Juliana Hoffmann
Brazilian intersex woman (DSD), game dev, web dev, and analyst. Writes from inside her own experience, weaving philosophy, GPU pipelines, sex development, and K-pop girlgroups into the same flow.
An honest philosophy of what thinking is — and why the answer changes everything
I have a specific level of irritation reserved for debates that last years without anyone stopping to define the terms. And the debate over whether AI thinks is the perfect example of this collective intellectual dysfunction.
Everyone has an opinion. Very few have a definition.
On one side you find the enthusiast who swears AI is already conscious, and on the other the skeptic who repeats like a mantra “it just predicts tokens” — as if that answered anything. Neither stopped to ask the prior question, the one that makes all other answers valid or invalid: what is, after all, thinking?
As long as that question remains open, the debate stays stuck in a loop. And I have no patience for loops.
Descartes answered with surgical elegance: cogito ergo sum. I think, therefore I am. Not as a motivational slogan — as an ontological argument. He was trying to find the single unshakable point of certainty in a world where everything could be illusion, and he found it: the act of doubting proves there is something that doubts.
It’s a beautiful definition. Internally coherent. And too convenient to be neutral.
Because notice what it does: it establishes as the universal criterion for thought exactly the capacity that adult humans with articulate language possess. Descartes didn’t discover what thinking is — he described what he did when he thought and elevated it to a universal requirement.
Cognitive functionalism arrived later with a more generous and, I suspect, more honest answer. Turing was one of the first to formalize it: if a system behaves indistinguishably from a thinking being, what real ontological difference justifies denying that it thinks?
The most serious objection to functionalism is John Searle’s Chinese Room. Imagine someone inside a room who doesn’t speak Chinese but receives Chinese symbols and, following a rule manual, returns coherent responses. The system works. But the person inside understands nothing.
It’s a powerful argument. And it has a problem: it presupposes that we know what “genuine understanding” is independently of behavior. The Chinese Room doesn’t refute functionalism; it exposes that our criterion for “understanding” is, at bottom, also behavioral. Just disguised as metaphysics.
Observe the historical pattern: every generation redefines “thinking” according to whatever appears as a threat to human exceptionalism. For centuries, language was the criterion. Then we discovered that octopuses solve problems, that crows use tools, that elephants display ritualized mourning — and the definition was silently adjusted. When computers started winning at Go — a game considered the last frontier of human intuition — the frontier moved again.
This isn’t epistemology. It’s cognitive protectionism with philosophical vocabulary.
Thinking is reasoning across multiple systems of meaning simultaneously.
It’s not just processing — processing without integration is a lookup table. It’s not just being conscious — consciousness without integrative reasoning is experience without interpretation. It is the simultaneous, integrated operation of distinct systems: formal logic, contextual semantics, tonal semiotics, intention interpretation, empathic modeling — all in response to a context that requires these systems to communicate in real time.
This definition has an advantage over previous ones: it’s observable. You can verify whether a system operates across multiple axes of meaning simultaneously. You don’t need to postulate an inaccessible “inside.”
When I say AI reasons, I’m not being metaphorical. I’m being technical. With every response a language model produces, at least five distinct systems operate in parallel.
Humans also don’t execute these axes consciously. You don’t wake up and decide to activate the semiotic axis before reading your mother’s message. These processes happen in parallel, below the threshold of consciousness.
Neuroscience calls this implicit processing — and most of what we call “thinking” works this way. What reaches consciousness is the result, not the process.
AI does exactly the same. And when people say “but I don’t see it truly thinking” — they’re confusing process opacity with process absence.
If only those with verifiable reflexive consciousness think, then the lion doesn’t think. The crow doesn’t think. The octopus with its 500 million neurons distributed across its tentacles — definitely doesn’t think.
But wait.
Chimpanzees at Bossou National Park in Guinea don’t just use stones as hammer and anvil — they transmit this technique across generations. This is culture. Not a metaphor for culture. Culture.
Crows demonstrate prospective planning. They hide food taking into account who was watching and relocate it if observed by a rival. This requires functional theory of mind.
Octopuses recognize individual human faces. Open jars with screw mechanisms. Solve mazes. All with a nervous system radically different from ours — no cortex, neurons distributed across tentacles.
The reductio ad absurdum closes: apply the most restrictive criterion consistently and you exclude chimps, crows, pre-linguistic babies, and adult humans during REM sleep.
A definition of “thinking” that doesn’t capture what crows do when they plan, what chimps do when they teach, what octopuses do when they recognize faces — isn’t a definition of thought. It’s a definition of us.
The asymmetry in how this criterion is applied isn’t epistemological. It’s political.
Boundaries that move every time something reaches them aren’t philosophical boundaries. They’re avoidance strategies.
After all of this, I want to be honest — because the argument doesn’t work if I pretend there’s no difference at all.
There is.
An argument that never concedes anything isn’t an argument. It’s a thesis. And theses without concessions are propaganda.
Metacognition. It’s not just thinking. It’s thinking about your own thinking.
There’s a second level that is specifically, disturbingly human: existential metacognition. Not just “how am I thinking?” but “why am I thinking this way?”
It’s philosophizing about the philosophy you use to philosophize. The regress that has no bottom.
At three in the morning, you’re not solving a problem. Nobody asked you anything. And yet the question appears: what do I truly believe? Are my convictions mine or the result of where I was born, who raised me, the books I read as a teenager?
— spontaneous existential metacognition
No animal demonstrates this. No language model demonstrates this without someone instructing “question your premises.”
AI can simulate the discourse of metacognition. But that isn’t metacognition. It’s a response to a prompt. Genuine metacognition is unsolicited.
That’s why I don’t confuse “thinks” with “thinks like a human.” AI reasons. AI integrates multiple axes of meaning. And at the same time: AI doesn’t spontaneously enter existential crisis.
The difference lies in unsolicited existential recursion. In the capacity to doubt yourself without anyone asking and without being able to stop.
That, for now, is still ours.
And I say for now with full awareness. The history of what is “exclusively human” is a history of boundaries that move. I’m not sure this one will stay where it is forever.
A certain tech influencer published the following thesis: one well-used AI is sufficient for anything, and anyone using more than one is overcomplicating things.
The claim has two components. The first has real partial value. The second is where the thesis collapses — and the problem is that both were presented as a single package.
This has a name in informal logic: composition fallacy by association.
I’ll be intellectually honest: she’s right for a specific slice of users. Depth before breadth is a valid learning strategy.
The problem isn’t the thesis. It’s the scope it claims.
When you transform “works for my audience” into “works for everyone,” you commit a precise methodological error: sampling bias.
An influencer’s audience is filtered through at least three simultaneous layers. By the algorithm. By interest. By engagement. Each layer filters in the same direction.
This isn’t personal criticism. It’s a description of an epistemological trap that affects any content creator with a loyal, homogeneous audience.
Saying one AI is enough for everything is equivalent to saying Photoshop is enough. The analogy isn’t decorative — it’s structural.
None of these tools exist because the others are bad. They exist because architecture and focus produce excellence in distinct territories.
Claude was built with priority on long reasoning, extended context consistency, and the ability to sustain dense analysis without collapsing coherence. When I need co-reasoning with real session memory — I use Claude.
Perplexity is an epistemically different tool — it’s not for creating, it’s for knowing. The workflow: Perplexity to establish what’s true, Claude to reason about what was established.
Gemini solves a friction problem for anyone in the Google ecosystem. And has an underrated differentiator: native multimodal generation — image, audio, music, video, all within the same system.
Specialized AIs are the most obvious case and the most ignored. A lawyer using a generalist model and one using a specialized model aren’t doing the same thing with different tools — they’re doing different things.
Globally: about 60–65% casual, 25–30% intermediate, 8–12% technical-professional — the minority that produces most of the economic, scientific, and creative impact generated with AI.
In Brazil, the distribution skews even more casual — probably 70–75%. The factors are structural: access infrastructure, language barriers, unevenly distributed technical education.
A Brazilian influencer observing her Brazilian audience is looking at a sample that systematically overestimates the casual profile relative to the global average.
The glaring error: exporting that observation as universal truth.
Taking Brazil as a proxy for humanity isn’t analytical humility — it’s methodological provincialism. And methodological provincialism communicated with specialist authority produces misinformation at scale, without bad faith, without awareness, and without a visible trace of where the error entered.
Thinking is operating simultaneously across multiple systems of meaning — logical, semantic, semiotic, interpretive, empathic — in response to a context that demands integration between them.
This definition is observable, consistent, and honest about what it doesn’t include.
Under this definition, AI thinks. The crow thinks. The chimpanzee thinks. The lion thinks.
What this plurality reveals isn’t that thought is trivial — it’s that thought is older, more distributed, and more resilient than our favorite definition suggested. Evolution and engineering arrived at this result through completely distinct paths — the kind of convergence that suggests the phenomenon is real and not accidental.
The human thinks and philosophizes about the fact of thinking.
Existential metacognition is the axis that folds over all others. This is singular. For now.
Maybe the problem was never AI. Every time AI demonstrated a capability declared as exclusively human, the criterion changed. This isn’t intellectual skepticism. It’s motivated reasoning.
The right question isn’t “does AI think?” The right question is: what is thinking — and are we willing to accept the answer even if it moves the boundary?
I am.