# Text > I was an everted person, with my tiny, fragmented body situated at the center of my own distended brain. It was in this unlikely configuration that I began to explore myself. > — Ted Chiang, Exhalation --- > [!NOTE] > please note, the following, contentwise and structurally, is great. In terms of tone/style, vocabulary, register and jargon, it's appalling. # PART ONE: The Defense of the Instrument *Clearing the air — Establishing the decoupling of cognition from value.* ### I. The Opening Can LLMs produce philosophy? I do philosophy with LLMs. Extensively. A significant portion of my working day is spent in reciprocal dialogue with these systems. But the question of whether LLMs can "do philosophy" is perhaps a category error. A more pressing concern is whether they can produce philosophy of sufficient quality to be useful, and whether we, as a community of practitioners, ought to adopt them. I write as an extensive user—not one who has outsourced the cognitive labor, but one who has found in these systems a new way to conduct the activity of thinking itself. As you will see, this is less a matter of having the machine "do" the work, and more a matter of using it to "evert" and refine the process of inquiry. ### II. Decoupling Cognition and Value Just to be clear: I do not believe these systems "think" in any sense analogous to humans or animals. Their architecture—next-token prediction over a multi-dimensional vector space—is fundamentally unlike the biological brain. They are not minds. However, we are now in an age where philosophically valuable text need not be produced by a system which understands that text. **The Functional Reliability Move (Butter Chicken)**: We evaluate a tool by its performance, not its biography. If an LLM provides a reliable recipe for Butter Chicken, the recipe is "good" because it works, regardless of whether the generator "intended" to feed me or understood the chemistry of spices. Similarly, if an LLM produces a text that identifies a valid logical dependency or a non-obvious counter-example, that text is "good philosophy" because it functions as philosophy. Reliability is a property of the output, not the process. ### III. The Safety of the Discipline (Conceptual Transparency) Of course, technical unreliability—the risk of "hallucination"—is a diminishing problem. Modern systems (January 2026) are significantly more reliable and have direct access to external knowledge bases. But even if they were not, the worry is largely misplaced in our discipline. In empirical fields, one requires a "truthful witness" (testimony). But philosophy is an **auto-evaluating** or **self-authenticating** practice. Understanding an argument *is* verifying it. We do not "check the facts" of a thought experiment; we "follow the logic." Because philosophical understanding is a matter of direct conceptual inspection, a bad argument is visible the moment it is comprehended. The philosopher is the verification mechanism. ### IV. Recombination Realism (The Novelty Objection) Skeptics often claim that LLMs only remix and cannot create. However, recombination is not a bug; it is the fundamental feature of all creative systems. Drawing on the work of Boden and Gaut, we can see that "Creation Ex Nihilo" is an incoherent myth. All human breakthroughs are transformations or recombinations of existing conceptual spaces. If humans are "pattern-dwellers," the LLM's recombinatorial nature is not a limitation, but a shared architectural reality. The value lies in the "flair" of the selection—which brings us to the human in the loop. --- # PART TWO: The Lab of the Everted Brain *The mechanism — Framing thought as an interactive activity.* ### I. The Labor of Legibility To get philosophy out of a machine, you cannot be lazy. You must define premises, set registers, and articulate the "Ultimate Question" with exquisite precision. This is what Keith Frankish calls **"deliberative mastication."** The act of making your thought "legible" to the system forces a level of clarity that solitary thinking often avoids. The intellectual work is in the *steering*. ### II. The Frankish Cycle (Reciprocal Prompting) Frankish defines conscious thought as an activity (the "activity view"). It follows a loop: Produce symbols → Perceive → Interpret → Respond. In this workflow, the cycle is split. I *produce* the prompt; the LLM provides a *perceivable symbol* that is genuinely outside my own immediate resources; I *perceive and interpret* that stimulus, which "prompts" my next move. The LLM acts as a participant in my cognitive loop. It provides **Cognitive Traction**. ### III. The Friction of Failure Even "bad" or non-optimal responses are useful. They provide the "tire on the road" resistance (traction) that lets thought move forward. A flawed output forces the philosopher to articulate exactly *why* it fails, driving the understanding deeper. ### IV. Conceptual Triangulation (The Everted Self) My actual workflow involves bouncing ideas between several models (Claude, Gemini, ChatGPT). I am not seeking an "oracle"; I am using these systems to externalize my own thinking. As Ted Chiang’s anatomist sees his own brain working through a periscope, I see my own philosophical thinking externalized on the screen—everted, observable, and open to cold, analytical inspection. --- ## Future Project Note: "These Words Are Mine" > These words are mine, no one else's. I love you, I love you, I love you. > — Natasha Bedingfield, "These Words" (2004) - I do philosophy with LLMs. Extensively. As in, a significant part of a working day for me will be spent bouncing between various chats with various chatbots. - As far as I can tell I am in a minority. Most philosophers I know well either use it a bit, but more often for fairly mundane things, or dislike AI so much that the very idea is anathema. Another common reaction is surprise that LLMs are fit for such a purpose. - I'll talk more specifically about what I am doing a little bit later. - Why do I use them? I mean, there has been a modest increase in getting papers finished, %%this not well formed%% - I think that in January 2026, they have become a very good way of *getting better at philosophy, or enhancing my philosophical understanding*. %%this not well formed%% - As far as I can tell I am in a minority. Most philosophers I know well either use it a bit, but more often for ffailrly mundane things, or dislike AI so much that the very idea is anathema. Another common reaction is surprise that LLMs are fit for such a purpose. - - Can LLMs do philosophy? In a sense, this is not quite the right question to ask. If you ask ChatGPT or Claude or Gemini a philosophical question, they'll certainly have a go. Two more interesting questions are: Can LLMs produce _good_ philosophy? And: Should they? Even if they can, should we philosophers be using them? In what follows I will have more to say about the first question than the second. Regarding the second, I should say, to get it straight from the start: I am a philosopher and I use these systems extensively — not just to write, but to think. You will see how extensively in a minute, and hopefully this will persuade you that what's happening is doing philosophy _with_ an LLM, rather than the LLM doing the philosophy for me. - Now, just to get this out of the way. I don't think that these things are anything like actual things that think. Like humans or animals. - [Explained succinctly that this is primarily because the architecture of LLMs is so very much unlike the architecture of our brains. Obviously there's a lot more to be said here but I'm not going to be focusing on that today, I'm just going to take it as read. These things are not minds or pseudo-minds or anything like minds.] - but I don't think this prevents them from offering boundless opportunities for philosophers who want to get better at being philosophers. ## new plan made from gemini ### Part 1: The Skeptical Threshold (Clearing the Air) **Purpose**: To dismantle the reflexive dismissal of LLMs in philosophy. You aren't arguing that LLMs are "smart," but that they are "useful" within the specific epistemic constraints of the discipline. #### **Argument A: The Butter Chicken Defense (Intent vs. Reliability) %% I think this is actually very muddled. Rather than framing it as a response to an objection, this should actually be more along the lines of making my claim clear: I am not saying that LLMs are capable of doing actual philosophy, or any sort of thinking whatsoever, I do not think that they are agents. but I do not think that this impedes their capacity to produce good philosophy. that is, and this line should be added somewhere: we are now in an age where philosophically valuable text (mutatis mutandis spoken words) need not necessarily be produced by a system which *understands* said text. this seems like an important point. and I think this part of part 1 should be rewritten completely around this idea%% *   **The Objection**: "LLMs are just stochastic parrots; they have no intent to tell the truth, so their output is philosophically worthless." *   **The Move**: Decouple the *intent of the process* from the *reliability of the product*.  *   **Elaboration**: Use the **Butter Chicken** analogy to show that we evaluate tools by their performance, not their biography. If a recipe works, it is a "good recipe" regardless of whether it was generated by a chef or a probability distribution. In philosophy, if a text provides a valid counter-argument or a useful distinction, it is "good philosophy" regardless of the generator’s "inner life." Reliability is a functional property of the output. #### **Argument B: Philosophy as a Self-Verifying Discipline** %%after the rewrite of the section preceding this one, it's most likely that everything here will have to be rewritten quite substantially as well. The ideas are good though as once i have established that i am not making claiming about llm cognition, the next thing a reader will think about is hallucinations.%% %% also, another couple of things that should be mentioned in passing is LLMs in 2026 hallucinate a lot less than they did in 2022, like a lot less. I'll find some stats somewhere. The second thing is that LLMs in 2026 also have access to Google and access to other databases of knowledge. They can therefore in many circumstances look up facts. I say most of these should be mentioned in passing because, yeah, that's not the most interesting part of this section. The most interesting part is the 'auto-evaluative' stuff. altohugh i am still scrabbling around for a label for this. Also, I wonder if this auto evaluative stuff might be tied into the self prompting stuff.%% *   **The Objection**: "LLMs hallucinate facts; you can't trust them." *   **The Move**: Distinguish between *empirical testimony* and *philosophical argument*. *   **Elaboration**: In science, you must trust the "testimony" of a lab report. In philosophy, you **auto-evaluate**. Understanding an argument *is* verifying it. If an LLM suggests a logical transition, the philosopher doesn't "check the facts"—they evaluate the logic. This makes philosophy uniquely resilient to "hallucinations." A bad argument is visible the moment it is understood. #### **Argument C: The Recombination Realism (Reduced Novelty Objection)** %% this is okay at the moment, although yeah I'm gonna sleep on it and see how it can be improved. One thing to mention or to remember is that in the last day or so there was a quick substack by a guy saying that of course llms can create new knowledge, but he was struggling to see how they could produce new concepts. I don't want to repsond extensively to this idea here, but it isworth linking too.%% *   **The Objection**: "LLMs can only remix training data; they can't produce the 'genuinely new' moves philosophy requires." *   **The Move**: Collapse the distinction between "genuine novelty" and "sophisticated recombination." *   **Elaboration**: Draw on **Boden** and **Gaut** to argue that "Creation Ex Nihilo" is a myth. Human creativity—including Kripkean or Kantian breakthroughs—always proceeds by transforming or recombining existing conceptual spaces. If humans are "pattern-recombiners," then the LLM's recombinatorial nature is not a limitation, but a shared architectural feature. --- ### Part 2: The Everted Philosopher (The Mechanism) **Purpose**: To move from "what the machine does" to "what the human does with the machine." This section defends the user’s agency and shows that AI-augmented philosophy is a more intensive version of traditional thinking. #### **Argument A: The Reciprocal Prompting Loop (The Frankish Connection)** *   **The Core Idea**: Use **Keith Frankish’s "activity view"** of inner speech to explain why LLM dialogue counts as "thinking." *   **The Move**: Frame the LLM as a participant in the "Produce → Perceive → Interpret → Respond" cycle. *   **Elaboration**:      *   Frankish argues that Type 2 reasoning (conscious, effortful thought) is an *intentional activity* where we prompt ourselves with symbols.      *   In your workflow, the **Reciprocal Prompting** loop splits this cycle: you *produce* the prompt (intentional act); the LLM *responds* with content genuinely outside your own immediate cognitive resources; you *perceive and interpret* that response, which then *prompts* your next move.     *   This isn't "outsourcing" thought; it's a "double-extension" of the very mechanism (internalized dialogue) that constitutes human thinking. #### **Argument B: The Articulation Demand (Prompting as Intellectual Labor)** %% I wonder if this should go earlier, or if the structure of this part of the text should be changed some way. The reason why is... I wonder if this part of the text should begin by first sort of just suggesting a few ways in which an LLM can be used so that it is not cheating, in scare quotes. OK, and also just to sort of pump the intuitions that using an LLM can be philosophically... Yeah, can be work which is philosophically robust work. You know what I'm trying to get at here? And I'm wondering if the articulation demand is just one idea. By the way, I hate the label of the articulation demand. We need to change that. I mean the mirror of this is also understanding and evaluating the responses of the LLM. Note that bad responses by the LLM or non-optimal responses can sometimes, maybe even more than sometimes, be just as useful. OK, and then this last little bit can then lead us on to the self-prompting Keith Frankish stuff. %% *   **The Core Idea**: Addressing the "laziness" charge. *   **The Move**: Explaining what you want to an LLM is a high-level philosophical exercise. *   **Elaboration**: To get "good philosophy" out of a model, you cannot be vague. You must articulate premises, define the required register, and set the dialectical stakes with extreme precision.  *   **Point**: This "Articulation Demand" is a form of **deliberative mastication** (Frankish). The act of making your thought "legible" to the system forces a level of clarity that solitary "head-thinking" often avoids. Right now, as you transcribe and explain these ideas to me, you are doing the "work" of philosophy. #### **Argument C: Triangulation and the Everted Self** %% this should be framed as a result of what's been described in argument A and argument B. By the way, they shouldn't be called arguments here. It's not quite in keeping with the style of the paper.%% *   **The Core Idea**: The phenomenology of the "Thinking Screen." *   **The Move**: Describe the "Multi-LLM" workflow as a method of **Conceptual Triangulation**. *   **Elaboration**: You aren't asking an "oracle"; you are bouncing ideas between multiple models (Claude, Gemini, ChatGPT) to see where the logic holds and where it breaks. This is **Cognitive Eversion**: seeing the "periscope" of your own thinking externalized on the screen (Ted Chiang).  *   **Conclusion of Part 2**: You are still "doing the work." You are the evaluator, the synthesizer, and the director of the loop. The LLM is the "traction" (the tire on the road) that lets the thought move forward. --- ### **Synthesis of Ideas for Part 1 and Part 2** | Feature of Your Workflow | Skeptical Objection Addressed | Theoretical Support | | :--- | :--- | :--- | | **Butter Chicken Logic** | No "Intent" to tell truth | Output Reliability vs. Process Intent | | **Auto-evaluation** | Hallucination risk | Philosophy as a self-verifying discipline | | **Recombination Realism** | Lack of genuine novelty | Boden/Gaut: Creativity as Transformation | | **Frankish Loop** | Outsourcing thought | Conscious thinking as "Internalized Dialogue" | | **Articulation Demand** | Intellectual laziness | Prompt-craft as "Deliberative Mastication" | | **Cognitive Eversion** | Loss of Authorship | Ted Chiang: Thought made observable on screen | --- ### **Drafting Notes on Voice and Style** *   **Dryness**: Avoid laudatory adjectives for the AI (e.g., instead of "this amazing reasoning engine," use "this functional simulation of a reasoning trace"). *   **The "We" vs. "I"**: In Part 2, lean into the "I"—the personal workflow description makes the "The Everted Philosopher" argument feel grounded rather than speculative. *   **Footnote placement**: Use the **Freddy deBoer** quote as a footnote to the "thinking with" claim in Part 2, aligning your experience with his observations on the shift in writing practices. **Does this elaborate structure for the first two sections feel like the right "fine-grained" step?**  <final_answer> - The essay is now structured as a three-part argument: **Skeptical Threshold**, **The Everted Philosopher**, and **The Provocation**. - Part 1 clears the air by using the **Butter Chicken defense**, the **Auto-evaluating register** of philosophy, and a **Recombinatorial account of creativity**. - Part 2 uses **Frankish’s cyclical model** to define "Reciprocal Prompting" and frames the **Articulation Demand** as the core intellectual labor of the user. - The "Novelty Objection" is integrated into Part 1 to prevent Part 2 from feeling defensive. - Pacing remains a point of consideration, but the conceptual links to the Frankish paper are now firm. </final_answer> # Notes ## Working Notes (18 Jan 2026) **The two-part thesis:** - (a) Philosophers can use LLMs to increase their *own* understanding - (b) Speculatively: maybe philosophers can be taken out of the loop entirely --- # PART ONE: LLMs Can Help Philosophers Generate Understanding *The cognitive loop argument — LLMs participate in generating understanding through reciprocal prompting* --- ### Setup for Skeptics: Why Think LLMs Can Reason At All? **Why butter chicken works**: Recipes are abundantly represented in training data. When you ask for butter chicken, the model pattern-matches against thousands of recipes. High probability of decent output because the target is well-represented. **But philosophy isn't like that**: There's no "latent philosophical tract embedded in GPT-5.2." You're not extracting something sitting there fully-formed. So why would the model produce anything philosophically worthwhile? **The response — LLM reasoning capacities**: **How CoT works (brief account for non-technical readers)**: Model generates tokens sequentially; each token becomes context for the next. Prompting for steps produces scaffolding the model can use. Extended thinking modes automate this. **Benchmarks as evidence**: Math competitions, bar exams, coding tasks. Outputs demonstrably contain correct reasoning chains with high reliability on well-defined problems. Establishes that LLMs *can* produce reliable reasoning, making (a) and (b) plausible. *It doesn't matter if this is "real" reasoning or not — what matters is that it gets the job done.* --- ### Why Philosophy Is Uniquely Suited to LLM Collaboration **Philosophy's anti-testimonial nature**: You can't just trust expert testimony in philosophy; you have to evaluate the arguments yourself. This means the LLM-trust question is different from the expert-trust question. When a philosopher reads LLM output, they're not taking it on testimony — they're evaluating the arguments. The same cognitive processes that would generate the philosophy are used to judge it. **Philosophy vs. other disciplines on hallucination**: In science or history, if an LLM hallucinates a fact (a study that doesn't exist, an event that didn't happen), you need external verification. You either trust or check. But in philosophy, the "checking" is the reading. When you evaluate an argument, you're not verifying against external facts — you're assessing whether the reasoning holds, whether there are counterexamples, whether the moves are valid. A philosopher who understands what they're reading is already doing the verification. --- ### Core Theoretical Move: Reciprocal Prompting "We are prompting the LLMs, but when the text returns, they are prompting us." LLMs participate in [[Keith Frankish]]'s cyclical reasoning loop: - Produce symbols → Perceive them → Interpret as posing subproblem → Form beliefs → Produce further symbols → Repeat In LLM dialogue, the cycle is split between human and machine. The LLM plays the role that autonomous (Type 1) processes would play in inner speech — but can generate content genuinely outside your cognitive resources. **Self-prompting framing**: The LLM is a tool for prompting *yourself*. You're not outsourcing thinking; you're externalizing it and receiving material that drives your cognition forward. **Inner speech connection**: Like Frankish's cyclical loop, but split across two participants. The LLM generates content genuinely outside what you'd produce alone. **Cognitive eversion / thinking on screen**: Doing philosophy with an LLM involves seeing your thought process externalized, like the anatomist seeing his own brain work. Callback to epigraph. --- ### The Articulation Demand (Against the "Laziness" Objection) Working with LLMs is cognitively demanding, not lazy. You have to: - Explain what you mean in huge detail to make it legible to the LLM - Push back when something seems off - Synthesize across multiple LLM responses (triangulation) - Decide what's worth keeping and what to discard The articulation work *is* intellectual work. The process of making your thought clear enough to prompt the LLM forces clarity in your own thinking. This ties directly into reciprocal prompting — the LLM forces you to think more precisely. --- ### Handling the "Process-Trust" Objection **Process-trust asymmetry argument**: "You can't trust LLM output because it wasn't aimed at truth" **BUTTER CHICKEN counter**: Process-aim doesn't predict reliability. The recipe works regardless of the LLM's "intentions." If an LLM gives me a butter chicken recipe and following it results in good butter chicken, the recipe worked — whether or not the LLM was "aiming" at truth. --- # PART TWO: Could Philosophers Be Taken Out of the Loop Entirely? *Provocative speculation — what if the loop could run without us?* --- ### The Provocation Given that: - LLMs can produce reasoning chains - The loop generates understanding through reciprocal prompting - Philosophy is verified by reading/evaluating, not external fact-checking ...do we actually need humans in it? Could LLMs prompt *each other* and generate philosophical understanding? Could an LLM prompt *itself* through iterative refinement and build accurate dependency maps without any human in the loop? --- ### Objections to "Philosophers Out of the Loop" (and Counters) | Objection | Counter | | ---------------------------------------------------------------------------------------- | ---------------------------------------------------------------------------------------------------------------------------------- | | **Stakes/Caring**: Humans care about getting it right; LLMs don't have anything at stake | BUTTER CHICKEN — the recipe works regardless of caring | | **Evaluation**: Who evaluates the output? | Same process for generating and judging — the cognitive machinery that would produce philosophy is the machinery that evaluates it | | **Verification**: How do you know it's right? | By reading it. Same as how you verify human philosophy. | | **Novelty**: LLMs can only recombine existing patterns | See below | --- ### The Novelty Objection (Elaborated) The objection distinguishes: - **"Genuine novelty"** — creating outside the pattern-space - **"Mere recombination"** — working inside the pattern-space The claim: LLMs can only do the latter; real philosophy requires the former. ### Why "Outside the Pattern Space" Doesn't Hold Up **Boden's three types of creativity:** 1. **Combinational** — familiar ideas combined in unfamiliar ways 2. **Exploratory** — exploiting existing conceptual spaces 3. **Transformational** — altering the rules of existing conceptual spaces All three operate within/against existing conceptual spaces. None requires accessing something "outside." **Key points from the literature:** - "The alternative idea of creation ex nihilo, as instanced in one view of God's act of creation, is close to incoherent." (Gaut/Kieran anthology) - Even transformational creativity must be "intelligibly close to the previous way of thinking" if it is to be accepted (Boden) - Gaut: flair is about *how* you combine/explore/transform, not about accessing outside-space - Humans also work by recombination — nobody creates ex nihilo **The upshot**: If the novelty objection is to have bite against LLMs, it would need to identify something about *how* humans recombine/explore/transform that LLMs cannot replicate — not claim that humans escape pattern-space altogether.