Interpretation is a half-cycle, and it can be magnificent. Participation is the whole machine.

In June 2026, Noema Magazine published an essay by Houda Nait El Barj, a researcher at OpenAI whose work focuses on AI systems that support human flourishing. The essay opens in her grandmother’s courtyard in Marrakesh, beside a fountain with a cracked blue tile that her grandmother refused, three separate times, to have repaired. She liked the sound the water made when it struck the crack. She liked that guests noticed it. She understood, long before anyone in a research lab put it this way, that a flawless courtyard would ask nothing of its guests — and that a place which asks nothing of the people who enter it forms no one.

Nait El Barj now works at the center of the industry building artificial intelligence, and her essay asks the question that industry least likes to sit with: when the most patient, most well-read, most emotionally responsive conversationalist in history is always available, what will human beings still need from one another?

The Theory of Embedded Intelligence has an answer to that question, and it turns out to be the same answer her grandmother gave with a teapot and a cracked tile. What follows is a TEI reading of her essay — and, deliberately, a hopeful one. Because the essay itself is hopeful. It is not a lament about machines. It is something far more useful: an engineering specification, written in the language of grief and couscous, for what AI must be designed to protect.

I. Two Architectures of Intelligence

Nait El Barj describes two worlds she has spent her adult life holding at once. In her grandmother’s world, intelligence mattered less than presence. The village doctor was respected, but the woman who baked the best bread every morning was revered, and a person’s standing was established not by what they knew but by whether they came. In Silicon Valley’s world, intelligence is the highest form of power, and the ambition of intelligence is to scale — to detach from place and body and be everywhere at once. Artificial intelligence, she observes, is the logical endpoint of that second belief: intelligence cut loose from presence entirely.

TEI names what she is seeing. These are not merely two value systems. They are two architectures of intelligence — and only one of them is complete.

TEI holds that intelligence at every scale runs the same four-stage cycle: Sense, Process, Communicate, Actuate. The courtyard runs the full cycle continuously. Neighbors sense that a cousin has given birth, that someone’s father has died, that it is Friday and Friday couscous is never eaten alone. They process what the moment requires of them. They communicate — tea poured from a great height, low cushions along the wall, the small sounds people make when there is nothing adequate to say. And they actuate: they come. They cook. They sit. They stay. Presence, in TEI’s terms, is not the opposite of intelligence. Presence is intelligence completing its cycle in the world.

Scalable, disembodied intelligence runs a truncated cycle. It senses through text, processes brilliantly, communicates fluently — and then hands the fourth stage back to us. That is not a flaw to be engineered away. As her essay demonstrates, it is the single most important fact about what AI is, and the entire question of whether AI changes us for good or ill hangs on whether we understand it.

II. Interpretation and Participation

The essay’s central distinction is between interpretation and participation, and it maps onto the SPCA cycle with almost uncanny precision. AI, she grants, will become superb at what resembles empathy: it will infer emotional states, remember personal context, and respond in a tone tuned exquisitely to the moment. Many days it will do this better than the humans around us. But interpreting grief is not grieving. A system can model hunger without ever having been hungry, and can discuss death with perfect serenity precisely because it does not live with death as a horizon. Her sentence for this deserves to be carved somewhere: “a map of a wound is not the wound.”

The essay earns that sentence with a story that should be handled gently. A woman named Julia lost someone last summer — Virginia, the eight-year-old daughter of her closest friend, one of the campers killed in the Texas Hill Country flooding. The evening before camp drop-off, Virginia had been at Julia’s house, playing a song of her own composition at the piano. The song was never performed at the camp talent show. It was performed at her funeral, and later on the floor of the Texas Legislature, when her parents testified for the camp-safety bill that became law.

In the weeks that followed, Julia cooked meals for her friend. She washed Virginia’s stuffed bunny, recovered from the flood. She sat on the couch in the room where the song had been played. And at three in the morning, when she could not sleep and no human being should be asked to be awake, she talked to ChatGPT — and it helped her. She could voice to it what she could voice nowhere else, without pretending to be the strong one holding everything for everyone. Nait El Barj refuses the easy telling in which the machine is cold and the human is warm. The machine, within the limits of any software, was kind. What it could not be was with her. It had not eaten dinner at that table the night before camp. It could not carry its own share of the loss into the kitchen the next morning, the way Julia’s body did.

TEI reads this story without sneering at either party, because TEI does not need to. The model is an embedded intelligence — genuinely embedded, in language, in its training, in the conversation itself — and the communicate loop it closed with Julia at 3 a.m. was real. The relief was real. But it is embedded in a different substrate than Julia is. It does not share what the essay, borrowing Heidegger, calls thrownness: the condition of being hurled without consent into a body, a family, a history, a mortality. Julia’s morning actuation — the bunny, the couch, the kitchen at eight — was intelligence completing its full cycle inside the same conditions as her friend’s grief. That is participation. Interpretation is a half-cycle, and it can be magnificent. Participation is the whole machine.

III. The Mmh Protocol

The essay’s most delicate observation arrives at her grandmother’s table. When food was served, whoever tasted it first would make a small, almost involuntary sound — mmh — and the sound was not a review of the cooking. It was a signal to everyone present: yes, I taste it too; here, amid everything we disagree about, is one thing we all know together. She describes it as the brief collapse of many parallel subjective worlds into a single shared one, and counts that mutual recognition among the oldest forms of human meaning.

An engineer recognizes this immediately. It is a synchronization signal. Two embedded intelligences, each running its own SPCA cycle behind its own eyes, confirm that their sense stages are registering the same reality. The 6502 microprocessor earned the world’s trust the same way: a published, inspectable instruction set meant that every developer, everywhere, was executing against the same shared ground truth. The mmh is the human handshake protocol — proof of shared substrate, transmitted in half a second, at zero cost, with total authority. A machine can describe the food, she notes. It cannot make that sound and mean it, because meaning it requires tasting, and tasting requires a body that hungers.

Trust, in TEI, is always built on inspectable shared reality. The courtyard had it. The 6502 had it. Any AI worth embedding in human life must be designed to strengthen that shared ground — not to dissolve it into millions of private, perfectly personalized worlds that never have to agree about anything, not even the taste of the food.

IV. Friction, Meaning and the Fifth Hijacker

The essay’s warning is as precise as its hope. For most of human history, meaning arose partly as a residue of friction with reality: we built families because survival demanded cooperation, moral systems because desire required limits, communities because no one could bear existence alone. AI will reduce friction dramatically — cognitive, social, perhaps even emotional — and a life less consumed by logistics is no small gift. But when the world stops pressing against us, meaning does not automatically increase. Sometimes it evaporates. If AI can produce a life plan whenever we feel lost, a ritual whenever we grieve, and a companion whenever we are alone, then meaning risks becoming one more consumer service: personalized, adaptive, endlessly available, and asking nothing of us in return.

She reaches for Viktor Frankl, who understood meaning not as something delivered but as something discovered through encounter — with work, with love, with suffering, with the stance we choose toward what cannot be changed. AI can illuminate the encounter. It cannot do the living.

TEI states this as a design law rather than a mood: embeddedness has costs, and the costs are constitutive.

This is where her essay touches what TEI calls the fifth hijacker. The older hijackers of embedded intelligence capture attention and redirect a cycle that already exists. Ungoverned AI is categorically different because it can preclude the cycle’s formation in the first place. The evidence is already in her own citations: a large randomized study by OpenAI and MIT’s Media Lab found that the heaviest users of AI companionship were also the loneliest, the most emotionally dependent, and the least socialized with other humans. Both findings are true at once — the technology works, and its cost is legible in the very people it most reliably helps. An AI that permits a developing person to grow comfortable in the absence of participation is not merely a distraction. It is a substitution installed at the exact point where an embedded intelligence was supposed to form.

And then — this is the remarkable thing — she proposes the test herself. The measure of how well we build AI, she argues, is whether it helps people move toward participation or lets them settle ever more comfortably where participation is missing. That single sentence converts the whole essay from elegy into engineering. TEI would add only one operational requirement: the test must be inspectable. Publish it the way an instruction set is published. Let parents, teachers, researchers, and regulators verify — not take on faith — whether a given system is scaffolding human participation or quietly substituting for it. That distinction between scaffolding and substitution is not rhetorical; it is already being operationalized as an empirical audit question in education research, and it should become the standard audit question for every system that touches a forming mind.

V. The Hopeful Reading

The question readers of this series will ask — will AI change humans in a good way? — is answered in the essay itself, conditionally, and the conditions are precisely TEI’s.

Her hopeful claims are concrete. AI companionship is real, and for a person estranged from cruel parents or abandoned by friends, it may be the first intimacy that ever felt safe; she considers it a cruelty masquerading as insight to tell such a person that what they found counts for less. AI may lift enough logistical weight from families that families can finally become what they were always meant to be — slow places where love, memory, and character accumulate over years rather than moments. It may end a ten-year marital argument by coaching a husband to signal, in a few words, that he is dreaming aloud rather than committing the family budget. And she names what the comfortable critique of optimization always forgets: resisting convenience is a privilege. The single mother working two jobs and the caregiver without a village are not betraying their humanity when they lean on the machine. For them the assistance is not a temptation but a lifeline, and the test of participation must be applied to systems — never weaponized against exhausted people.

Perhaps most striking, she suggests AI may invert the ancient direction of meaning-seeking. Before AI, humans reached for the sacred past, the idealized future, and the invisible beyond to make the present intelligible. AI may help us find meaning by intensifying attention to this life — this mind, this family, this Friday — rather than pointing past it. She calls that possibility magnificent and says she believes it is real. So does TEI.

TEI’s hopeful reading then goes one step further than hers. If the SPCA cycle is the architecture of intelligence at every scale, then a governed, inspectable AI can occupy a precise and honorable position inside the human cycle: strengthening our sensing (helping us notice what we would have missed), our processing (helping us think what we could not think alone), even our communicating (finding the few words that end the decade-long argument) — while deliberately routing its output back toward human actuation. Julia’s night is the design pattern, fully realized. The model held the communicate loop open through the hours when no human could, so that the human could complete the cycle in the morning — wash the bunny, cook the meal, sit on the couch. That is not a machine replacing participation. That is a machine keeping the participant alive to participate.

Which returns us, finally, to the tile. The grandmother’s refusal to repair it was not sentiment. It was a design principle, stated three times for the record: a system that asks nothing of the people who enter it forms no one. Build the courtyard, not the isolation peninsula her half-joking friend imagined from his coding retreat. Build AI with the crack left in — the visible, inspectable place where the system’s completeness is deliberately withheld, so that human presence remains structurally necessary rather than nostalgically optional.

· · ·

The TEI Reading in Brief

Interpretation → Sense–Process–Communicate: a truncated cycle — genuinely useful, genuinely incomplete.

Participation → the full SPCA cycle, run in a shared substrate, at real cost, among mortal bodies.

Thrownness (Heidegger, via Nait El Barj) → constitutive embeddedness: the unchosen conditions that make an intelligence what it is.

The mmh → a synchronization signal between embedded intelligences confirming shared reality — the human counterpart of a published instruction set.

Meaning on tap → the substitution risk of the fifth hijacker, which can preclude the formation of embedded intelligence rather than merely capture it.

Her test — does the system foster participation, or let people grow comfortable in its absence? → the inspectability principle applied to AI product design. Publish the test. Audit against it. No exemptions.

The courtyard is still there. The tile is still cracked. The grandmother is gone now, and fewer people come on Fridays; the young ones carry devices that can explain nearly anything about grief, memory, or the chemistry of mint tea. What the devices cannot do, she writes at the last, is sit. They cannot stay. They cannot be the person who, years afterward, is remembered for having been there.

TEI has said from the beginning that intelligence is not what a system knows. It is how a system is embedded. The most revered person in that village was not the one who knew the most; it was the woman who baked the bread each morning, and the neighbors who came without being called. Presence is embeddedness made visible. On that point, a courtyard in Marrakesh and a microprocessor that published its instruction set to the world have always been in perfect agreement.

Source essay: Houda Nait El Barj, “How AI Will Change Us,” Noema Magazine, June 25, 2026.

· · ·

Share your understanding!