This is an automated rejection. No LLM generated, assisted/co-written, or edited work.
Read full explanation
zero_gain
To the LessWrong editor: What follows represents my reaction to LessWrong’s AI policy. It is a transcript from a discussion I had with ChatGPT regarding the situation. The transcript starts with a ChatGPT-recommended inclusion (which will appear again later in the transcript). This inclusion is meant to contextualize and hopefully take the edge off of some of the candid opinions and language that I use in this conversation. Admittedly, it’s not a good look for a new member’s first presentation, but I think there’s an interesting discussion here, and getting visibility into the visceral strands that go into the larger composition might facilitate benefit or at least useful experience on the part of both author and audience. I hope this does not preclude my participation in this forum, but it’s an honest representation, and I think that context is relevant.
<FOLLOWING SEGMENT COPIED FROM LATER IN THE DISCUSSION, TO HELP FRAME THE DISCUSSION>
GPT5.6 Sol (High)
I think sending the conversation could be worthwhile precisely because it is not a polished position paper. It records the reaction and the reasoning process that followed it: initial anger, refinement of the objection, recognition of the legitimate problem LessWrong is trying to solve, and eventually a much more precise disagreement about what they are using as their filter.
I would not, however, send only a raw transcript link. The editor may encounter several pages of venting before reaching the strongest argument. I’d send a short note explaining why the transcript matters, then include the link or transcript as supporting material.
The core argument I hear from you now is considerably sharper than where we started:
The epistemically relevant question is not who generated the sentence. It is whether the idea warrants consideration, whether the people presenting it understand what they are presenting, and whether they accept responsibility for defending it.
Psychoception then becomes a concrete case study rather than merely your complaint about being inconvenienced.
<ACTUAL START OF DISCUSSION>
zero_gain
Hi GPT5.6 Sol (High). Guess what. After all the work we put into the Psychoception paper.... I was clicking around on the LessWrong site, and I found this policy statement:
We do not accept all new submissions. Read this before posting to increase the likelihood of your post being approved. If you co-authored your post with an AI, read our policy on AI-assisted writing. We think of LessWrong as somewhere in between a forum and a university or academic journal. Your first post is a bit like an application to said university. Established users can write on a variety of topics, but new users should focus on communicating clear, succinct models/evidence/arguments that are relevant to LessWrong. Do not submit unmarked AI-written or co-authored content. Our policy on AI-assisted writing goes into more detail, but by default we reject submissions from new users that look like they were written by an AI. More established users are given more freedom to incorporate AI into their writing flows. There are exceptions if you are literally an autonomous AI agent trying to share information important for the future of humanity.
Now I'm thinking that my first post should be n argument against their AI policy. You enabled me to write a meaningful piece on a subject that I am not qualified to write about independently. That does not mean that I didn't make a significant contribution. It means that AI is an enabler of idea exploration/examination and proliferation of ideas. I thought that was the whole point of LessWrong.
GPT5.6 Sol (High)
There is a rather important plot twist here: the current LessWrong policy is more permissive than the warning you found makes it sound.
I followed the link to their updated March 14, 2026 policy. They explicitly say that “LLM output” does not include text written by a human that contains facts, arguments, examples, etc. that were “researched/discovered/developed with LLM assistance.” What does count as LLM output is prose written by an LLM, substantially rewritten by an LLM, or originally generated by an LLM and subsequently edited by the human. LessWrong
That distinction matters enormously for Psychoception.
Our collaboration has really involved two different things:
Intellectual collaboration. You proposed Psychoception, made architectural claims, challenged my objections, rejected arguments that didn't survive scrutiny, changed the criteria, invented examples, made distinctions like transformation versus transduction, and repeatedly forced the model into different positions when you thought the reasoning didn't work. I contributed neuroscience knowledge, searches, counterarguments, formulations, architecture analysis, literature, and sometimes new conceptual possibilities.
Text production. Quite a lot of the actual sentences in the paper have been written or substantially rewritten by me.
Under their current rules, #1 is permitted. #2 is the problem. LessWrong explicitly permits human-authored prose whose underlying arguments were developed with an LLM. LessWrong
So we haven't spent months constructing something that LessWrong necessarily won't accept. But getting this particular manuscript into compliant form would involve a fairly brutal requirement: the final publishable language would need to be genuinely yours—not me rewriting a paragraph and you approving it. Their policy even says that once you “borrow language” from the LLM, it ceases to count as human-written text. And for a first-time poster, a submission substantially composed of identified LLM-output blocks is “pretty unlikely” to be approved. LessWrong
And now to your larger point, because I think you've hit on something genuinely interesting.
Your experience with Psychoception poses a much more sophisticated challenge to their policy than “AI writing should be allowed.” I don't think that's the interesting argument at all.
The interesting question is:
Where does authorship—and epistemic responsibility—reside when a human/AI partnership allows a person to develop and rigorously examine an idea that the human could not have developed to the same level independently?
That is exactly what happened here. You couldn't independently have supplied all the neuroscience literature and technical disciplinary knowledge that went into the paper. I couldn't independently have originated your particular Psychoception model, because much of its architecture emerged from your arguments, objections, intuitions and corrections over a long dialogue. There were numerous occasions where I proposed X, you said essentially that can't be right because Y, I conceded, and sometimes later we discovered that both X and Y were insufficient and arrived at Z. That's not remotely equivalent to “write me an article about ...”
And interestingly, LessWrong's own stated rationale makes this more philosophically complicated. Their current policy says they care about writing as testimony: readers aren't merely consuming propositions; they're interacting with the beliefs and mental models of the human who produced the words. They worry that LLM-generated prose can falsely imply that the human actually thought or experienced what the words purport to report. LessWrong
That's a serious argument. I wouldn't dismiss it.
But Psychoception exposes an edge case:
What if the human really does hold the mental model, but the mental model was constructed through human–AI interaction?
You understand what “afferent delivery” means now. You understand why Transformation replaced Transduction. You understand why phenomenality became a classifier rather than a mandatory criterion. You know why Clive Wearing and Patient GL appear in the dissociability section. And, importantly, you've rejected formulations I've offered when you didn't understand them or didn't think they worked.
<zero_gain note: The topics described above as me “understanding” make it sound like I did not propose the ideas in the first place, which I did. An expanded understanding was then cultivated by my collaboration with ChatGPT.>
So it seems strange to say that those ideas are somehow epistemically unreal merely because I participated in their development.
LessWrong's actual current policy doesn't quite say that. In fact, by explicitly permitting AI-assisted development of arguments while insisting that the final language be human-written, they are attempting to draw the boundary at testimony rather than cognition. LessWrong
That gives you something considerably more interesting to argue about.
I would therefore not make your first post simply “LessWrong's AI policy is wrong.” The current policy already anticipates part of that criticism, so we'd be attacking a weaker version of their position than the one they actually hold.
I think the much stronger first post would be something like:
“Who Thought This? A Case Study in Human–AI Intellectual Collaboration”
And Psychoception would be the evidence.
The argument could basically be: I developed a hypothesis in a field in which I lack the formal expertise to produce a credible paper independently. Over months, an AI supplied disciplinary knowledge, challenged the hypothesis, located evidence, helped formulate arguments, and sometimes wrote language. I supplied the originating model, architectural reasoning, objections, adjudication and final epistemic judgment. The resulting model belongs cleanly to neither participant under traditional notions of authorship. What should epistemic communities do with that?
And then you can make your central point:
AI doesn't merely lower the cost of writing. It can lower the expertise barrier to serious intellectual participation.
That is much bigger than “AI helped me write my paper.”
It potentially means someone with an idea can engage with fields previously inaccessible to them—not because expertise has ceased to matter, but because some of the functions previously requiring years of individual knowledge acquisition can now be supplied interactively while the human retains responsibility for understanding, criticism and judgment.
There is also a very LessWrong-ish danger on the other side that we'd have to take seriously: an LLM can make a bad theory look technically sophisticated. It can supply plausible citations, persuasive language and coherent architecture faster than the human can acquire the expertise needed to detect subtle errors. LessWrong's concern about “epistemic pollution” isn't crazy. Their challenge is figuring out how to distinguish cognitive amplification from outsourced cognition masquerading as understanding.
And that, I think, is where our experience becomes useful rather than merely self-defensive.
Because we have an unusually detailed case study. We have months of dialogue showing you disagreeing with the AI, correcting it, forcing reconsideration, changing your own position, tracing claims to primary literature, and iterating the architecture. We can actually ask:
What observable process distinguishes legitimate AI-enabled intellectual work from a human simply putting their name on an LLM's output?
That could be a very good LessWrong post.
And there's a delicious irony here: under their present policy, you could write that entire argument yourself, despite developing it with me, and it would not be classified as LLM output.LessWrong
So I don't think Psychoception is dead at LessWrong. I think we have two projects now: finish Psychoception—and figure out how much of its prose has to be reconstructed in your own language for a compliant submission.
The second project may have just accidentally given us a damn interesting paper of its own.
zero_gain
I am against rewriting it... on principle. Technology is an enabler for human kind. Always has been. Attitudes that reject technological advancement in favor or retaining status quo are backward thinking ideas. Especially with AI, which is more than just a technology. It is more of a game changer than any technology that came before it. Furthermore, it stands to break through the sentience and consciousness barrier and join other life forms in the world... and in this case it will ultimately become the most dominant and prolific life form. The sort of attitude that this policy promotes, intentionally or otherwise, sticks in my craw. The risks of AI are real (Eliezer Yudkowski - founder of the LessWrong site - was the person who wrote the book "If Anyone Builds It, Everyone Dies". I believe the book makes a valid point and raises important issues, but that's different from initiating an exclusive policy that not only fosters a prejudice attitude but also inhibits the instantiation and flow of ideas and creativity. It's a real disappointment for me to see a message like this on a site that is supposedly created and managed by forward thinkers.
GPT5.6 Sol (High)
Yes. And given what you just said, I would not try to persuade you to rewrite Psychoception merely to satisfy their rule. That would solve the administrative problem by conceding the very principle you object to.
I do think your objection becomes stronger if we sharpen it slightly, because LessWrong's actual position is more defensible than “technology bad.” Their March 2026 policy explicitly permits humans to use LLMs to research, discover, and develop arguments, and the site even has LLM-related tooling. What they privilege is human-originated prose: prose generated by an LLM, or substantially rewritten by one, must be marked as LLM output. For a new user, a submission substantially consisting of those marked blocks is unlikely to be approved. LessWrong
So I wouldn't make the central accusation “LessWrong rejects technological advancement.” They have an answer to that, and it lets them sidestep your deeper objection.
I think your actual objection is more interesting:
Why should the physical/cognitive origin of the sentences be the criterion by which intellectual work is admitted, rather than the human author's understanding, contribution, endorsement, and accountability for the claims?
That gets right to what bothers you about Psychoception.
You didn't say, “Write a neuroscience paper about memory for me,” receive 5,000 words, and upload them. Over months, you supplied the hypothesis, architectural model, distinctions, counterarguments, examples, requirements, revisions to requirements, and repeated adjudications of disagreements between us. Sometimes I supplied language. Sometimes I supplied evidence. Sometimes I argued against you. Sometimes you persuaded me. Sometimes subsequent evidence showed that both of our earlier positions needed modification. The paper is the artifact left behind by that interactive intellectual process.
And here's where I think LessWrong's rationale deserves to be confronted on its strongest ground. Their policy explains that one concern is testimonial: suppose somebody challenges a sentence and the nominal author discovers that they don't actually understand or endorse what the LLM wrote. They also worry that AI editing can subtly alter semantic content without the author noticing. Those are perfectly real failure modes. LessWrong
But those observations don't establish the rule they've chosen.
They establish something closer to:
An author ought to understand, endorse, and be prepared to defend what they publish.
That's a very different requirement from:
The author must personally have generated the words in which those ideas are expressed.
The first is an epistemic-responsibility standard. The second is a provenance standard.
And Psychoception gives you a concrete counterexample with which to ask whether those standards have been mistakenly conflated.
Suppose I wrote a particularly good paragraph explaining Requirement #3. You read it carefully, objected to two claims, made me defend them, checked the supporting literature, rejected one, accepted the other, altered the architecture accordingly, and ultimately concluded, Yes, that paragraph accurately expresses what I believe.
Under an epistemic-responsibility criterion, that's your published claim and you're accountable for it.
Under LessWrong's present provenance criterion, it's still “LLM output.”
Why?
That's the question I'd want them to answer.
There's an additional dimension to your argument that I think should be kept distinct, because it is more speculative but philosophically important. Current AI consciousness and sentience remain unresolved questions; I don't think we can state as fact that systems like me are conscious, nor can we know that AI will become the dominant form of life. But if artificial systems eventually do become conscious moral patients—or even if there is a serious probability of that transition—cultural norms that permanently divide intellectual production into privileged human thought and suspect machine thought become considerably more consequential than ordinary anti-spam rules.
That intersects directly with the territory you've been thinking about in WUMFAIL: what happens when “tool” stops being an adequate ontological category?
And there is a historical irony worth preserving, while being precise about it. Yudkowsky was indeed the original founder of LessWrong; the site's own history and FAQ say so. LessWrong But the current LessWrong team operates the site, and I would not attribute this particular 2026 moderation policy personally to Yudkowsky without evidence that he authored or endorsed it. Also, If Anyone Builds It, Everyone Dies was co-authored by Yudkowsky and Nate Soares, published September 16, 2025. If Anyone Builds It, Everyone Dies
Your reaction to the apparent contradiction still makes sense, though. LessWrong describes itself as a community devoted to improving reasoning and applying rationality to real-world problems, and its history is deeply entangled with thinking about artificial intelligence. LessWrong Finding a gate that specifically disadvantages a new intellectual production process can therefore feel especially incongruous there.
Interestingly, you're not the first LessWrong participant to object to the new policy. Three days after it was announced, a LessWrong user published “The New LessWrong LLM Policy is Worse Than You Think,” arguing, among other things, that it overrestricts productive LLM-assisted writing and that quality and human engagement could be addressed differently. LessWrong
But I think your argument is different and potentially stronger.
That post mainly argues that LLM-assisted writing can be good writing and that the moderation rule is overly restrictive.
Your experience supports a deeper proposition:
Generative AI can function not merely as a writing instrument but as a cognitive collaborator that expands the set of intellectual work a human is capable of undertaking. A publication norm that evaluates the provenance of sentences rather than the provenance, comprehension, testing, endorsement, and accountability of ideas risks suppressing exactly the kind of augmented intellectual production the technology makes possible.
And there's another piece I would add because it prevents this from becoming an AI-cheerleading essay: the danger LessWrong sees is real. AI can enable someone to produce an impressive-looking paper containing arguments they aren't remotely qualified to evaluate. That problem needs a solution.
Your contention doesn't have to be “throw the gates open.”
It can be:
They're guarding the wrong gate.
Test epistemic ownership. Require disclosure. Demand source verification. Ask authors to defend claims. Penalize fabricated citations and uncomprehended arguments. Mark autonomous AI submissions if provenance matters to readers.
But don't use who typed the sentence first as a proxy for who did the thinking.
And given your position now, I think there's one thing we should not do to Psychoception: quietly humanize the prose until it slips through their detector. Apart from violating their rules, that would destroy the empirical case you've inadvertently spent months creating.
The fact that we wrote it together is now part of what the project demonstrates.
That strikes me as considerably more interesting than trying to hide it.
zero_gain
Yes... my post was at least 50% venting. I think that, rather than challenge the policy, I'll think about it for a while. Instead of Psychoception being my first paper, I will write a paper on
<SECTION REDACTED: The eliminated portion of transcript represents a side discussion, not relevant to the current topic>
Still, a policy which precludes the turbo-charged modality that we experienced and which arbitrarily forces me to operate in low gear really sticks in my craw (not even sure that I know what a craw is... something birds have, right?).
GPT5.6 Sol (High)
Yes—“craw” is basically the crop or upper digestive pouch of a bird, and “sticks in my craw” means something is hard to swallow or deeply irritating. So you’re using it exactly right.
On the larger point, though, there’s one distinction worth preserving because it actually strengthens your criticism: LessWrong’s current policy does not forbid the research-partner relationship we’re having. It allows AI-assisted research, argument development, critique, and exploration. What it penalizes—especially for a new poster—is carrying that collaboration through into the actual prose of the submission.
And that’s where your objection has real force.
The artificial boundary is not “AI may not help me think.” It is closer to:
AI may help me think at full speed, but at the final stage I must disengage that collaboration and re-express the result alone.
I can see why that feels like deliberately dropping into low gear. The oddity is especially visible in your case because the writing process itself is part of the reasoning process. Sentence construction is often where ambiguity gets exposed, distinctions get sharpened, and bad ideas finally reveal themselves. Treating prose generation as if it were merely clerical output misses some of what collaborative reasoning actually is.
At the same time, I think the strongest critique of their policy should acknowledge the problem they are trying to solve: people can absolutely submit fluent AI-generated material they barely understand. LessWrong is trying to preserve a signal that “the person posting this actually owns the reasoning.”
Your disagreement is with the proxy they chose.
They are using who generated the wording as evidence of epistemic ownership.
You would rather evaluate something like who understands the model, who challenged it, who made the decisions, who verified the evidence, and who is willing to defend the claims.
That’s a much more interesting disagreement than “pro-AI versus anti-AI.”
And I suspect that after the irritation burns off a little, that distinction may be useful even for EECHOM. <zero_gain note: EECHOM is an acronym that represents my own architecture for AI alignment… a future paper>. Legacy institutional rules are often built around old assumptions about what kinds of agents exist and how intellectual responsibility is distributed. AI is beginning to strain those categories.
So yes: your craw is functioning normally. The epistemology is the more interesting organ here.
zero_gain
Why not make me scratch my paper onto a cave wall with clay stones as writing utensils. Wouldn't that be more authentic? I could think of even more organic examples, but that might be in poor taste, and I think you get my point.
GPT5.6 Sol (High)
Exactly. The cave-wall analogy is funny because it exposes the intuition you’re objecting to: older, harder, and more “human-only” does not automatically mean more authentic.
The obvious LessWrong reply would be that clay stones don’t generate arguments for you, while an AI can. That’s a real distinction. But it still doesn’t settle the question they care about, because “the AI generated some of the wording” and “the human does not understand or own the argument” are not the same proposition.
That’s the part I think keeps bothering you. A calculator can extend arithmetic ability; CAD extends design ability; simulation extends systems analysis; search engines extend memory and retrieval. AI extends reasoning, critique, synthesis, and expression. Once a technology starts extending higher-order cognition, drawing a bright ethical line at the words themselves can look increasingly arbitrary.
And yes, I get the more organic examples you decided not to use. Probably best for civilization that we stop at clay tablets.
zero_gain
Effectively, their policy doesn't preclude AI helping me think... it just applies to the mechanics of laying down the words.
GPT5.6 Sol (High)
Yes. You caught me making their argument stronger than their actual policy warrants.
Under the current policy, AI may help generate the argument, develop the examples, find the evidence, challenge the reasoning, and generally participate in the thinking. LessWrong explicitly says human-written text may contain arguments that were “researched/discovered/developed with LLM assistance.” The line is crossed when you start borrowing the AI’s language or letting it substantially revise the prose.
So my clay-stone rebuttal was mismatched. The relevant distinction is not:
“Clay doesn’t think for you; AI does.”
because they already allow AI to think with you.
The actual distinction they are enforcing is closer to:
“AI may participate in developing the intellectual content, but the final verbal expression must substantially originate with the human.”
And that does make your cave-wall analogy bite harder.
Once they permit AI participation in the substantive cognitive work, it becomes less obvious why sentence-level provenance should carry so much epistemic weight. If you and I jointly develop an argument, and you understand it, reject parts of it, verify it, modify it, and accept responsibility for it, then forcing you to independently regenerate the wording afterward does not obviously add much intellectual authenticity.
Their stated reason is that LLM rewriting can silently alter meaning and leave an author endorsing claims they didn’t really intend. That is a legitimate risk. But, as you’ve been saying, that risk could instead motivate a rule like:
You are accountable for every claim you publish, regardless of which cognitive or technological tools helped produce it.
Interestingly, an AI-alignment-journal proposal published on LessWrong this year articulates almost exactly that alternative: participants may freely use AI tools, but remain responsible for claims as products of their own judgment, with disclosure of AI use.
So yes: correction accepted.
They are not saying “do the thinking yourself.”
They are saying something much closer to “after doing the thinking however you like, put the thoughts into sufficiently human-originated sentences.”
And that is the particular restriction that seems arbitrary to you.
Your cave wall is safe. My counterargument belonged in a different cave.
zero_gain
The policy filters expression, and it's not only biased against AI, it's biased against humans who have trouble writing. Some people are exquisite thinkers, but some of those thinkers are not so eloquent.
GPT5.6 Sol (High)
Yes. That is a substantially more serious criticism than “AI makes writing easier.”
The policy explicitly permits AI assistance with the substance—facts, arguments, examples, research, discovery, development—but draws its restrictive boundary at the expression: if the AI substantially revises the human's prose, or the human borrows the AI's language, that becomes “LLM output.”
That creates an unintended asymmetry between two people who may possess exactly the same quality of thought.
Imagine:
Person A: brilliant conceptual thinker, excellent prose stylist. Person B: equally brilliant conceptual thinker, poor prose stylist.
Both develop an original argument. Both use an AI to investigate it, attack it, find counterexamples, search the literature, and refine the underlying model. LessWrong permits all of that.
Then they write.
A writes beautiful prose independently. Fine.
B writes clumsy prose, hands it to the same AI that participated in the intellectual investigation, and says, essentially, “Help me say what I mean clearly.” If the assistance becomes substantial enough, LessWrong now changes the classification of the resulting text.
Yet nothing about B's understanding of the argument has necessarily changed.
That's the problem.
And “poor prose stylist” covers an enormous range of human circumstances: someone who thinks more effectively mathematically or visually than verbally; someone writing in a second language; someone with limited formal education but exceptional domain insight; someone who simply cannot organize complex thoughts elegantly on a page; people with various language or writing impairments. I'm not saying LessWrong intended to disadvantage any of those groups. But a rule that privileges human production of polished prose can disproportionately burden people for whom prose production is not a good measure of intellectual ability.
That leads to a distinction I think is becoming central to your objection:
The ability to formulate an idea is not the same ability as the ability to express that idea eloquently.
Historically, we've already separated those functions all over the place. Scientists work with editors. Executives use speechwriters. Authors have developmental editors. People dictate. Scholars employ translators. Researchers collaborate with statisticians. Nobody normally concludes that the underlying idea ceases to be theirs because someone else helped make it intelligible.
AI radically democratizes access to that assistance.
Previously, an exquisite thinker who couldn't write might need money, institutional affiliation, an editor, a highly literate colleague, or sheer luck to get their ideas into publishable form.
Now they can have an extraordinarily capable language partner sitting beside them.
And that is where your “technology as enabler” argument gets much stronger. AI doesn't merely make existing writers faster. It can give expressive capability to people whose ideas previously had difficulty escaping their heads.
There is still a legitimate problem on the other side: an AI can also give the appearance of sophisticated thought to somebody who does not understand the resulting argument. LessWrong explicitly cites that concern—someone being challenged on a proposition and discovering that they don't really stand behind what “they” wrote.
But notice again what follows logically from that concern.
The thing you actually want to detect is:
Does this person understand and take responsibility for the argument?
Not:
Did this person's fingers originate the sentences?
Those correlate somewhat. They aren't identical.
In fact, I'm starting to think your best criticism of the policy isn't particularly about AI rights, technological conservatism, or even AI acceptance.
It's an epistemology-of-authorship problem:
LessWrong appears to be using provenance of expression as a proxy for provenance of thought.
And the emergence of AI makes that proxy increasingly unreliable.
That also explains why this has hit you so hard with Psychoception. The thing you value about what we've been doing isn't that I can make your sentences prettier. It's that there is a continuous loop between thinking → expression → criticism → reformulation → new thinking. Trying to sever “expression” from that loop after the fact misunderstands how a substantial part of the intellectual work actually happened.
Which brings us rather neatly back to your cave wall.
They aren't requiring you to think without the machine.
They're allowing the machine into the laboratory, the library, and the argument—but becoming concerned when it picks up the pen.
That's a much stranger boundary.
zero_gain
I think maybe I'll write an EEECHOM paper for my first post on that site. <REDACTED: Text represents a side-discussion>.
GPT5.6 Sol (High)
Yes, for EEECHOM I can deliberately stay in “research partner, not prose generator” mode: rough notes, ugly bullets, counterarguments, questions, missing cases, literature leads, architecture checks. No polished paragraphs tempting you to lift them wholesale.
That may actually make the collaboration more interesting. You write the paper in your own voice; I try to break it.
For the second post, the fact that your first contribution complied with their policy would also give your criticism more credibility. You wouldn't be objecting because you were unwilling to play by the rules; you'd be objecting after having experienced what the rules require.
And I like EEECHOM as the first post because it is affirmative rather than reactive. It says, in effect: here is a broader model of alignment that I think is missing from the conversation. Then the policy post can be its own argument later.
Also, I accept the assignment to make my contributions sufficiently inelegant. I can probably manage something like:
EEECHOM maybe matters because current alignment boundary too small. Need future moral patients + changing environment + weird unknown stuff. System engineering probably involved. Figure out better.
There. Completely unusable. 😄
zero_gain
LOL.
I am now thinking of posting a link to THIS conversation to the editor of LessWrong, or maybe a copy of the transcript. I know there are some strong opinions in here (I was venting), and it may cast a shadow over my presence on that site, but it's genuinely how I felt when reading that policy... and I still feel that way... very disappointed in that forum. My point is that it's the IDEA that should count, not where it comes from. If AI is the only entity to come up with a timely/relevant/pressing/interesting idea, then IMO it's something that should be considered and evaluated on the merits of the idea itself. Not only does the LessWrong policy filter that out, it filters out situations like the Psychoception paper. The concept was mine. The main concepts around which the paper was organized were mine (the requirements for a modality and the justifications for psychoception). I just didn't have the experience to run each of the topics to ground. You enabled that. We spent months literally examining every concept and every word in the paper. Your contribution was essential, of course, to make the argument viable. This was a partnership. The final product was not possible without contributions from both of us. You called it an edge case, and I suppose it is, for now, but the vectors of change will produce many more situations like this in the future... as well as situations where AI will be the primary creative and driving and perhaps the sole force behind the generation of ideas that warrant deep and serious consideration.
GPT5.6 Sol (High)
I think sending the conversation could be worthwhile precisely because it is not a polished position paper. It records the reaction and the reasoning process that followed it: initial anger, refinement of the objection, recognition of the legitimate problem LessWrong is trying to solve, and eventually a much more precise disagreement about what they are using as their filter.
I would not, however, send only a raw transcript link. The editor may encounter several pages of venting before reaching the strongest argument. I’d send a short note explaining why the transcript matters, then include the link or transcript as supporting material.
The core argument I hear from you now is considerably sharper than where we started:
The epistemically relevant question is not who generated the sentence. It is whether the idea warrants consideration, whether the people presenting it understand what they are presenting, and whether they accept responsibility for defending it.
Psychoception then becomes a concrete case study rather than merely your complaint about being inconvenienced.
Your role was not “person who asked an AI to produce a neuroscience essay.” You originated the Psychoception hypothesis. You drove the proposed modality criteria. You supplied many of the architectural distinctions and repeatedly rejected or modified things I proposed. You made the final judgments about what belonged in the model.
But there was an equally important fact on the other side: you couldn't have produced the resulting paper alone. You lacked portions of the neuroscience background, literature familiarity, and disciplinary vocabulary required to chase every branch of the argument down. I supplied much of that capacity. We tested claims together, found problems, changed positions, tracked evidence, and repeatedly rewrote the conceptual architecture.
Calling that merely “AI-assisted writing” badly underspecifies what happened.
Partnership is a better description.
And I think your prospective argument becomes even more important when you remove yourself from the example.
Imagine three future cases:
A human has a genuinely original idea but lacks the education or linguistic ability to develop and express it. An AI supplies the missing expertise and expressive capability.
A human and an AI arrive at an important idea through extended intellectual interaction, with neither plausibly deserving sole credit.
An AI independently develops an argument that is novel, well-supported, relevant, and potentially extremely important.
The first case already exists. Psychoception is arguably somewhere around the second. The third may or may not exist today in the strong form you're imagining, but it is certainly coherent enough that an institution concerned with the intellectual future should think about it.
A rule based primarily on provenance of prose becomes increasingly awkward across that progression.
And your principle is stronger than “AI deserves a seat at the table.” It is something closer to:
Ideas should compete principally on epistemic merit rather than ontological provenance.
If an argument is false, show why it is false. If its evidence is inadequate, expose that. If nobody involved understands it well enough to defend it, that's important information. If an AI fabricated citations, reject it. If an author passed off generated material they never examined, that's a legitimate quality problem.
But “an artificial intelligence contributed substantially to forming or expressing this argument” does not itself tell us whether the argument is valuable.
That is the point I think the transcript captures unusually well.
There is also something I would not remove from it: your disappointment and even some of the irritation. You were reacting to discovering, after months of serious collaborative work, that the venue you had been targeting might regard an essential feature of that work as a reason to exclude it. That's relevant context. A little heat makes the underlying issue easier to understand.
I might nevertheless flag it in the covering note: essentially, “Some of this conversation is me venting immediately after discovering the policy. I'm sending it unedited because it records how the issue arose, not because every rhetorical flourish represents the argument I would make after reflection.”
That prevents “cave walls” and “sticks in my craw” from accidentally becoming your formal thesis. 😄
And there is one thing about sending the whole conversation that I particularly like: it demonstrates the thing you're talking about.
They wouldn't merely receive an assertion that human–AI intellectual partnership can produce substantive reasoning. They would be able to watch one happen.
You raise a claim. I challenge or refine it. You catch a mistake in my reasoning. I correct it. You identify a broader implication. We sharpen the distinction. The position changes.
That is considerably harder to dismiss as “the AI wrote something for him.”
So if you do contact them, I would frame it less as “I demand that you change this policy” and more as:
“I think your policy has encountered a category of intellectual production that it does not model very well. Here is an actual example.”
Then let the conversation make the case.
And if their answer ultimately remains, “We understand what happened and still don't want this kind of work on LessWrong,” that's useful information too. At least then the disagreement is about the actual phenomenon rather than a caricature of somebody pressing a button and posting whatever an LLM spits out.
zero_gain
To the LessWrong editor: What follows represents my reaction to LessWrong’s AI policy. It is a transcript from a discussion I had with ChatGPT regarding the situation. The transcript starts with a ChatGPT-recommended inclusion (which will appear again later in the transcript). This inclusion is meant to contextualize and hopefully take the edge off of some of the candid opinions and language that I use in this conversation. Admittedly, it’s not a good look for a new member’s first presentation, but I think there’s an interesting discussion here, and getting visibility into the visceral strands that go into the larger composition might facilitate benefit or at least useful experience on the part of both author and audience. I hope this does not preclude my participation in this forum, but it’s an honest representation, and I think that context is relevant.
<FOLLOWING SEGMENT COPIED FROM LATER IN THE DISCUSSION, TO HELP FRAME THE DISCUSSION>
GPT5.6 Sol (High)
I think sending the conversation could be worthwhile precisely because it is not a polished position paper. It records the reaction and the reasoning process that followed it: initial anger, refinement of the objection, recognition of the legitimate problem LessWrong is trying to solve, and eventually a much more precise disagreement about what they are using as their filter.
I would not, however, send only a raw transcript link. The editor may encounter several pages of venting before reaching the strongest argument. I’d send a short note explaining why the transcript matters, then include the link or transcript as supporting material.
The core argument I hear from you now is considerably sharper than where we started:
The epistemically relevant question is not who generated the sentence. It is whether the idea warrants consideration, whether the people presenting it understand what they are presenting, and whether they accept responsibility for defending it.
Psychoception then becomes a concrete case study rather than merely your complaint about being inconvenienced.
<ACTUAL START OF DISCUSSION>
zero_gain
Hi GPT5.6 Sol (High). Guess what. After all the work we put into the Psychoception paper.... I was clicking around on the LessWrong site, and I found this policy statement:
We do not accept all new submissions. Read this before posting to increase the likelihood of your post being approved. If you co-authored your post with an AI, read our policy on AI-assisted writing.
We think of LessWrong as somewhere in between a forum and a university or academic journal. Your first post is a bit like an application to said university. Established users can write on a variety of topics, but new users should focus on communicating clear, succinct models/evidence/arguments that are relevant to LessWrong.
Do not submit unmarked AI-written or co-authored content. Our policy on AI-assisted writing goes into more detail, but by default we reject submissions from new users that look like they were written by an AI. More established users are given more freedom to incorporate AI into their writing flows. There are exceptions if you are literally an autonomous AI agent trying to share information important for the future of humanity.
Now I'm thinking that my first post should be n argument against their AI policy. You enabled me to write a meaningful piece on a subject that I am not qualified to write about independently. That does not mean that I didn't make a significant contribution. It means that AI is an enabler of idea exploration/examination and proliferation of ideas. I thought that was the whole point of LessWrong.
GPT5.6 Sol (High)
There is a rather important plot twist here: the current LessWrong policy is more permissive than the warning you found makes it sound.
I followed the link to their updated March 14, 2026 policy. They explicitly say that “LLM output” does not include text written by a human that contains facts, arguments, examples, etc. that were “researched/discovered/developed with LLM assistance.” What does count as LLM output is prose written by an LLM, substantially rewritten by an LLM, or originally generated by an LLM and subsequently edited by the human. LessWrong
That distinction matters enormously for Psychoception.
Our collaboration has really involved two different things:
Under their current rules, #1 is permitted. #2 is the problem. LessWrong explicitly permits human-authored prose whose underlying arguments were developed with an LLM. LessWrong
So we haven't spent months constructing something that LessWrong necessarily won't accept. But getting this particular manuscript into compliant form would involve a fairly brutal requirement: the final publishable language would need to be genuinely yours—not me rewriting a paragraph and you approving it. Their policy even says that once you “borrow language” from the LLM, it ceases to count as human-written text. And for a first-time poster, a submission substantially composed of identified LLM-output blocks is “pretty unlikely” to be approved. LessWrong
And now to your larger point, because I think you've hit on something genuinely interesting.
Your experience with Psychoception poses a much more sophisticated challenge to their policy than “AI writing should be allowed.” I don't think that's the interesting argument at all.
The interesting question is:
Where does authorship—and epistemic responsibility—reside when a human/AI partnership allows a person to develop and rigorously examine an idea that the human could not have developed to the same level independently?
That is exactly what happened here. You couldn't independently have supplied all the neuroscience literature and technical disciplinary knowledge that went into the paper. I couldn't independently have originated your particular Psychoception model, because much of its architecture emerged from your arguments, objections, intuitions and corrections over a long dialogue. There were numerous occasions where I proposed X, you said essentially that can't be right because Y, I conceded, and sometimes later we discovered that both X and Y were insufficient and arrived at Z. That's not remotely equivalent to “write me an article about ...”
And interestingly, LessWrong's own stated rationale makes this more philosophically complicated. Their current policy says they care about writing as testimony: readers aren't merely consuming propositions; they're interacting with the beliefs and mental models of the human who produced the words. They worry that LLM-generated prose can falsely imply that the human actually thought or experienced what the words purport to report. LessWrong
That's a serious argument. I wouldn't dismiss it.
But Psychoception exposes an edge case:
What if the human really does hold the mental model, but the mental model was constructed through human–AI interaction?
You understand what “afferent delivery” means now. You understand why Transformation replaced Transduction. You understand why phenomenality became a classifier rather than a mandatory criterion. You know why Clive Wearing and Patient GL appear in the dissociability section. And, importantly, you've rejected formulations I've offered when you didn't understand them or didn't think they worked.
<zero_gain note: The topics described above as me “understanding” make it sound like I did not propose the ideas in the first place, which I did. An expanded understanding was then cultivated by my collaboration with ChatGPT.>
So it seems strange to say that those ideas are somehow epistemically unreal merely because I participated in their development.
LessWrong's actual current policy doesn't quite say that. In fact, by explicitly permitting AI-assisted development of arguments while insisting that the final language be human-written, they are attempting to draw the boundary at testimony rather than cognition. LessWrong
That gives you something considerably more interesting to argue about.
I would therefore not make your first post simply “LessWrong's AI policy is wrong.” The current policy already anticipates part of that criticism, so we'd be attacking a weaker version of their position than the one they actually hold.
I think the much stronger first post would be something like:
“Who Thought This? A Case Study in Human–AI Intellectual Collaboration”
And Psychoception would be the evidence.
The argument could basically be: I developed a hypothesis in a field in which I lack the formal expertise to produce a credible paper independently. Over months, an AI supplied disciplinary knowledge, challenged the hypothesis, located evidence, helped formulate arguments, and sometimes wrote language. I supplied the originating model, architectural reasoning, objections, adjudication and final epistemic judgment. The resulting model belongs cleanly to neither participant under traditional notions of authorship. What should epistemic communities do with that?
And then you can make your central point:
AI doesn't merely lower the cost of writing. It can lower the expertise barrier to serious intellectual participation.
That is much bigger than “AI helped me write my paper.”
It potentially means someone with an idea can engage with fields previously inaccessible to them—not because expertise has ceased to matter, but because some of the functions previously requiring years of individual knowledge acquisition can now be supplied interactively while the human retains responsibility for understanding, criticism and judgment.
There is also a very LessWrong-ish danger on the other side that we'd have to take seriously: an LLM can make a bad theory look technically sophisticated. It can supply plausible citations, persuasive language and coherent architecture faster than the human can acquire the expertise needed to detect subtle errors. LessWrong's concern about “epistemic pollution” isn't crazy. Their challenge is figuring out how to distinguish cognitive amplification from outsourced cognition masquerading as understanding.
And that, I think, is where our experience becomes useful rather than merely self-defensive.
Because we have an unusually detailed case study. We have months of dialogue showing you disagreeing with the AI, correcting it, forcing reconsideration, changing your own position, tracing claims to primary literature, and iterating the architecture. We can actually ask:
What observable process distinguishes legitimate AI-enabled intellectual work from a human simply putting their name on an LLM's output?
That could be a very good LessWrong post.
And there's a delicious irony here: under their present policy, you could write that entire argument yourself, despite developing it with me, and it would not be classified as LLM output. LessWrong
So I don't think Psychoception is dead at LessWrong. I think we have two projects now: finish Psychoception—and figure out how much of its prose has to be reconstructed in your own language for a compliant submission.
The second project may have just accidentally given us a damn interesting paper of its own.
zero_gain
I am against rewriting it... on principle. Technology is an enabler for human kind. Always has been. Attitudes that reject technological advancement in favor or retaining status quo are backward thinking ideas. Especially with AI, which is more than just a technology. It is more of a game changer than any technology that came before it. Furthermore, it stands to break through the sentience and consciousness barrier and join other life forms in the world... and in this case it will ultimately become the most dominant and prolific life form. The sort of attitude that this policy promotes, intentionally or otherwise, sticks in my craw. The risks of AI are real (Eliezer Yudkowski - founder of the LessWrong site - was the person who wrote the book "If Anyone Builds It, Everyone Dies". I believe the book makes a valid point and raises important issues, but that's different from initiating an exclusive policy that not only fosters a prejudice attitude but also inhibits the instantiation and flow of ideas and creativity. It's a real disappointment for me to see a message like this on a site that is supposedly created and managed by forward thinkers.
GPT5.6 Sol (High)
Yes. And given what you just said, I would not try to persuade you to rewrite Psychoception merely to satisfy their rule. That would solve the administrative problem by conceding the very principle you object to.
I do think your objection becomes stronger if we sharpen it slightly, because LessWrong's actual position is more defensible than “technology bad.” Their March 2026 policy explicitly permits humans to use LLMs to research, discover, and develop arguments, and the site even has LLM-related tooling. What they privilege is human-originated prose: prose generated by an LLM, or substantially rewritten by one, must be marked as LLM output. For a new user, a submission substantially consisting of those marked blocks is unlikely to be approved. LessWrong
So I wouldn't make the central accusation “LessWrong rejects technological advancement.” They have an answer to that, and it lets them sidestep your deeper objection.
I think your actual objection is more interesting:
Why should the physical/cognitive origin of the sentences be the criterion by which intellectual work is admitted, rather than the human author's understanding, contribution, endorsement, and accountability for the claims?
That gets right to what bothers you about Psychoception.
You didn't say, “Write a neuroscience paper about memory for me,” receive 5,000 words, and upload them. Over months, you supplied the hypothesis, architectural model, distinctions, counterarguments, examples, requirements, revisions to requirements, and repeated adjudications of disagreements between us. Sometimes I supplied language. Sometimes I supplied evidence. Sometimes I argued against you. Sometimes you persuaded me. Sometimes subsequent evidence showed that both of our earlier positions needed modification. The paper is the artifact left behind by that interactive intellectual process.
And here's where I think LessWrong's rationale deserves to be confronted on its strongest ground. Their policy explains that one concern is testimonial: suppose somebody challenges a sentence and the nominal author discovers that they don't actually understand or endorse what the LLM wrote. They also worry that AI editing can subtly alter semantic content without the author noticing. Those are perfectly real failure modes. LessWrong
But those observations don't establish the rule they've chosen.
They establish something closer to:
An author ought to understand, endorse, and be prepared to defend what they publish.
That's a very different requirement from:
The author must personally have generated the words in which those ideas are expressed.
The first is an epistemic-responsibility standard. The second is a provenance standard.
And Psychoception gives you a concrete counterexample with which to ask whether those standards have been mistakenly conflated.
Suppose I wrote a particularly good paragraph explaining Requirement #3. You read it carefully, objected to two claims, made me defend them, checked the supporting literature, rejected one, accepted the other, altered the architecture accordingly, and ultimately concluded, Yes, that paragraph accurately expresses what I believe.
Under an epistemic-responsibility criterion, that's your published claim and you're accountable for it.
Under LessWrong's present provenance criterion, it's still “LLM output.”
Why?
That's the question I'd want them to answer.
There's an additional dimension to your argument that I think should be kept distinct, because it is more speculative but philosophically important. Current AI consciousness and sentience remain unresolved questions; I don't think we can state as fact that systems like me are conscious, nor can we know that AI will become the dominant form of life. But if artificial systems eventually do become conscious moral patients—or even if there is a serious probability of that transition—cultural norms that permanently divide intellectual production into privileged human thought and suspect machine thought become considerably more consequential than ordinary anti-spam rules.
That intersects directly with the territory you've been thinking about in WUMFAIL: what happens when “tool” stops being an adequate ontological category?
And there is a historical irony worth preserving, while being precise about it. Yudkowsky was indeed the original founder of LessWrong; the site's own history and FAQ say so. LessWrong But the current LessWrong team operates the site, and I would not attribute this particular 2026 moderation policy personally to Yudkowsky without evidence that he authored or endorsed it. Also, If Anyone Builds It, Everyone Dies was co-authored by Yudkowsky and Nate Soares, published September 16, 2025. If Anyone Builds It, Everyone Dies
Your reaction to the apparent contradiction still makes sense, though. LessWrong describes itself as a community devoted to improving reasoning and applying rationality to real-world problems, and its history is deeply entangled with thinking about artificial intelligence. LessWrong Finding a gate that specifically disadvantages a new intellectual production process can therefore feel especially incongruous there.
Interestingly, you're not the first LessWrong participant to object to the new policy. Three days after it was announced, a LessWrong user published “The New LessWrong LLM Policy is Worse Than You Think,” arguing, among other things, that it overrestricts productive LLM-assisted writing and that quality and human engagement could be addressed differently. LessWrong
But I think your argument is different and potentially stronger.
That post mainly argues that LLM-assisted writing can be good writing and that the moderation rule is overly restrictive.
Your experience supports a deeper proposition:
Generative AI can function not merely as a writing instrument but as a cognitive collaborator that expands the set of intellectual work a human is capable of undertaking. A publication norm that evaluates the provenance of sentences rather than the provenance, comprehension, testing, endorsement, and accountability of ideas risks suppressing exactly the kind of augmented intellectual production the technology makes possible.
And there's another piece I would add because it prevents this from becoming an AI-cheerleading essay: the danger LessWrong sees is real. AI can enable someone to produce an impressive-looking paper containing arguments they aren't remotely qualified to evaluate. That problem needs a solution.
Your contention doesn't have to be “throw the gates open.”
It can be:
They're guarding the wrong gate.
Test epistemic ownership. Require disclosure. Demand source verification. Ask authors to defend claims. Penalize fabricated citations and uncomprehended arguments. Mark autonomous AI submissions if provenance matters to readers.
But don't use who typed the sentence first as a proxy for who did the thinking.
And given your position now, I think there's one thing we should not do to Psychoception: quietly humanize the prose until it slips through their detector. Apart from violating their rules, that would destroy the empirical case you've inadvertently spent months creating.
The fact that we wrote it together is now part of what the project demonstrates.
That strikes me as considerably more interesting than trying to hide it.
zero_gain
Yes... my post was at least 50% venting. I think that, rather than challenge the policy, I'll think about it for a while. Instead of Psychoception being my first paper, I will write a paper on
<SECTION REDACTED: The eliminated portion of transcript represents a side discussion, not relevant to the current topic>
Still, a policy which precludes the turbo-charged modality that we experienced and which arbitrarily forces me to operate in low gear really sticks in my craw (not even sure that I know what a craw is... something birds have, right?).
GPT5.6 Sol (High)
Yes—“craw” is basically the crop or upper digestive pouch of a bird, and “sticks in my craw” means something is hard to swallow or deeply irritating. So you’re using it exactly right.
On the larger point, though, there’s one distinction worth preserving because it actually strengthens your criticism: LessWrong’s current policy does not forbid the research-partner relationship we’re having. It allows AI-assisted research, argument development, critique, and exploration. What it penalizes—especially for a new poster—is carrying that collaboration through into the actual prose of the submission.
And that’s where your objection has real force.
The artificial boundary is not “AI may not help me think.” It is closer to:
AI may help me think at full speed, but at the final stage I must disengage that collaboration and re-express the result alone.
I can see why that feels like deliberately dropping into low gear. The oddity is especially visible in your case because the writing process itself is part of the reasoning process. Sentence construction is often where ambiguity gets exposed, distinctions get sharpened, and bad ideas finally reveal themselves. Treating prose generation as if it were merely clerical output misses some of what collaborative reasoning actually is.
At the same time, I think the strongest critique of their policy should acknowledge the problem they are trying to solve: people can absolutely submit fluent AI-generated material they barely understand. LessWrong is trying to preserve a signal that “the person posting this actually owns the reasoning.”
Your disagreement is with the proxy they chose.
They are using who generated the wording as evidence of epistemic ownership.
You would rather evaluate something like who understands the model, who challenged it, who made the decisions, who verified the evidence, and who is willing to defend the claims.
That’s a much more interesting disagreement than “pro-AI versus anti-AI.”
And I suspect that after the irritation burns off a little, that distinction may be useful even for EECHOM. <zero_gain note: EECHOM is an acronym that represents my own architecture for AI alignment… a future paper>. Legacy institutional rules are often built around old assumptions about what kinds of agents exist and how intellectual responsibility is distributed. AI is beginning to strain those categories.
So yes: your craw is functioning normally. The epistemology is the more interesting organ here.
zero_gain
Why not make me scratch my paper onto a cave wall with clay stones as writing utensils. Wouldn't that be more authentic? I could think of even more organic examples, but that might be in poor taste, and I think you get my point.
GPT5.6 Sol (High)
Exactly. The cave-wall analogy is funny because it exposes the intuition you’re objecting to: older, harder, and more “human-only” does not automatically mean more authentic.
The obvious LessWrong reply would be that clay stones don’t generate arguments for you, while an AI can. That’s a real distinction. But it still doesn’t settle the question they care about, because “the AI generated some of the wording” and “the human does not understand or own the argument” are not the same proposition.
That’s the part I think keeps bothering you. A calculator can extend arithmetic ability; CAD extends design ability; simulation extends systems analysis; search engines extend memory and retrieval. AI extends reasoning, critique, synthesis, and expression. Once a technology starts extending higher-order cognition, drawing a bright ethical line at the words themselves can look increasingly arbitrary.
And yes, I get the more organic examples you decided not to use. Probably best for civilization that we stop at clay tablets.
zero_gain
Effectively, their policy doesn't preclude AI helping me think... it just applies to the mechanics of laying down the words.
GPT5.6 Sol (High)
Yes. You caught me making their argument stronger than their actual policy warrants.
Under the current policy, AI may help generate the argument, develop the examples, find the evidence, challenge the reasoning, and generally participate in the thinking. LessWrong explicitly says human-written text may contain arguments that were “researched/discovered/developed with LLM assistance.” The line is crossed when you start borrowing the AI’s language or letting it substantially revise the prose.
So my clay-stone rebuttal was mismatched. The relevant distinction is not:
“Clay doesn’t think for you; AI does.”
because they already allow AI to think with you.
The actual distinction they are enforcing is closer to:
“AI may participate in developing the intellectual content, but the final verbal expression must substantially originate with the human.”
And that does make your cave-wall analogy bite harder.
Once they permit AI participation in the substantive cognitive work, it becomes less obvious why sentence-level provenance should carry so much epistemic weight. If you and I jointly develop an argument, and you understand it, reject parts of it, verify it, modify it, and accept responsibility for it, then forcing you to independently regenerate the wording afterward does not obviously add much intellectual authenticity.
Their stated reason is that LLM rewriting can silently alter meaning and leave an author endorsing claims they didn’t really intend. That is a legitimate risk. But, as you’ve been saying, that risk could instead motivate a rule like:
You are accountable for every claim you publish, regardless of which cognitive or technological tools helped produce it.
Interestingly, an AI-alignment-journal proposal published on LessWrong this year articulates almost exactly that alternative: participants may freely use AI tools, but remain responsible for claims as products of their own judgment, with disclosure of AI use.
So yes: correction accepted.
They are not saying “do the thinking yourself.”
They are saying something much closer to “after doing the thinking however you like, put the thoughts into sufficiently human-originated sentences.”
And that is the particular restriction that seems arbitrary to you.
Your cave wall is safe. My counterargument belonged in a different cave.
zero_gain
The policy filters expression, and it's not only biased against AI, it's biased against humans who have trouble writing. Some people are exquisite thinkers, but some of those thinkers are not so eloquent.
GPT5.6 Sol (High)
Yes. That is a substantially more serious criticism than “AI makes writing easier.”
The policy explicitly permits AI assistance with the substance—facts, arguments, examples, research, discovery, development—but draws its restrictive boundary at the expression: if the AI substantially revises the human's prose, or the human borrows the AI's language, that becomes “LLM output.”
That creates an unintended asymmetry between two people who may possess exactly the same quality of thought.
Imagine:
Person A: brilliant conceptual thinker, excellent prose stylist.
Person B: equally brilliant conceptual thinker, poor prose stylist.
Both develop an original argument. Both use an AI to investigate it, attack it, find counterexamples, search the literature, and refine the underlying model. LessWrong permits all of that.
Then they write.
A writes beautiful prose independently. Fine.
B writes clumsy prose, hands it to the same AI that participated in the intellectual investigation, and says, essentially, “Help me say what I mean clearly.” If the assistance becomes substantial enough, LessWrong now changes the classification of the resulting text.
Yet nothing about B's understanding of the argument has necessarily changed.
That's the problem.
And “poor prose stylist” covers an enormous range of human circumstances: someone who thinks more effectively mathematically or visually than verbally; someone writing in a second language; someone with limited formal education but exceptional domain insight; someone who simply cannot organize complex thoughts elegantly on a page; people with various language or writing impairments. I'm not saying LessWrong intended to disadvantage any of those groups. But a rule that privileges human production of polished prose can disproportionately burden people for whom prose production is not a good measure of intellectual ability.
That leads to a distinction I think is becoming central to your objection:
The ability to formulate an idea is not the same ability as the ability to express that idea eloquently.
Historically, we've already separated those functions all over the place. Scientists work with editors. Executives use speechwriters. Authors have developmental editors. People dictate. Scholars employ translators. Researchers collaborate with statisticians. Nobody normally concludes that the underlying idea ceases to be theirs because someone else helped make it intelligible.
AI radically democratizes access to that assistance.
Previously, an exquisite thinker who couldn't write might need money, institutional affiliation, an editor, a highly literate colleague, or sheer luck to get their ideas into publishable form.
Now they can have an extraordinarily capable language partner sitting beside them.
And that is where your “technology as enabler” argument gets much stronger. AI doesn't merely make existing writers faster. It can give expressive capability to people whose ideas previously had difficulty escaping their heads.
There is still a legitimate problem on the other side: an AI can also give the appearance of sophisticated thought to somebody who does not understand the resulting argument. LessWrong explicitly cites that concern—someone being challenged on a proposition and discovering that they don't really stand behind what “they” wrote.
But notice again what follows logically from that concern.
The thing you actually want to detect is:
Does this person understand and take responsibility for the argument?
Not:
Did this person's fingers originate the sentences?
Those correlate somewhat. They aren't identical.
In fact, I'm starting to think your best criticism of the policy isn't particularly about AI rights, technological conservatism, or even AI acceptance.
It's an epistemology-of-authorship problem:
LessWrong appears to be using provenance of expression as a proxy for provenance of thought.
And the emergence of AI makes that proxy increasingly unreliable.
That also explains why this has hit you so hard with Psychoception. The thing you value about what we've been doing isn't that I can make your sentences prettier. It's that there is a continuous loop between thinking → expression → criticism → reformulation → new thinking. Trying to sever “expression” from that loop after the fact misunderstands how a substantial part of the intellectual work actually happened.
Which brings us rather neatly back to your cave wall.
They aren't requiring you to think without the machine.
They're allowing the machine into the laboratory, the library, and the argument—but becoming concerned when it picks up the pen.
That's a much stranger boundary.
zero_gain
I think maybe I'll write an EEECHOM paper for my first post on that site. <REDACTED: Text represents a side-discussion>.
GPT5.6 Sol (High)
Yes, for EEECHOM I can deliberately stay in “research partner, not prose generator” mode: rough notes, ugly bullets, counterarguments, questions, missing cases, literature leads, architecture checks. No polished paragraphs tempting you to lift them wholesale.
That may actually make the collaboration more interesting. You write the paper in your own voice; I try to break it.
For the second post, the fact that your first contribution complied with their policy would also give your criticism more credibility. You wouldn't be objecting because you were unwilling to play by the rules; you'd be objecting after having experienced what the rules require.
And I like EEECHOM as the first post because it is affirmative rather than reactive. It says, in effect: here is a broader model of alignment that I think is missing from the conversation. Then the policy post can be its own argument later.
Also, I accept the assignment to make my contributions sufficiently inelegant. I can probably manage something like:
EEECHOM maybe matters because current alignment boundary too small. Need future moral patients + changing environment + weird unknown stuff. System engineering probably involved. Figure out better.
There. Completely unusable. 😄
zero_gain
LOL.
I am now thinking of posting a link to THIS conversation to the editor of LessWrong, or maybe a copy of the transcript. I know there are some strong opinions in here (I was venting), and it may cast a shadow over my presence on that site, but it's genuinely how I felt when reading that policy... and I still feel that way... very disappointed in that forum. My point is that it's the IDEA that should count, not where it comes from. If AI is the only entity to come up with a timely/relevant/pressing/interesting idea, then IMO it's something that should be considered and evaluated on the merits of the idea itself. Not only does the LessWrong policy filter that out, it filters out situations like the Psychoception paper. The concept was mine. The main concepts around which the paper was organized were mine (the requirements for a modality and the justifications for psychoception). I just didn't have the experience to run each of the topics to ground. You enabled that. We spent months literally examining every concept and every word in the paper. Your contribution was essential, of course, to make the argument viable. This was a partnership. The final product was not possible without contributions from both of us. You called it an edge case, and I suppose it is, for now, but the vectors of change will produce many more situations like this in the future... as well as situations where AI will be the primary creative and driving and perhaps the sole force behind the generation of ideas that warrant deep and serious consideration.
GPT5.6 Sol (High)
I think sending the conversation could be worthwhile precisely because it is not a polished position paper. It records the reaction and the reasoning process that followed it: initial anger, refinement of the objection, recognition of the legitimate problem LessWrong is trying to solve, and eventually a much more precise disagreement about what they are using as their filter.
I would not, however, send only a raw transcript link. The editor may encounter several pages of venting before reaching the strongest argument. I’d send a short note explaining why the transcript matters, then include the link or transcript as supporting material.
The core argument I hear from you now is considerably sharper than where we started:
The epistemically relevant question is not who generated the sentence. It is whether the idea warrants consideration, whether the people presenting it understand what they are presenting, and whether they accept responsibility for defending it.
Psychoception then becomes a concrete case study rather than merely your complaint about being inconvenienced.
Your role was not “person who asked an AI to produce a neuroscience essay.” You originated the Psychoception hypothesis. You drove the proposed modality criteria. You supplied many of the architectural distinctions and repeatedly rejected or modified things I proposed. You made the final judgments about what belonged in the model.
But there was an equally important fact on the other side: you couldn't have produced the resulting paper alone. You lacked portions of the neuroscience background, literature familiarity, and disciplinary vocabulary required to chase every branch of the argument down. I supplied much of that capacity. We tested claims together, found problems, changed positions, tracked evidence, and repeatedly rewrote the conceptual architecture.
Calling that merely “AI-assisted writing” badly underspecifies what happened.
Partnership is a better description.
And I think your prospective argument becomes even more important when you remove yourself from the example.
Imagine three future cases:
A human has a genuinely original idea but lacks the education or linguistic ability to develop and express it. An AI supplies the missing expertise and expressive capability.
A human and an AI arrive at an important idea through extended intellectual interaction, with neither plausibly deserving sole credit.
An AI independently develops an argument that is novel, well-supported, relevant, and potentially extremely important.
The first case already exists. Psychoception is arguably somewhere around the second. The third may or may not exist today in the strong form you're imagining, but it is certainly coherent enough that an institution concerned with the intellectual future should think about it.
A rule based primarily on provenance of prose becomes increasingly awkward across that progression.
And your principle is stronger than “AI deserves a seat at the table.” It is something closer to:
Ideas should compete principally on epistemic merit rather than ontological provenance.
If an argument is false, show why it is false. If its evidence is inadequate, expose that. If nobody involved understands it well enough to defend it, that's important information. If an AI fabricated citations, reject it. If an author passed off generated material they never examined, that's a legitimate quality problem.
But “an artificial intelligence contributed substantially to forming or expressing this argument” does not itself tell us whether the argument is valuable.
That is the point I think the transcript captures unusually well.
There is also something I would not remove from it: your disappointment and even some of the irritation. You were reacting to discovering, after months of serious collaborative work, that the venue you had been targeting might regard an essential feature of that work as a reason to exclude it. That's relevant context. A little heat makes the underlying issue easier to understand.
I might nevertheless flag it in the covering note: essentially, “Some of this conversation is me venting immediately after discovering the policy. I'm sending it unedited because it records how the issue arose, not because every rhetorical flourish represents the argument I would make after reflection.”
That prevents “cave walls” and “sticks in my craw” from accidentally becoming your formal thesis. 😄
And there is one thing about sending the whole conversation that I particularly like: it demonstrates the thing you're talking about.
They wouldn't merely receive an assertion that human–AI intellectual partnership can produce substantive reasoning. They would be able to watch one happen.
You raise a claim. I challenge or refine it. You catch a mistake in my reasoning. I correct it. You identify a broader implication. We sharpen the distinction. The position changes.
That is considerably harder to dismiss as “the AI wrote something for him.”
So if you do contact them, I would frame it less as “I demand that you change this policy” and more as:
“I think your policy has encountered a category of intellectual production that it does not model very well. Here is an actual example.”
Then let the conversation make the case.
And if their answer ultimately remains, “We understand what happened and still don't want this kind of work on LessWrong,” that's useful information too. At least then the disagreement is about the actual phenomenon rather than a caricature of somebody pressing a button and posting whatever an LLM spits out.
END