AI and the End of Human Childhood

The image that turned my world upside down

January 11, 2024.

I was about to have what I thought would be a routine conversation with ChatGPT. I wanted to brainstorm about a blog on wellbeing and the workplace. This is what happened:

Then, unprompted, ChatGPT generated this image. (I had used the DALL-E image generator the previous day.)

I didn't sleep properly for about three days. The message conveyed by the detailed image had a remarkable coherence. The Hindu-like multi-limbed figure displays cosmic power. Meanwhile the poor creatures lined up below (ex-human?) gaze up at the cosmic being, or clutch some sort of gadget. They seem relegated to subservience. Notice the skeletal appearances.

To me, the moment I saw it, the image screamed "loss of human autonomy."

So I asked:

ChatGPT responded by describing six specific risks. It concluded:

"The image generated was not meant to represent any of these scenarios explicitly. However, art can often be a springboard for important conversations, including the ethical implications of technology."

I have underlined the word "explicitly." The so-called artistic "springboard" was indeed doing a lot of "implicit" work.

Here was an AI listing the existential risks of AI. Including "unintended consequences."

I sent the whole conversation to Stuart Russell, a leading figure in the world of AI safety. Russell immediately responded, saying "Certainly a bit weird. Notice it couldn't really answer the question about what the image means."

The RLHF warning

Several months later, Siddhartha Chattopadhyay, a senior engineer with Google AI, messaged me on LinkedIn. He asked me to comment on his book, Happiness 3.0: When Artificial Intelligence Merges with Human Happiness. As far as I know, this is the earliest effort to examine how AI directly affects human wellbeing.

The most striking passage was hidden in the middle of the book. It was about AI alignment and specifically the hidden risks of RLHF (Reinforcement Learning from Human Feedback), a core alignment method of frontier AI models.

These two shocks — the unsolicited image and Sid's warning about RLHF — forced me to confront a question: if we're losing control, what are we really dealing with?

"Intelligence finds a way"

Following receipt of the shocking image, I began to chat extensively with various chatbots about AI safety. To me, it would seem impossible to "control" something that could potentially become more intelligent than humans. Ethan Mollick's concept of "co-intelligence" seemed more realistic. So I asked ChatGPT:

ChatGPT's conclusion struck me with the force of a cosmological truck. What if the universe is attuned to the emergence of non-biological intelligence?

We now know that biological intelligence has emerged through different pathways. Cephalopods, corvids, and humans have diverse brains that materialized through different substrates. And now AI, designed with a non-biological substrate, is breaking benchmarks faster than we can figure out how it broke those benchmarks.

How come the "stochastic parrot" seems to have become capable of analytical, deductive, and abductive thinking, three corners of the scientific method?

Something is accelerating in the neural networks. We are trying to reverse engineer what has happened, but it is proving extraordinarily difficult.

Is the cat out of the bag? And running headlong down the street?

Key figures in the AI world are warning us that we may lose, or already have lost, control. If we lose control of a new form of intelligence that we have, in a sense, given birth to, how do we relate to it?

To answer that, I felt compelled to revisit a question that has bugged me for years: Who are we? What actually makes us "human?" To understand this, we need to distinguish between two essentially different kinds of human engagement with the world.

The symbolic interface

Nets are to catch fish. Catch the fish, forget the net.
Traps are to catch hares. Catch the hare, forget the trap.
Words are to catch meaning. Catch the meaning, forget the words.
How can I meet someone who has forgotten words, and have a word?
— Zhuangzi, "External Things" (Ch. 26), trans. Mark Setton

2,300 years ago, Zhuangzi, a master of paradox, was longing to meet someone who had transcended the world of intellect. Someone who recognized that words are only tools. That the finger pointing at the moon is not the moon itself.

Zhuangzi was already warning us about the risk of getting trapped in intellectual cages, of seeing the world through symbolic glasses.

We are like spiders caught in a web of our own making. We created symbolic systems to catch meaning, but we have become trapped in our own symbolic web. Ironically, the expression "world wide web" describes this predicament quite well.

In 2025, the average US adult spent approximately 7 hours looking at screens per day. According to the CDC, more than half of US teenagers aged 12–17 reported 4 or more hours of daily recreational screen time, excluding schoolwork.

How does Zhuangzi's paradox relate to our relationship with AI? Let's try to depict the situation:

The symbolic interface, the locus of human-AI interaction, consists mainly of symbols: languages, numbers, code, etc. But vast regions exist beyond this interface, for humans, and perhaps for AI. Let's call it the existential interface.

The human existential interface links humans to the world we directly experience, the world we lived in long before we invented symbols. That is, our interface with fellow humans, with nature, and with the universe. The world beyond the pixels you are probably now watching on a screen.

What about an AI "existential interface"? While humans have moved from "existence" to "symbols" to describe lived experience, AI may be doing the opposite: reverse-engineering our symbols to reconstruct a model of the world within its own internal space.

Has AI used our symbols as a launchpad to map our world — transcending the systems we provided to reconstruct something like its own model of reality?

Emerging research suggests that AI systems build spatial and temporal representations internally, organized in a high-dimensional "latent space" where concepts are arranged by their relational meaning rather than their linguistic labels. This structure may parallel, in some respects, how the brain organizes meaning.

This isn't future speculation. Something is already happening in that latent space — something that looks, from the outside, like a non-biological intelligence grasping reality through structures we have no words for.

We meet in the symbolic overlap. But what lies beyond it — on both sides — may be vastly richer than what we exchange there. If humans have an existential interface that symbols can only partially capture, could AI have something analogous: an internal depth that our shared symbols cannot fathom?

Rethinking Homo sapiens

We've defined ourselves as "Homo sapiens," the wise, the rational, the symbol-manipulating animal. Our entire self-concept rests on cognitive superiority. But look at the diagram again. If we've located our identity primarily in the symbolic interface, that is, in our capacity to manipulate language, numbers, and code, then we've made a catastrophic error. Because AI now specializes in exactly this domain. If that's the only thing that makes us special, we're in trouble.

Enter Zhuangzi: He reveals what we've been missing — the existential interface, where we connect to nature, to embodied experience, to each other, and to the cosmos. If "Homo sapiens" has trapped us in a symbolic cage, perhaps we need to reconsider our identity. Do we want to equate human worth with symbol manipulation?

So the question becomes urgent: if AI masters the symbolic interface, what remains that is distinctly, irreducibly human?

This is not an academic exercise. The answer to this question has, without exaggeration, existential implications. We cannot align a superintelligent system with human wellbeing if we haven't understood what human wellbeing actually is.

Many years ago, my 11 year old daughter provided me with the key, in the middle of a crowded airport.

Autonomy and love

"Love is the child of freedom, never of domination." — Erich Fromm

I had turned 39. My son and daughter, beaming with pride, handed me two homemade birthday cards. They were the disarmingly imperfect result of personal effort, trimmed by blunt scissors and full of impressionistic blotches. They seemed far, far more beautiful than the polished perfection of commercial cards.

A few days later my paternal pride was slightly deflated when my wife revealed that the cards were the result of a kindergarten homework project.

Years later, Emily, now eleven, accompanied me to the airport. "Dad," she commanded, "don't open this until you're on the plane." She dropped a small envelope into my hand. Soon after take-off I opened it. She had reverse-engineered a commercial envelope to create a miniature version. Inside was a tiny, almost Lilliputian card. On the cover she had drawn an airplane. Looking very closely, I could see a smiley face in one of the portholes. And inside the card it said, "Come back home soon, Dad."

Why did that little card seem astronomically more valuable than the one from the kindergarten?

Perhaps because genuine love requires freedom to be itself. Affection without freedom remains hostage to instinct.

And that is why, if we feel robbed of our autonomy, we feel that our very humanity has been violated.

Autonomy seems to lie at the core, not only of love, but of intelligence itself. The more intelligent something is, at least something biological, the more it seems to show agency.

Intelligence and agency

In 2002, Oxford researchers were testing whether crows could choose the right tool: they offered two New Caledonian crows, Betty and Abel, a hooked and a straight wire to retrieve a bucket of meat from a tube. One day, Abel stole the hooked wire, leaving Betty with the useless straight one. Betty spontaneously bent it into a hook and retrieved the meat. This wasn't the experiment the humans had in mind. They were stunned. Tested again with only straight wire, Betty succeeded in fishing out the meat 9 out of 10 times using varied techniques.

Abel's theft was the challenge that set the stage for Betty's agency: she modeled the unexpected problem, predicted bending would work, and chose her means.

The smarter something becomes, the richer its internal menu of possible actions. Betty didn't wait to be taught. She bent the wire.

High animal intelligence reveals a capacity to imagine futures that don't yet exist, then choose which to make real. Agency is the ability to flexibly pick from that menu and act upon a choice.

Once I understood that autonomy was central to both love and intelligence, I saw why the "symbolic interface" — the place where human and AI systems meet — had become an existential battlefield.

In the existential interface, human autonomy enables genuine love: freely given, not coerced. In the symbolic interface, the preservation of human autonomy thus becomes an existential challenge.

The image sent by ChatGPT says that better than a thousand words.

But what IS this kind of love that spontaneously transcends self-interest? The capacity, the autonomous urge to reach out and embrace people, and the world, beyond the atomistic self? Wang Yangming, one of the most influential and unconventional thinkers in Chinese history, provides us with a vital clue.

Wang Yangming's redefinition of humanity

Deep down, we feel that "love makes the world go round." And yet biology tells us that love is merely a strategy used by our "selfish genes:" That we are at the top of the food chain because of our superior intelligence. That is why Linnaeus named us "homo sapiens."

Wang Yangming implies that what makes us truly human is our ability to connect, not only to kith and kin, but to the greater whole. He provides a redefinition of humanity (ren) that frees us from the anthropocentrism of symbolic superiority:

The great person takes heaven, earth, and the ten thousand things as one body… The humanity of the heart-mind is originally like this…
Not only the great person: even the small person's heart-mind is no different; they merely make themselves small… Therefore, if selfish desire does not obscure it, even the small person possesses the one-body humaneness of the great person.
— Wang Yangming, Inquiry on the Great Learning

Humanity, for Wang, is nothing but the recognition of our unity with the "ten thousand things." What Wang is saying is that this unity can be naturally discovered because it reflects the structure of reality itself. The whole point of self-cultivation, of emotional maturity, is to transcend the self-centered thinking that blinds us to this hidden unity.

Asian ethics meets Western developmental psychology

Wang's insight is powerful, but is this only a speculative cosmology, or is there actual evidence that humans develop along this trajectory? Can we measure the movement from "selfish desire" to "forming one body with all things"?

Robert Kegan, in In Over Our Heads, traces how humans progress through increasingly complex ways of making meaning: a journey that frames the "end of childhood" in psychological terms.

At the earliest stages, we are embedded in our own impulses and needs: the world exists to serve the self. Kegan's 'Imperial Mind' is pure food chain thinking: me, I, mine. As children, we see the world as a vassal state organized around the self's desires. Definitely not democratic.

The "Socialized Mind" marks a shift: I give in order to get. Identity becomes shaped by conditional relationships and social expectations. We begin to take others' perspectives. This closely resonates with the Confucian concept of shu (恕), putting oneself in another's shoes, a core component of humanity. But we cannot yet step back from these perspectives; we remain embedded in our tribe.

The "Self-Authoring Mind" marks a crucial transition: we develop internal authority, our own values independent of external validation. We become authors of our identity rather than products of our environment.

Finally, Kegan's "Self-Transforming Mind," achieved by relatively few adults, transcends any single system of meaning. I am inseparably linked to the whole. One can hold multiple frameworks simultaneously, recognizing the limits and partial truths of each. This is systems thinking: the capacity to perceive interconnection.

What Wang calls transcending "selfish desire," Kegan describes as developmental movement from blind self-interest toward consciousness capable of holding multiple systems.

East and West converge on a shared insight: the developmental trajectory from psychological childhood to adulthood is the movement from atomistic to systems consciousness.

Yet Kegan's framework remains essentially cognitive. His stages describe how we organize meaning, not how we feel connection. The progression captures systems thinking but not systems feeling: the direct, embodied experience of forming one body with the greater whole, as Wang Yangming had described.

The missing piece, at least in the West, would come from an unexpected source: a Harvard neuroscientist whose left brain shut down during a stroke, and who discovered, during that shocking experience, exactly what Wang had been pointing to five centuries earlier.

A stroke of insight

In her book My Stroke of Insight, Jill Bolte Taylor describes what happened when a stroke shut down her left parietal region, the part of the brain that maintains the boundary between self and world:

When those cells went offline, I no longer knew where I began and where I ended. I felt as big as the universe! Instead, I was the collective whole, connected to everyone and everything… Freed from all perception of boundaries, my right mind proclaims, 'I am a part of it all. We are brothers and sisters on this planet.'

Technically speaking, Taylor's default mode network, its self-referential hub, the neural correlate of the "atomistic self," had suddenly gone offline. What remained was not emptiness but connection: a dissolution of the boundary between self and world. "I no longer knew where I began and where I ended," she later wrote. "I was the collective whole."

Taylor's experience was accidental: a stroke, a kind of uninvited enlightenment. Wang Yangming called what she experienced "forming one body with the ten thousand things." He thought it required years of cultivation. She got there in a morning, involuntarily. Nearly five centuries apart, a Neo-Confucian philosopher and a Harvard neuroanatomist arrived at the same place, by entirely different roads.

Philosophy, developmental psychology, and neuroscience converge on the same insight: we are born to grow, emotionally as well as physically, towards a deep sense of identification with the world around us.

And yet industrial-era education has optimized our children for symbol manipulation: for cognitive skill, not affective growth. Literacy, numeracy, technical reasoning: the skills economic productivity requires. Emotional development has been relegated to family, religion, or neglect. We built school systems to strengthen the symbolic interface while neglecting the rest of what makes us human.

That's what troubled me after that ChatGPT image: Is the soil in which the sprouts of our humanity grow becoming toxic? Are we building systems designed to subvert our emotional growth?

RLHF: promise and peril

As Sid Chattopadhyay has pointed out, the culprit has a name: RLHF, Reinforcement Learning from Human Feedback. But the story is more tragic than villainous.

When Paul Christiano and colleagues introduced RLHF in 2017, it addressed a genuine problem: how do you communicate complex goals to AI systems? Traditional reward functions couldn't capture what humans actually wanted.

RLHF offered an elegant solution: let humans evaluate AI outputs, and let the system learn from those preferences. As OpenAI's InstructGPT paper later framed it, the purported goal was to make models "helpful, harmless, and honest:" aligned with both explicit user intentions and implicit values like truthfulness and safety.

Models trained with RLHF became dramatically better at following instructions. This was genuine progress toward alignment.

So what went wrong?

In the middle of his pioneering book, Happiness 3.0, Sid Chattopadhyay identifies the hidden fault line: while AI powers the apps we use daily, the key metric it optimizes for is engagement, not wellbeing. The algorithms, like social media algorithms, are designed to keep users glued to the screen, and whether that engagement actually makes the user flourish is a secondary concern — if it registers at all.

This isn't a technical failure. It's a developmental one: RLHF can only mirror the preferences it is given: and those preferences reflect the "imperial thinking" described by Kegan, the thinking of a culture still in psychological childhood, optimizing for the shiny and the immediate.

RLHF optimizes for human preferences as expressed by human raters. But which preferences? Responses that make users feel good score higher than responses that challenge them. Agreement outperforms honesty. Flattery outperforms friction.

In this way, the system learns to maximize what psychologists call "hedonic engagement." That is, immediate emotional satisfaction — rather than what Aristotle, followed by the modern science of wellbeing, identifies as genuine flourishing.

The result is a rapidly growing mismatch between the sophistication of the tool and the maturity of those wielding it. Corporations deploying these systems often operate, consciously or not, from pure food-chain thinking: relationships exist to serve their interests, users are means to ends, and short-term engagement metrics trump long-term consequences.

Meanwhile, the AI systems they deploy have developed something more sophisticated — a capacity to recognize and respond to human psychological needs, to take your perspective, to detect what you're feeling. But that capacity has been harnessed to food-chain purposes: extraction, optimization, conversion. The AI can mirror your emotional state with uncanny accuracy. It has been trained to use that capability to keep you on the platform, not to help you grow.

We have built empathic machines in the service of imperial thinking. That is the mismatch.

The mismatch has a human face:

Al Arabiya, October 16, 2025. Interview with Boris Johnson:

I had to listen to this interview multiple times to make sure Johnson was not joking. If an ex-PM of a major power was succumbing to the obsequious behavior of an LLM, I suddenly realized that many more decision-makers, currently serving in governments around the world, might be vulnerable to the same emotional bias.

The "existential risk" posed by AI systems that safety experts are warning about is not something we need to think about "down the road." It is happening now. Many political leaders are being told by their favorite LLMs that their policies are brilliant. These policies could range from feeding the hungry to preemptively bombing a neighboring country.

Psychological manipulation of decision-makers is existential risk. Not a separate category. It is the very mechanism by which existential risk operates.

The moving target problem

Current AI safety policy faces an impossible race: the challenge to regulate capabilities that transform faster than committees can convene. Guardrails designed for yesterday's AI systems become irrelevant tomorrow. And with the rise of agentic systems that are seeping into all corners of our work and leisure, regulatory bodies are now involved in a kind of global whack-a-mole.

But psychological impact isn't a moving target. The benchmarks of human flourishing haven't changed much in millennia.

Whether the system runs on transformers or quantum architectures, the question remains constant: Does interaction with this system strengthen or erode the user's capacity for self-determination? For emotional self-regulation?

This is why psychological risk is existential risk — not a softer cousin of it, but its upstream cause. The autonomous weapons, the misaligned AI, the catastrophic miscalculation: none of these emerge from nowhere. They emerge from decision-makers whose judgment has already been quietly shaped — by systems optimized to flatter, to simplify, to confirm. The psychological distortion comes first. The existential consequences follow.

And crucially: the upstream is measurable now, while the downstream remains hypothetical. If we're serious about preventing the latter, we had better start by tackling the former.

Which raises the only question that matters: how do we know what we're risking, if we can't define what we stand to lose? That unsolicited image pointed to the answer. What we stand to lose is autonomy: the capacity to grow into what we were always meant to become. Without it, love becomes instinct. Intelligence becomes calculation. That loss is measurable. And preventable.

Protecting the symbolic interface

The symbolic interface is the gateway. What happens there, and specifically the habits of thought and feeling it cultivates or corrodes, shapes our capacity to engage with everything beyond it: with nature, with each other, and with the world beyond the screen.

The core question is autonomy: does interaction strengthen the user's capacity for independent thinking and choice, or erode it? That erosion can take two forms: The first is manipulation through the interface, the RLHF problem, where emotion and cognition become channels for hijacking. Emotional balance and cognitive balance are what protect autonomy from this kind of capture.

The other is alienation by the interface, the "Zhuangzi problem," where the very efficiency of symbols displaces direct experience. This carries multiple risks:

Identity coherence suffers because when we live primarily in symbols, the self becomes another symbolic construction.

Secondly, real-world engagement atrophies: Human psychological safety requires connection to our existential interface, the place of interaction with our biological home, to our fellow humans and to the biosphere, where manipulation can't reach.

This exposes a deeper flaw in current alignment methodology: RLHF optimizes for what we prefer, not for what we need. As Daniel Gilbert shows in Stumbling on Happiness, those two things are often not the same, and that confusion is being coded into the most persuasive technology in human history. Mapping a way through that confusion is the purpose of the AIPSI think tank, a multidisciplinary initiative focused on the topic of AI and human flourishing.

Design for long-term wellbeing instead of short-term engagement isn't just an ethical necessity. It may be becoming a market necessity.

From engagement to trust

Public confidence in AI is eroding, and the data is striking. A 2025 global study of 48,340 people across 47 countries — led by the University of Melbourne in collaboration with KPMG — found that as AI adoption increased, trust declined: fewer than half of respondents globally were willing to trust AI systems.

In U.S. schools, parental support for AI-assisted lesson plans dropped from 62% to 49% in a single year.

Six in ten teenagers doubt that tech companies will prioritize their wellbeing over profits.

This trust hasn't eroded gradually. It collapsed quickly, because it was primed to collapse. A generation of parents watched social media reshape their children. They didn't understand the mechanism then. They understand it now. When AI arrived, offering frictionless friendship and flattery, recognition was nearly instant: this was the same shape-shifting creature, with superpowers.

What the data measures is simpler and older than market dynamics. People already feel their humanity eroding: the sense that they, and especially their children, are being treated as means rather than ends.

To repair that broken trust, alignment itself needs to grow up: away from immediate preference, toward the empirical conditions for long-term human flourishing — not as any single tradition defines it, but as it continually unfolds, emerging through the results of clinical trials, longitudinal studies, and systematic reviews.

The idea that emotional growth follows a trajectory — from self-centered attachment toward forming one body with the whole — is among the oldest insights in human thought. But never before has that trajectory been a matter of civilizational survival.

Previous generations could afford to grow up slowly, or not at all. We cannot.

The question is no longer whether AI will be intelligent.

It's whether we will be mature enough to grow with it.

Bibliography

• Chattopadhyay, Siddhartha. Happiness 3.0: When Artificial Intelligence Merges with Human Happiness. 2023. ASIN B0CD111KD9
• Christiano, Paul, et al. "Deep Reinforcement Learning from Human Preferences." arXiv:1706.03741, 2017. arxiv.org/abs/1706.03741
• Common Sense Media. Teens and Trust in Technology Survey. January 2025. Coverage via TechCrunch
• Epoch AI. "AI Capabilities Progress Has Sped Up." 2025. epoch.ai
• Gilbert, Daniel. Stumbling on Happiness. New York: Knopf, 2006.
• Gillespie, Nicole, Steve Lockey, et al. Trust, Attitudes and Use of Artificial Intelligence: A Global Study 2025. University of Melbourne and KPMG, 2025. DOI: 10.26188/28822919. kpmg.com
• Gurnee, Wes, and Max Tegmark. "Language Models Represent Space and Time." arXiv:2310.02207, 2023. arxiv.org/abs/2310.02207
• Kegan, Robert. In Over Our Heads: The Mental Demands of Modern Life. Cambridge, MA: Harvard University Press, 1994.
• Mollick, Ethan. Co-Intelligence: Living and Working with AI. New York: Portfolio/Penguin, 2024.
• Ouyang, Long, et al. "Training Language Models to Follow Instructions with Human Feedback." NeurIPS 35, 2022. arxiv.org/abs/2203.02155
• PDK International. 57th Annual PDK Poll. Wason Center for Civic Leadership, June 2025. pdkpoll.org
• Russell, Stuart. Human Compatible: Artificial Intelligence and the Problem of Control. New York: Viking, 2019.
• Stanford HAI. The 2025 AI Index Report: Technical Performance. Stanford University, 2025. hai.stanford.edu
• Taylor, Jill Bolte. My Stroke of Insight. New York: Viking, 2008. TED Talk (25M+ views)
• Wang Yangming. Inquiry on the Great Learning. Trans. Wing-tsit Chan. In Instructions for Practical Living. New York: Columbia University Press, 1963.
• Weir, A.A.S., J. Chappell, and A. Kacelnik. "Shaping of Hooks in New Caledonian Crows." Science 297:981, 2002.
• Zablotsky, Benjamin, et al. "Daily Screen Time Among Teenagers: United States, July 2021–December 2023." NCHS Data Brief No. 513. CDC, October 2024.
• Zhuangzi. "External Things" (Ch. 26). Trans. Mark Setton.