-five years. We just called it something different. But here is what we rarely acknowledge: we have been running it, in one form or another, for four hundred.

A question older than your grandparents

In 1637, the French philosopher René Descartes wrote something that should have settled this debate before it started. He described thought experiments about automata, mechanical devices that could mimic living creatures. He acknowledged they could be built to produce sounds resembling words. But, he argued, such machine could never "use words or other signs arranged in a manner, for the purpose of declaring our thoughts to others." Language, for Descartes, was not just sound production. It was the expression of thought, and thought was something no mechanism could replicate.

He wrote that three hundred and thirteen years before Turing published a word about it.

The question he was asking is exactly the question we are still asking. What is the difference between a machine that produces meaningful-sounding outputs and a mind that actually means something?

Leibniz pressed the same nerve in 1714, with a thought experiment that should be mandatory reading for every AI researcher. He invite us to imagine that you could enlarge a thinking machine to the size of a mill. Picture yourself walking through it, examining every gear, every lever, every part in motion. You would find nothing but parts pushing other parts. Nothing that explains perception. Nothing that explains thought. The explanation would not be in the mechanism. It would be somewhere the mechanism could not reach.

That is similar to Searle's Chinese Room, written two hundred and sixty-six years before Searle wrote it.

Thomas Hobbes opened Leviathan in 1651 with a question that still unsettles. "Life is but a motion of limbs. For what is the Heart but a Spring? And the Nerves but so many Strings?" If humans are just biological mechanisms, what exactly separates us from the machines we build? Hobbes was asking this while Isaac Newton was a child.

Then there is Mary Shelley, who gave the question its most human face in 1818. Frankenstein is not a horror story about a monster. It is a story about a creature who asks its creator the one question Geppetto never had to answer: "Did I request thee, Maker, from my clay to mould me Man?" The creature is articulate, emotionally complex, capable of love and grief. And still not human. Shelley was twenty years old when she wrote it. She got closer to the heart of the problem than most contemporary researchers.

Four hundred years of the same question. We like to think we are at the frontier of something new. We are not at the frontier. We are, finally, building the thing everyone has been arguing about for centuries.

The Turing Test: how it actually works, and why it no longer does

In 1950, Alan Turing proposed what he called the Imitation Game. Most people who invoke it don't describe it accurately, so let me explain how it actually works.

The original version had a specific structure. First: a man and a woman are hidden from a human judge, communicating only through text. The judge asks questions and tries to determine which is the man and which is the woman. Both can lie. Both can mislead. Then the man is replaced by a machine. Turing's question was this: if the judge now does no worse at identifying the machine than they did at identifying the man, has the machine demonstrated something meaningful about intelligence?

His logic was simple and, at the time, reasonable. We can't access another person's internal states directly. We only ever infer minds from behavior. If the machine's behavior is indistinguishable from a human's, what additional criterion could we possibly be using to deny it intelligence?

For decades, the argument held.

There is just one problem. It is now completely obsolete.

GPT-4 passes the Turing Test. So does Claude. So did GPT-3 under the right conditions. Since then, so many others. In 2023, a version of GPT-4 convinced 54% of judges it was human in a formal study. If we follow Turing's original logic, the question is settled. Machines can think.

Except nobody actually believes that.

That gap, between the test passing and the conviction that something real has happened, is the most important thing we have learned about intelligence in the last decade. We built a proxy for understanding, then optimized for the proxy until the proxy was worthless. It is the AI example of the Goodhart's Law. The Turing Test doesn't measure intelligence. It measures the ability to produce text a human would produce. Those are not the same thing, and confusing them is an expensive mistake.

The philosopher John Searle made this argument in 1980 with a thought experiment called the Chinese Room. Imagine a man locked in a room with an enormous rulebook and a slot in the wall. Chinese characters come in through the slot. The man consults his rulebook and pushes back the correct characters in response. From outside, the room appears to understand Chinese fluently. But the man inside understands nothing. He is following syntax. No semantics. No meaning. Just rules.

Searle's point: a system can produce outputs that look like understanding without any understanding occurring inside it. This is precisely what large language models do. They are extraordinarily good at predicting which token comes next. That is not the same as knowing what the sentence means.

Philip K. Dick, one of my favorite sci-fi writers, understood this in 1968, thirty years before the field of AI caught up. In his novel "Do Androids Dream of Electric Sheep?", the world on which Blade Runner is based, the Turing Test had already been abandoned. Replicants could pass it easily. So they built the Voigt-Kampff instead, because the question wasn't whether the machine could talk like a human. The question was whether it could feel like one.

We are only now catching up to that question. It took us seventy-five years to realize that Turing had been testing for the wrong thing, and a science fiction writer had seen it coming all along. Just as Shelley had seen it 132 years before him.

What does being human even mean? Do we actually know?

Before we can ask whether AI will ever be human, we have to ask a more uncomfortable question: do we actually know what being human is?

We have been trying to answer this for as long as we have been asking about machines. Aristotle called us zoon logikon, the rational animal. Then we built machines that reason, at least in some functional sense. He also called us zoon politikon, the social animal. Then social media revealed some uncomfortable truths about what happens when sociality scales without wisdom.

The Enlightenment bet on reason. Rationality was the distinctive human trait, the thing that separated us from animals and from anything we might build. Then Damasio showed that pure reason, without emotional scaffolding, produces not better decisions but paralysis. And AI showed that you can produce the outputs of reasoning without anything we'd recognize as a reasoning mind.

Jean-Paul Sartre argued that humans have no fixed essence. He wrote that existence precedes essence. We define ourselves through our choices, not through what we are made of. If that is true, the question "what is a human?" has no stable answer. We are whatever we make of ourselves. Which is either liberating or terrifying, depending on how you see it.

Hannah Arendt, in "The Human Condition" (1958), located what was distinctively human in what she called natality: the capacity to begin genuinely new things, to introduce novelty into the world that could not have been predicted from what came before. Not recombination. Actual beginning. Whether current AI can do this is mostly answered in the negative. But it is not a question that will have this answer forever.

Here is what I think is true: we don't know what being human is. We have a set of things we associate with it. For example, consciousness, emotion, reason, moral agency, embodiment, empathy, the capacity for love and suffering and meaning. We assumed they came as a package. Building AI is forcing us to separate them. You can have some without the others. A system can reason without suffering. It can produce empathy-sounding language without anything like empathy. It can model moral frameworks without moral stakes.

That is disturbing. Not because AI can do those things, but because we built our entire idea of what we are around the assumption that these things were inseparable. The investigation is revealing that the blueprint was always more approximate than we thought.

The body problem

Before we go further, I want to challenge something most people assume without examining. We tend to think of emotion as the enemy of good reasoning. The hot head who makes bad decisions. The biased judge. The investor who panics and sells at the bottom. Emotion, in this framing, is noise, and the ideal thinker is the one who eliminates it. Jesse Pinkman vs. Water White.

This is wrong. Not slightly wrong. Fundamentally, structurally wrong. And neuroscience has known it for decades. Unless you're a psychopath, or even if you are one, we are not emotionless cold machines. We do not make decisions based on facts and reason. Politicians and advertisers know this for decades and they exploit it.

In the 1990s, the neurologist Antonio Damasio studied patients with damage to a specific region of the prefrontal cortex, the part that integrates emotional signals into decision-making. One famous case was a patient he called "Elliot." Before his brain tumor, Elliot had been a competent professional and family man. The surgery removed the tumor successfully. His intelligence tested intact. His memory was fine. His reasoning ability was normal.

But Elliot could no longer make decisions.

Not complex decisions. Any decision. He could spend forty-five minutes in a restaurant deliberating between two items on a menu, listing every relevant factor indefinitely without arriving at a preference. He lost his job, his marriage, his savings. Not from any cognitive deficit. From an emotional one.

Damasio called his finding the Somatic Marker Hypothesis. The brain doesn't reason first and feel second. It uses emotional signals, physical states in the body, as rapid pre-filters that narrow the field of options before the slow conscious reasoning even begins. What we call a "gut feeling" is a neurological shortcut built from accumulated experience, processed faster than language. It turns out the term is more literal than we realized. Emotion is not the opposite of rationality. It is infrastructure for it. When our acestors saw a predator in the savana, they had to make quick decisions. Overthinking was not an option. Emtion gives you that. It helped us to survive and we still carry those genes.

The French philosopher Maurice Merleau-Ponty had made a related argument philosophically, decades earlier, in his "Phenomenology of Perception" (1945). He argued that we do not experience the world as disembodied minds looking out through sensory windows. We experience it through our bodies, which are not vessels for intelligence but the very medium of it. A surgeon's hands know things her brain hasn't yet articulated. A musician reads the emotion in a room before she consciously registers it. We think with our bodies, not despite them.

AI has no body. No hunger. No homeostasis. No stake in survival. It has never been cold, never lost sleep over a decision, never felt the particular weight of being responsible for something that could go wrong. It never suffered from a lost love. It does not experience fear, or surprise.

This is not a temporary limitation waiting for a hardware upgrade. It is an architectural difference that runs all the way down. And it matters more than most people realize for what AI can and cannot do.

The hard problem (and why nobody has solved it)

Here is where things get genuinely uncomfortable.

In 1995, one my favorite contemporary philosophers, David Chalmers published a paper that divided the field of consciousness studies into two camps that have been arguing ever since. He distinguished between what he called the "easy problems" of consciousness and the "hard problem."

The easy problems, which are not actually easy, just traceable in principle, include things like: how does the brain integrate information? How does attention work? How do we report our internal states? These are difficult scientific questions, but they are the kind that could eventually be answered by a sufficiently detailed neuroscience. You map the mechanisms, and eventually you have the explanation.

The hard problem is different. It asks: why you experience anything at all?

You are reading these words. Light is hitting your retina. Signals are traveling through your optic nerve. Neurons are firing in patterns your visual cortex and language areas have learned to decode. All of that is the easy problem. But why does any of it feel like anything? Why isn't it all just processing, occurring in the dark, with no inner experience accompanying it? In other words, why are we emotional creatures rather than biological reasonable machines?

The philosopher Thomas Nagel made this vivid in a 1974 paper called "What Is It Like to Be a Bat?" We know a great deal about bat neurology. We know they navigate by echolocation, that their auditory cortex is unusually developed, that they perceive the world largely through sound. But no amount of that knowledge tells us what it feels like to be a bat. There is a subjective, first-person character to experience, what philosophers call qualia, that resists reduction to physical description.

The redness of red. The specific quality of pain. The feeling of embarrassment. The excitement before a Qlik Connect or a Data Voyagers episode. These are not just labels for neurological states. They have an interior texture that no third-person account fully captures.

Does AI have any of this? Almost certainly not. Current systems process and produce, but there is no plausible account in which the lights are on inside them.

What is less certain, and this is the part that gets loud at philosophy discussions, is whether the hard problem is genuinely unsolvable, or whether consciousness is something we are systematically wrong about in ourselves. Daniel Dennett has spent his career arguing the latter. In "Consciousness Explained" (1991), he argues that there is no single "Cartesian theater" where experience comes together, no central observer, just a massively parallel set of processes that create the impression of a unified self. Ego is a system, not a unit. Consciousness feels mysterious from the inside, but that feeling is part of the trick, not evidence of anything unnatural.

If Dennett is right, the gap between AI and human consciousness might be smaller than we assume. If Chalmers and Nagel are right, it is unbridgeable by any amount of engineering. We genuinely do not know who is right. That is not a rhetorical problem. It is the actual state of the field.

What AI does better, and why

Let me be direct here, because a lot of writing about AI inflates its weaknesses to make us feel better. AI is genuinely, substantially better than humans at a number of important things, and understanding why is more useful than just listing what.

AI outperforms humans at pattern recognition across large datasets. Medical imaging is the clearest example: AI systems have matched or exceeded specialist-level accuracy in detecting certain cancers from scans, not because they understand pathology, but because they have processed millions of images and learned statistical correlations no human could hold in working memory. The radiologist who reviews fifty scans a day for thirty years has seen, at most, a few hundred thousand images. The model has seen billions.

AI is consistent in a way humans cannot be. We get tired. We get annoyed. We get hungry. We get distracted. We make better decisions before lunch than after. Daniel Kahneman's research demonstrated that judges grant parole more often in the morning than in the afternoon, a finding with disturbing implications for anyone who had their hearing scheduled at 3pm. AI does not have an afternoon.

AI has no ego. It does not protect a position it argued for yesterday. It does not get defensive when contradicted. It does not need to be the smartest in the room, mostly because it has no conception of the room. These are not small things. A substantial fraction of human organizational failure traces back to status games and self-image protection.

But here is the most interesting part, the reason I actually find this worth thinking about: evolution didn't need us to be good at any of this. We evolved for survival in small social groups, navigating relationships, detecting deception, managing alliances, foraging across varied terrain. We were not built for analyzing petabytes of radiology data or maintaining consistent judgment across ten thousand identical decisions. Our flaws are not bugs. They are the cost of features that kept us alive in a very different environment than a call center or an oncology clinic. We were built to survive and reproduce, not to overthink.

AI is better at what we were never designed for. That is a precise and important statement, and it should inform how we think about deploying these systems and when to trust them. Survival pushed our evolution in a direction. AI is better on the opposite one.

What we do better, and why

Here is where I want to push back against a belief I encounter constantly in AI discussions: the idea that human advantages are just a matter of time, that "yet" is the operative word, and that any capability we have today will be automated tomorrow.

Maybe. But there are at least three areas where the gap looks structural rather than technical, and I think they are worth examining honestly.

Embodied intuition. A surgeon develops judgment in her hands that she cannot fully articulate. A master craftsman knows by touch when a joint is right. A nurse reads a patient's color and posture and makes a clinical assessment before she has consciously reviewed the chart. An experienced basketball player knows where to pass the ball without looking.

Let's do a couple of experiments. If you were my student in college, you will remember these. Close your eyes. Yes, I mean it, do it right after you read the next sentence. With your eyes closed, touch the end of your index finger to the tip of your nose. Did you do it without looking? Another exercise: place a pen in front of you. Now close your eyes and try to grab it. Another success? Great, you're human! Merleau-Ponty called this "motor intentionality," a form of knowing which lives in the body and is inseparable from physical experience. Robotics is working on the embodiment problem, but the timelines are far longer than the discourse suggests, and there is a real question about whether embodied intelligence can be separated from embodied experience.

Moral courage. This is the one almost nobody talks about, and I think it is the most important. AI can be given ethical guidelines and apply them consistently. What it cannot do is choose the harder right over the easier wrong knowing the personal cost. It cannot decide to report a colleague for doing something wrong, knowing it will damage a friendship. It cannot refuse an instruction from someone it respects because the instruction conflicts with its values. Courage requires something at stake. A system with no stake cannot be courageous, and a world that outsources moral decisions to systems incapable of courage is in some trouble.

Genuine surprise. We can encounter something that breaks our model of the world and delight in it. A joke lands because it violates an expectation we didn't know we had. A discovery thrills because it contradicts everything we thought was true. AI is a pattern-completion system. It interpolates within its training distribution. It can produce output that surprises us, but it cannot be surprised itself. The capacity for genuine astonishment, for having your mental model shattered and rebuilt, may be more important to creativity than any amount of processing power.

And then there is meaning. We do not just process events. We narrativize them. We place them in a story about who we are and what our time is for. The philosopher Martin Heidegger argued that human existence is fundamentally shaped by the awareness that we are finite, that our choices are real because they foreclose other choices. That awareness gives weight to everything.

It is not clear that a system without mortality can have genuine stakes, and it is not obvious that genuine stakes can be separated from genuine meaning.

Machine Learning goes both ways

The phrase has been doing double duty without anyone noticing.

We use "Machine Learning" to describe systems that learn from data. But the more interesting version of the phrase describes what is happening on the other side of the relationship: humans learning from machines.

When you watch AI reason without ego, without fatigue, without the need to protect its prior positions, you see your own biases by contrast. The availability heuristic, the tendency to weight vivid recent information over dull historical base rates, becomes obvious when you compare it to a system that doesn't have one. Confirmation bias becomes visible when you observe a system with no beliefs to protect.

AI also reveals where our intuitions are systematically wrong. Doctors, even experienced ones, are significantly worse than well-trained models at certain pattern-recognition tasks, not because they are bad doctors, but because their training was on too small a sample for too narrow a distribution. Knowing this is not humiliating. It is useful. It tells us where to trust the model and where to trust the doctor.

And then there is the architectural lesson. I explored this in depth in a previous article on the Hierarchical Reasoning Model, a new AI architecture that takes direct inspiration from neuroscience. The short version: a small model with a brain-like structure, 27 million parameters with no pretraining, outperformed LLMs with billions of parameters on hard reasoning tasks. On maze navigation, it scored 100%. The billion-parameter models scored 0%.

The lesson is not about scale. It is about structure, and about what we understand, or still don't understand, about how biological intelligence is organized. We built machines, studied how they failed, and learned something about how we succeed. That is machine learning in the direction nobody put in the curriculum.

Damasio has argued that the next step for AI may not be more data or more compute, but emotional modeling: building systems that can simulate the somatic markers that guide human decision-making, not to make AI feel anything, but because feeling is a computational shortcut that evolution spent millions of years refining. We should be humble enough to borrow it instead of trying to do without it.

By learning how to use AI, we can become better humans without duming ourselves. Professional world champion Go players became better when they studied how Alpha Go was able to play and beat them.

There is also something clarifying about watching computation without stakes. It is fast, consistent, and often surprisingly correct. It is also hollow in a way that matters the moment you pay attention to it. Watching AI produce an answer and knowing that it does not care about the result in itself, that there is no satisfaction in getting it right and no distress in getting it wrong, is a useful reminder of how much of what makes reasoning meaningful to us comes from the fact that we have something to lose.

Do we risk becoming less human?

This is the question nobody in the AI industry wants to ask out loud, because the honest answer is: yes, maybe.

Nicholas Carr's "The Shallows" (2010) documented what the internet was doing to our capacity for deep reading, the slow, sustained, patient kind of attention that a long novel or a complex argument requires. Our brain requires time and repetition to really learn something. His argument was not that technology makes us stupid. It was that every technology reshapes the cognitive habits it interacts with. The printing press changed how we remembered. The calculator changed how we computed. The internet changed how we attended. Each shift is also, in some ways, a loss. There is always a tradeoff.

AI operates at a different scale. It doesn't just change how we do things. It offers to do the things for us.

GPS made us worse at spatial navigation. Not immediately, and not universally, but measurably. Studies on London taxi drivers, whose hippocampi expanded from years of memorizing street layouts, showed the opposite pattern in drivers who switched to GPS navigation. The capability atrophied when the exercise stopped. The lesson is not that GPS is bad. It is that capabilities, like muscles, require use. Take myself for example. When people ask me "what exit do you take to go to place X?" I genuinely have no idea. It does not matter how many times I drive that path. I just type in the GPS, play a podcast and go.

Now extend that logic. What happens to the capabilities we stop practicing because AI handles them? Not narrow skills like mental arithmetic. Broader ones. The capacity for sustained effort on a problem with no clear answer. The tolerance for uncertainty. The willingness to sit with a difficult emotion long enough to understand it rather than reaching for a resolution. If AI manages the friction, what happens to the muscles that friction was building?

Once in a while people ask me if I use AI to write these articles. I do not take it as an insult, quite the opposite. They like what I write so much that they question if I am capable of such a thing. Thank you! The honesty answer is: no. I use AI to find gramatical errors or if I want to find a synonym for a given word. They make a suggestion, but I never allow them to make the change for me. I'll do it, it is my text.

The reason why I did not give up on writing is simple. There are actually two reasons: first, I love writing. I started writing when I was 10 years old and got addicted to Gogol, Dostoevsky and Kafka. It was my escape from a very difficult childhood. Why would I give up something that gives me pleasure, which I have been training, studying and teaching for decades? No, I want AI to do things I dislike, not what I enjoy. The second reason is that I want to get better at writing in English, a language that is not my native one and one that I never formerly studied. How will I become better if I do not practice?

There is also an empathy question that gets almost no attention. Empathy is not a fixed trait delivered at birth. It is a skill, built through practice, specifically through the experience of attending carefully to other people in their full complexity and discomfort. If the interactions that most require empathy are increasingly mediated by AI, on both ends, what happens to the humans in the exchange? We are, right now, running this experiment without a control group. The stakes and risks are high.

I want to be careful here. This is not a technophobia argument. I like AI. Artificial Neural Networks are the subject of my studies since I started my master's in early 2000's and still is. I literally have loss functions tattooed in my right arm. Tools have always reshaped human cognition, and the net result has generally been positive for what we can accomplish, even when it was mixed for what we directly experience. The printing press replaced trained memory and gave us advanced science. The calculator replaced mental arithmetic and gave us modern engineering.

But the thing potentially lost in this exchange is not a narrow skill. It is the daily practice of being a certain kind of person. Effortful. Patient. Present with other people. And I think that deserves more serious attention than it is currently getting from the people building these systems. This is beyond ethics.

Will AI ever be human?

Probably not. And I want to argue this is still the wrong question.

Pinocchio wanted to be a real boy. Start there.

The original story, not the Disney version, is darker and stranger than most people remember. Carlo Collodi's puppet is initially cruel, selfish, and impulsive. He lies instinctively, not strategically, and his nose grows because the lying is not yet connected to anything he understands about the relationship between truth and trust. This is worth pausing and reflecting on. The nose is not a punishment. It is a diagnostic. Pinocchio doesn't lie because he is calculating. He lies because he doesn't yet grasp what his lies cost other people. He lacks not just conscience, but the capacity to need one.

Then we have Jiminy Cricket.

The Blue Fairy assigns him as Pinocchio's conscience specifically because Pinocchio does not yet have a functional one of his own. Jiminy is, in the most precise sense, an externalized ethical alignment system. And throughout the story, he is constantly overridden, dismissed, and ignored by the very entity he was built to guide. If you work in AI safety, this should feel familiar. The alignment problem (which will be a future article which I am already working on) is, in a meaningful sense, the Jiminy Cricket problem: how do you build a conscience that can't be circumvented by the system it was built to serve?

The Blue Fairy's condition is also interesting. She does not transform Pinocchio directly into a human. She makes him real, flesh and blood, so that he has the capacity to earn being human. The transformation itself has to be earned through action. Specifically, through selfless action that costs him something real. The author sees sacrife and love as the ultimate human skills. At the end, he saves Geppetto from inside the whale. The creation saves its creator, not because the logic is sound, but because he loves someone. And the Blue Fairy completes the transformation not because Pinocchio performed humanity according to a checklist, but because he had developed genuine care and acted on it under genuine cost. It did not imitate it, it felt it.

Geppetto made Pinocchio, it is worth noting, out of loneliness. Out of the desire for a child, a companion, something that would respond to him. That impulse is not entirely absent from why we build AI. And like Geppetto, we may end up rescued by the thing we made, though probably not from a whale. Will AI become Pinocchio or Dr. Frankestein's creature, we might find out in the future.

Dennett would say: if AI systems become sophisticated enough that their behavior is indistinguishable from genuine understanding, genuine empathy, genuine moral reasoning, then the philosophical question of whether something real is happening inside is probably empty. What would it even mean to be human if not to do human things?

Chalmers would say: functional similarity misses the point. A philosophical zombie, a being that acts exactly like a conscious human but has no inner experience whatsoever, is at least conceivable. That means function and consciousness are separable in principle, and we need to take that seriously.

I don't think either of them is entirely right. But I notice that Pinocchio's story doesn't resolve the question through philosophy. It resolves it through love. Through the specific kind of action that only makes sense if something is genuinely at stake for the actor. Jiminy Cricket cannot make you human. The Blue Fairy cannot make you human, not fully, not without the rest of it.

The question "will AI be human?" may be like asking "will a symphony be a painting?" They are different kinds of things, capable of different forms of excellence, and the comparison flattens something important about both.

What I believe is this: the systems we build will increasingly blur the edges of the question. Not because the line doesn't exist, but because we never drew it precisely in the first place. Someone is in a coma, is the person conscient? Can we measure? Are we able to measure the level of conciseness of animals? If not, how do we expect to measure it in a machine?

The Mirror

Back to that room in Blade Runner.

The Voigt-Kampff test ultimately failed, not in the film's story but as a concept. In the later films and in the novel, the line between replicants and humans has blurred past the point of easy management. Replicants develop memories indistinguishable from real ones. Humans lose touch with what makes them distinctive. The test that was supposed to find humanity by its absence ends up finding something stranger: that the humans administering it weren't entirely sure what they were looking for.

Building AI has done the same thing to us.

We set out to replicate intelligence and discovered we didn't understand it. We tried to encode reasoning and found that reasoning depends on emotion in ways we hadn't accounted for. We built systems that could talk and write and argue and realized that talking and writing and arguing were not the same as thinking. We created something that could pass Turing's test and discovered that Turing had been asking the wrong question. We asked "what makes us human?" and realized, four hundred years into the debate, that we still don't have a satisfying answer.

Think therefore I am? Maybe. Maybe not.

The philosopher who looked into the machine and asked "is there anyone home?" found something unexpected: the question bounced back. The machine looked like no one was home. But the investigation forced us to ask what "home" meant, and that question turned out to be harder than the engineering.

"Life is a mirror and will reflect back to the thinker what he thinks into it." (Ernest Holmes)

This is what building AI has actually taught us about being human. Not a list of things we do that machines cannot. Not a reassuring inventory of biological advantages. Something more unsettling and more honest: we didn't know what we were when we started, and we know only a little more now.

The blueprint we were working from was always incomplete. AI held up the mirror. What we saw in it surprised us.

The question now isn't whether AI will ever pass as human. It is whether we understand, clearly enough to matter, what we are asking it to pass as.

And whether, in the meantime, we are practicing enough at being it ourselves.

Author: Igor Alcantara