Wang has a rule about AI in his own writing process. The rule is one word.
Verboten.1
He says this, on the podcast, with the specific awareness that the word is a strong one. I will try as much as possible and perhaps even declare it verboten not to let AI get into my writing. I don't want it prompted to have any sentences that I could actually just use. I want the sentences to come from within myself.1
This is interesting because Wang is not an AI skeptic. He is, by his own account, using AI all the time.1 He uses it for restaurant recommendations. He uses it to talk about music. He uses it to think through novels he has just read. He uses it for art, for conversations about culture, for the kind of curiosity-driven exploration that used to require either a knowledgeable friend or a great library or both. He calls it a super tool.1 The praise is sincere.
But the moment AI approaches his writing, the door closes. Not for ideological reasons. Not because Wang thinks the model can't produce serviceable prose. The opposite, in fact — Wang's complaint about AI prose is that the sentences are generally super flat, and his concern is precisely that he might use them.1 The verboten is a defense against the temptation. He knows he could prompt the model to produce a sentence that would be good enough to drop into his manuscript, and he is closing that door before he opens it.
The rule is built around a specific theory of what writing is and what would be lost if AI got into it. Understanding the theory is the work of this page.
When Wang says I want the sentences to come from within myself, the phrase is doing more work than it sounds like.1 He is not making a Romantic claim about authentic self-expression. He is making a much more specific claim: the sentences in my writing are how my mind organizes itself, and if those sentences are produced by a system that is not my mind, the writing stops being a record of my mind and becomes a record of something else.
This is closer to the cognitive-neuroscience claim that thinking happens partly in the act of writing than to any aesthetic preference. The writer who outsources sentence-construction is outsourcing the cognition that the sentence-construction would have produced. They get the sentence, but they don't get the thought the sentence-construction would have surfaced — the half-formed insight that becomes visible only when the writer has to find the words for it, the constraint that produces a sharper formulation when the obvious one fails. None of this happens when the model produces the sentence. The output looks like writing; the underlying thinking has not occurred.
For Wang specifically, who has built his analytical voice on texture and zoom in/out across scales and sentence as structural anchor for an entire essay, the loss would be catastrophic. Wang's voice is constituted by the cognition that produces his sentences. If the model were producing the sentences, Wang would still be the author of the byline but not the author of the voice. The voice would degrade by exactly the amount that the model's contribution increased. Over time, with continued use, the voice would converge toward what models produce when prompted to write like Dan Wang — which is not what Dan Wang produces when Dan Wang writes.
The verboten is not a limit on AI capability. It is a defense of the cognition that the writing is supposed to be of.
The defense of writing happens alongside an enthusiastic embrace of AI in another domain. Wang's framing is that AI is primarily a form to enhance my consumption.1 The phrase is more careful than it first appears — he does not say "consumption" loosely; he is staking a claim about the kind of activity AI is structurally well-suited to.
What does enhanced consumption look like in practice?
Music. Wang knows little about music theory. He listens to Mozart, Beethoven, Mahler with affection but without the technical vocabulary to describe what he's hearing. With AI, he can ask: what should I listen to in this string quartet, what is happening in the second movement, why does this passage feel different from the rest of the piece? The AI provides a musicological gloss that Wang could not have produced himself and that, in his Ottawa childhood, would have required either a music professor or hours in a music library. In terms of a counterfactual of what I couldn't get before to what I can get now, it might be the place with the biggest delta in all of the AI world.1
Novels. Wang reread Stendhal's The Red and the Black in the summer of 2025. After finishing, he interrogated the AI about the novel — what literary scholars and critics have said, which interpretive frames have been developed, what the consensus and contested readings are. His own experience of the novel had been ecstatic and personal, necessarily individual. The AI conversation was not a substitute for that experience; it was a layer added on top of it, providing the critical apparatus that allows the reader to think about what they have just read.1
Restaurants and clothing. Pedestrian, in the literal sense — Wang asks AI for a noodle shop in a neighborhood, for a piece of clothing he is interested in. The AI is functioning as a knowledgeable friend who has read everything and remembers everything, available at any time of day.
Music and art conversation generally. I think the AI is really really good for thinking through a lot of these different sort of things.1
The pattern across these uses is that the AI is providing structured access to expert knowledge that the user is not themselves expert in, oriented toward the user's existing experience rather than producing new artifacts. Wang has the experience of reading Stendhal; the AI provides the critical apparatus. Wang has the experience of hearing Mozart; the AI provides the musicological gloss. Wang has eaten somewhere; the AI suggests a similar place. The AI is a layer on top of Wang's life, not a substitute for it.
This is what enhanced consumption means. The user keeps doing the experiencing — the reading, the listening, the eating, the looking — and the AI enriches the layer of understanding that wraps around the experiencing.
The boundary between enhanced consumption and substituted production is the operative line in Wang's framework. Why does it hold?
The boundary holds because the two activities have different relationships to the user's cognition.
Consumption is downstream of the user's experience. The user has read the novel. The user has eaten at the restaurant. The user has heard the piece. The AI's contribution is critical commentary, comparative context, expert framing — all of which are additive to the user's already-existing cognitive engagement with the object. The user is enriched by the AI's contribution, but the user's underlying cognitive activity is unchanged.
Production is the cognitive activity itself. The user is generating sentences, arguments, structural choices. If the AI substitutes for the user's sentence-generation, the user has not done the thinking the sentences encode. The output exists, but the thinking does not. The user is impoverished, even if the output is fine.
The same AI tool is acting on opposite sides of the boundary in the two cases. In consumption, the AI augments cognition. In production, the AI replaces it. Wang is not making an across-the-board claim about AI; he is making a boundary claim. Use it on this side. Don't let it cross to the other side.
The boundary may be hard to maintain in practice because the same interface is available for both. ChatGPT will discuss Stendhal with you and ChatGPT will write you a passage about Stendhal in the style of a particular essayist. The first is consumption-enhancement; the second is production-substitution. The user has to decide which mode they're in, and the decision has to be sustained against the constant temptation to slip from one to the other. Wang's verboten is the discipline that holds the boundary against the slippage.
Wang adds a separate but related argument about when AI should enter a person's life. The argument is generational and structural.
He frames it through Tyler Cowen's experience. The internet dropped into Tyler's lap about halfway through his life. He had already gotten his PhD. He had already taught at UC Irvine and George Mason before the internet became a really big thing in the late 90s.1 Tyler arrived at the internet with a fully developed cognitive apparatus — research skills, writing discipline, intellectual networks, the habit of going to libraries. The internet turbocharged what Tyler already was; it did not replace the cognitive infrastructure he had built before it arrived.
Wang applies the same logic to himself and AI. Wang is 33. ChatGPT blew up when he was 30. He has lived half my life without AI, and he proposes to let the super tool take over my life in the second half.1 The structure matters. The first half built the cognitive infrastructure — go to the library, check out the databases on FRED, do some spreadsheets, do some PowerPoint presentations, learn a lot of these basic skills, learn how to read1 — and now the AI arrives as a multiplier on top of those skills.
The contrast is with Stanford undergrads in 2026. Wang sees them at the Hoover Institution library, using ChatGPT to check through everything that they want to do and then once they're done with that they're watching some TikTok short videos.1 His judgment is gentle but firm: that just doesn't feel like that's the very best form of learning to me. The undergrads are skipping the cognitive infrastructure layer that Tyler and Wang built before AI arrived. They are letting the multiplier operate without the underlying skills it is supposed to multiply, and the result will be people who can deploy AI fluently without having developed the cognition the AI is supposed to amplify.
This is not a claim that the undergrads will fail. It is a claim that they are missing something. The experience of not knowing the answer, having to find it, building the muscle of finding it, is the foundation that makes the AI multiplier produce more rather than substitute for. Skip the foundation, and the multiplier multiplies less.
The framework converts cleanly into a set of practical heuristics for the writer or analyst working with AI.
Heuristic 1: Identify which side of the consumption-production boundary you are on right now.
When you sit down at the keyboard, the first question is whether this is a consumption session or a production session. Asking AI about a topic you want to understand is consumption. Generating a paragraph for a piece you are writing is production. The two activities should happen in different sessions, ideally on different days, because the cognitive mode required for each is different and the slippage between them is dangerous.
Heuristic 2: Use AI freely on the consumption side.
The conservative posture toward AI is wrong on the consumption side. AI conversation about novels, music, art, history, restaurants, clothing, philosophy, science is genuinely enriching, and the user who refuses to use AI on the consumption side is leaving signal on the table. Wang's enthusiasm here is not exaggerated — the music-and-novel use case really is one of the highest-leverage applications of current AI for non-specialists.
Heuristic 3: Maintain the verboten on the production side.
The user who wants to develop a distinctive voice, to do the cognitive work that writing is for, to produce sentences that record their own mind's organization — that user has to keep AI out of the sentence-generation step. Drafting, revising, structural decisions, sentence rhythm, word choice — all of this has to be human, all the way through, every time. Some users will choose to relax this constraint for utilitarian writing where voice is not the point (corporate emails, technical documentation, press releases). For voice-bearing writing, the verboten holds.
Heuristic 4: For new users, build the cognitive infrastructure first.
If you are young or newly engaging with serious cognitive work, the order matters. Spend two or three years using libraries, databases, spreadsheets, paper notes, longhand drafting, and other pre-AI cognitive tools before letting AI become your default. The infrastructure built in the pre-AI phase becomes the foundation that AI later multiplies. Skip the foundation and the multiplier multiplies a smaller thing.
1 Wang's account of his AI use runs from lines 408-462 of the transcript. Specific claims and locations: AI started after the book was done (line 408); GPT-3 disappointment, "03" breakthrough (line 408); "important about delivering one last big project before letting this super tool into my life" (line 412); Tyler Cowen internet-mid-life framing (line 416); Wang age 33, ChatGPT at age 30, "lived half my life without AI" (line 418); FRED database, spreadsheets, PowerPoint, learn to read as foundational skills (lines 420-426); Stanford undergrads ChatGPT + TikTok concern (lines 426-428); music as biggest delta (line 430); Stendhal Red and Black AI conversation (line 432); cultural experiences as personal/individual but AI good for thinking through (lines 434-438); "I view it primarily as a form to enhance my consumption" (line 438); "still love going through books and going through databases" (line 438); "verboten not to let AI get into my writing... I want the sentences to come from within myself" (line 440); ChatGPT's "generally super flat sentences" (line 462).
The framing has a clear tension with the broader vault apparatus on AI-as-creative-collaborator. The existing pages — Taste, Judgment, Labor Framework, VAST Voice Print Method, AI First Design Workflow — treat AI as a legitimate participant in the creative-production process, with the human providing taste, judgment, and aesthetic constraints while delegating substantial labor to the model. Wang's framework rejects this as a category error for voice-bearing writing: the labor of sentence-generation is constitutive of voice, and delegating it impoverishes the writer even when the output is fine.
The tension is not a contradiction. The two frameworks may both be correct for different writing forms. The taste-judgment-labor framework may apply well to writing where voice is not the central currency — corporate copy, marketing, technical documentation. Wang's framework may apply specifically to writing where voice is the central currency — essay, analysis, literary nonfiction, anywhere the writer's distinctive cognition is what readers are buying. The vault should hold both as live frames and apply them to different cases. The collision between them is a meaningful one and should be filed as such.
The deeper tension is over what production means. The taste-judgment-labor framework treats production as factorable into components, some of which can be delegated. Wang's framework treats production as a unitary cognitive activity that loses something specific when it is factored. Whether production is in fact factorable in the way the labor framework assumes is the live empirical question. The honest answer may be that production is factorable for some kinds of writing and unitary for others, and the writer's task is to know which kind they are doing.
Wang's framework converges with the Tyler-Cowen framing implicit in his own argument: AI as a tool that should arrive in a fully-developed cognitive life and turbocharge what is already there, rather than substituting for cognitive infrastructure that hasn't been built yet. The convergence is structural — both Wang and Cowen treat the user's pre-AI cognitive baseline as the input that determines whether AI use produces a multiplier or a substitute. The divergence is in the specific applications: Cowen has been more public about using AI in research and writing in ways Wang would not, suggesting that Cowen's verboten line may sit elsewhere or may not exist at all. The tension reveals that even within the same intellectual circle, the boundary between consumption-enhancement and production-substitution is drawn differently by different practitioners, and there is no consensus on where the right line is.
A more pointed tension is with the vault's existing AI-collaboration corpus, much of which was built around the assumption that AI can be a creative-production partner if used skillfully. Wang's framework is the strongest articulation in the vault of the opposite position: that AI cannot, by structural necessity, be a voice-production partner without degrading voice. The two positions are not equally weighted in the vault as it currently stands, and the addition of Wang's page corrects what may have been an over-weighting of the optimistic AI-as-collaborator framing. The vault now contains both positions and can hold them in productive tension.
The framework reaches into creative practice (the writing-craft implications), psychology (the cognitive mechanism by which sentence-generation produces thought), and cross-domain (the broader question of which cognitive activities are substitutable and which are not).
Creative Practice: Voice Cultivation Through Stylistic Models — the verboten rule and the voice-cultivation framework are companion pieces. Voice is built through deliberate exposure to master writers, manuscript copying, and trained intentionality at the sentence level (per the existing voice-cultivation pages). All of that work is exactly what the AI sentence-generation would substitute for. The handshake produces a clarifying claim: the AI verboten is not a freestanding stance; it is the necessary conservation of the work that voice-cultivation requires. A writer who follows the voice-cultivation discipline for years and then lets AI produce their sentences is undoing the cumulative result of their training. The two pages should be read as a unit: voice-cultivation is the construction; the verboten is the protection.
Eastern Spirituality: Sadhana Practice Hub — the discipline of not outsourcing the cognitive labor that the practice is for shows up across spiritual practice traditions. The bhakta who recites mantra cannot delegate the recitation to a recording without losing the practice's effect; the meditator cannot delegate the sitting to an algorithm without losing the meditation; the yogi cannot delegate the asana to a trainer-on-screen without losing the embodied transformation. The handshake reveals: certain cognitive activities are not factorable. The transformation produced by the activity is intrinsic to the doing of the activity, and there is no shortcut that preserves the transformation. Wang's writing case is one instance of a much broader principle that operates wherever embodied repetition produces cognitive change. The protection of the practice is the same logic in both cases: the verboten rule for AI in writing is the writing-domain analog of the spiritual injunction against shortcut-substitutes for sadhana.
Behavioral-Mechanics: Stage-Setting and Environmental Manipulation — the live-before-AI argument is a behavioral-mechanics claim about skill acquisition: the cognitive infrastructure built through pre-AI activity becomes the foundation that AI later multiplies, and skipping the foundation produces a different and lesser cognitive endpoint. The handshake produces an actionable claim for educators and parents: the developmental sequence matters. Children and young adults exposed to AI before they have developed reading, writing, library research, and database fluency may end up with a different cognitive architecture than those who develop the foundational skills first. The difference may not be visible in short-run task performance but may matter enormously for long-run cognitive depth. This is consistent with what teachers and parents anecdotally report and provides a structural framework for why.
Cross-Domain: labor-cognition substitutability (no dedicated vault page yet — flagged as gap; this is a broader epistemological question about which cognitive activities can be automated) — Wang's verboten line is one specific application of a broader epistemological question: which cognitive activities can be substituted by external systems without loss, and which cannot? The mainstream AI-optimist position is that most cognitive activity is substitutable as long as the human provides taste and judgment. The contrary position — to which Wang's framework belongs — holds that some cognitive activities are constitutive of the thinker rather than separable products of thought, and substituting them changes who the thinker is rather than merely what they produce. The handshake produces a long-run claim: the resolution of this debate matters not just for individual practice but for how societies structure education, work, and civic life around AI. If the optimist position is right, AI is a productivity multiplier and human cognition continues unchanged; if the verboten position is right, AI use without discipline can hollow out cognitive capacity at scale, and the cumulative effect is invisible until it is too late to reverse.
The four handshakes converge on a single claim: the boundary between AI-as-multiplier and AI-as-substitute is a real boundary, and where it falls determines whether AI use produces cognitive enrichment or cognitive impoverishment. The boundary is harder to maintain in practice than in principle, and the discipline of maintaining it — the verboten — is structurally similar to disciplines of practice protection across other domains.
The Sharpest Implication
If voice-bearing writing requires the verboten, then the AI optimist position — that skilled use of AI as creative collaborator can produce voice the way the writer would have on their own — is wrong in a specific way. It is not wrong about the output: AI-assisted writing can produce excellent prose. It is wrong about what the writer becomes through the process. The writer who has been letting AI produce sentences for two years does not have the same mind as the writer who has been producing them themselves. The output may look identical from outside; the cognitive endpoint of the writer is different. The implication is uncomfortable for any writer currently using AI in their drafting: even if your output is fine, you are different than you would have been, and the difference is not visible to you because the cognition you are not having is exactly the cognition that would have surfaced what you are not seeing. The verboten is therefore not a precaution against bad output; it is a precaution against an invisible degradation that the writer using AI cannot detect from inside the practice. The honest test of whether to maintain the verboten is whether you trust your future self to know what they have lost — and the trouble is that the future self produced by sustained AI-assisted writing may not know.
Generative Questions