Han calls ChatGPT a parrot. He uses the word more than once, across multiple books and interviews since 2022, and he does not mean it kindly. In Vita Contemplativa and in the follow-up essays he published as the large-language-model wave broke, he argued — with the consistency of a philosopher who has been making the same argument for thirty years — that whatever the machine is doing, it is not thinking. It is replicating. It is sorting, analyzing, recombining. It is a very sophisticated parrot with a large vocabulary and a good imitation of grammar, and this should not be confused with thought.
I agree with the philosophy. I disagree with the conclusion he draws from it. This article is about where exactly the disagreement falls, and why it matters for what we build next.
#Where Han is right
Let us start with the parts that are not controversial, or should not be.
A large language model does not have an interior. It does not have a point of view. It does not have something-it-is-like-to-be-one. It does not want anything, fear anything, care about anything, or have anything at stake in its output. It does not reach for an idea. It predicts a token. The prediction is often brilliant, frequently useful, sometimes uncanny. It is not thinking in the sense Han cares about, which is the sense in which something new is brought into the world that was not there before.
Han's argument, in the frame of the previous article in this series, is that thinking requires the vita contemplativa — the idleness, the pause, the counterpoint to stimulus and response. The machine has no such mode. Han is explicit: AI knows only on and off. It cannot be idle. It cannot refuse. It cannot, as he puts it, be stirred. There is no silence before its thought because there is no thought, in the strong sense, to precede.
This is not a technical claim about transformer architectures. It is a philosophical claim about what thinking is. You can disagree with Han's definition. But if you accept his definition — thinking as the bringing-forth of something new, which requires the contemplative mode as its ground — then it follows immediately that the machine cannot think. And by the same logic, every attempt to "use AI to think for you" is a category error. Thinking is not what AI does. AI does execution.
This is Han's first point, and he is right about it.
#Where Han is wrong
Where he is wrong — or rather, where I think he stops the argument too early — is in the conclusion he draws.
His framing, which I will caricature only a little, is this: AI cannot think, therefore AI is a threat to thinking, therefore the appropriate human response is to resist AI, return to the vita contemplativa, and protect the human capacity for real thought from the imitation thought the machine offers.
This is half right and half catastrophic.
The half that is right is the insistence that imitation thought must not be mistaken for real thought. If you let the machine produce your strategy, and you adopt the strategy without judgment, you have given up the thing that makes you a strategist. If you let the machine write your emails, and the emails no longer sound like you, you have quietly conceded a little of the self that the emails were part of. Every place where we take the machine's output and treat it as our own without judgment, we become, in a small way, the parrot's ventriloquist's dummy. Han is right to warn.
The half that is catastrophic is the implied conclusion — that AI, because it cannot think, is without legitimate use. That the correct posture is refusal.
This is backwards. And I can say so carefully, because I have spent the last two years building software in close conjunction with these models, and the shape of the opportunity they present is the exact opposite of what a strict Han-ian reading would suggest.
#What actually changes when the machine can execute
Han's frame was formed by the Leistungsgesellschaft — the performance society — in which the defining fact was that the human does both the deciding and the doing. The Leistungssubjekt is exhausted because it cannot separate these two. It exploits itself because the same person who chooses the work is the one who performs it, and there is no relief between the two.
The machine, if we let it, changes this structure for the first time in the history of knowledge work.
Consider a concrete example. A week ago a small software company I advise had to produce a set of technical specifications for a regulator. Ten years ago, this would have taken a person two weeks of sitting at a desk, writing paragraphs, cross-referencing, reformatting, checking. The decision of what to specify is, in this case, the interesting work. The specifying itself is shallow labor — mechanical, exhausting, uncreative. The specifying was about 95% of the hours.
With a good language model the specifying took an afternoon. Two senior engineers spent the morning deciding what had to be true about the system. They handed the decisions to the model. The model produced the document. They read it, caught three errors, made two edits, and submitted. The week they got back was not, as Han might fear, spent on further optimization. It was spent deciding better things. More time on the judgment. Less time on the transcription.
What happened, in the frame of this series, is that the vita activa — the active, executing part of the work — got delegated. The contemplative part, the judgment part, the deciding-what-is-worth-doing part, did not get delegated. It could not be delegated. And because the active part stopped dominating the calendar, the contemplative part finally had room.
This is the exact inversion of Han's fear. The machine, correctly used, does not colonize thinking. The machine takes the execution, which was suffocating thinking, and removes it — which returns thinking to the space where it belongs.
The catch, and this is decisive, is the phrase correctly used.
#The two wrong ways to use a machine that cannot think
The first wrong way is to delegate the judgment along with the execution. This is the error that produces the worst AI-assisted work, the reports that feel plausible and say nothing, the emails that are correct English and mean less than the writer intended, the proposals that meet the specification and miss the problem. In every case, what got delegated was not just the labor of producing the output but the labor of deciding what the output should be. The model, having no judgment, produced competent prose aimed at the average of the training distribution. The human, having delegated the judgment, could not tell whether the output was right because they had outsourced the criterion by which rightness would be measured.
The second wrong way is the opposite, and it is the one Han is most alert to. It is the use of AI to do more of the same — not to free the contemplative mode but to accelerate the performative mode. You now write twice as many proposals because the model writes the drafts. You answer four times as many emails because the model drafts the responses. You take on twice the work because the throughput is higher. The Leistungssubjekt does not rest; it scales. The machine has not liberated you. It has removed the friction that was, in some small way, the natural limit on your self-exploitation.
This is the scenario in which Han's warning is exactly correct. An AI-accelerated Leistungssubjekt is a more thorough cage than the analog version. The throughput rises. The infarct from the previous article becomes more severe, faster, and harder to detect because the external metrics — output, response time, volume — all improve. The machine has made the trap tighter.
If Han is describing a world, he is describing this one: the one in which AI is primarily used to scale the performance society. And he is right to be horrified.
#The right way
The right way is not complicated to state. It is very difficult to execute, which is, I suspect, why it is rare.
Delegate the execution. Keep the judgment. Hold the ratio deliberately, and notice the moment it begins to slide.
Every task you currently do can be analyzed into two components: the decision about what-is-worth-doing, and the labor of doing it. In the pre-AI era, these were tangled. You discovered what was worth doing partly by doing it. The execution was a form of thinking, often a good one, sometimes the only one available. This is why there is a generation of writers who insist that writing is thinking — because for them, in a pre-machine world, the typing and the thinking could not be separated without loss.
The machine separates them. For most tasks — not all, but most — the machine is now good enough at the labor that the labor can be handed off without losing the judgment. The question, which is now the only question that matters, is: what did I decide that the machine did not and could not decide?
If the answer is something substantial — I defined the problem, I chose the audience, I set the constraints, I named what success looks like, I edited the result against a criterion only I have — then the machine has served you, and you have done thinking. The hours the machine absorbed are hours returned to the contemplative mode, and you should use them for contemplation, not for more execution.
If the answer is nothing substantial, I just prompted and shipped, then you have become the parrot's dummy, and Han is right about you specifically. The cure is not to stop using the machine. The cure is to use it while refusing to let the judgment migrate into it.
#Why this produces a sharper human, not a dimmer one
Here is the part Han misses.
When the execution gets cheap, the judgment gets visible. In a world where everyone can produce a competent document in twenty minutes, the value is no longer in the document. The value is in which document was worth producing. The person who is good at deciding what to work on — who has taste, judgment, a point of view, a relation to the problem that the average cannot replicate — becomes more distinct, not less. The machine cannot have taste. The machine cannot be opinionated. The machine cannot refuse. Only a human can.
Put another way: the arrival of the parrot raises the value of the original voice. The presence of the machine that cannot think amplifies the worth of the one that can.
This is a version of the argument Han himself should welcome, because it lands exactly where his earlier books pointed. The contemplative mode, which he rightly named as endangered, becomes the scarcest and most valuable mode in a post-execution economy. The person who has cultivated the vita contemplativa does not compete with the machine. The person uses the machine as a lever against the vita activa that was consuming them, and returns to the mode in which they are most themselves.
Han is afraid the machine will complete the colonization of thinking. I believe the machine, correctly positioned, is the first tool in a long time that could roll the colonization back — not because it thinks for us, but because it takes the labor that was preventing us from thinking at all.
The question is whether we use the machine that way, or the other way.
#What this means for what we build
The tool I am building has a position on this, and the position follows from the argument above.
Particle does not generate your work for you. Particle does not draft your emails, summarize your meetings, produce your plans, write your posts, or pretend to be an AI-powered productivity assistant. Every time I have been tempted to add one of these features — and I have been tempted many times — I have stopped, because the moment Particle begins doing the deciding rather than helping you hold the deciding, Particle has crossed the line from tool to parrot, and become part of the problem this series describes.
What Particle does do is hold the architecture around the judgment. It gives the day a shape. It protects the hours in which judgment happens. It closes the day on purpose so the judgment does not leak into the night. It makes the decision visible — what you chose to work on, what you chose to leave, what you chose to end. In the delegation society that is arriving, the record of human judgment becomes the thing worth keeping. Particle keeps it.
The machine will draft the report. The machine will run the analysis. The machine will answer the email. Particle will hold the hour in which you decide whether any of it was worth doing.
#The closing
Han is right that the machine cannot think. He is right that imitation thought is dangerous if mistaken for the real thing. He is right that a civilization too close to the parrot forgets how to speak.
He is wrong, I believe, that the appropriate response is refusal. The appropriate response is correct positioning. Let the machine take the labor. Keep the judgment. Use the hours the machine returns to you for the mode Han spent a career defending — the contemplative mode that thinks. Do not scale the performance society with AI. Dismantle it with AI, by refusing to do more of the wrong thing faster.
The thing the machine cannot do is the thing that is now most worth doing. It was always most worth doing. We just could not see it, because the execution had been drowning it out.
The final article in this series is about what remains of the human when the machine has taken everything it can take. Han has one answer. I have a more constructive one. They are not, as it turns out, in conflict.