# Inner Speech and LLM Coupling - Research Discussion This is a research brainstorm from a Raycast AI conversation exploring parallels between [[inner speech]] and LLM dialogue. ## Ideas ### 1. Reciprocal Prompting "We are prompting the LLMs, but when the text returns, they are prompting us." This connects to [[Keith Frankish]]'s cyclical model of reasoning: - Produce symbols → Perceive them → Interpret as posing subproblem → Form beliefs → Produce further symbols → Repeat In LLM dialogue, the cycle is split: - You produce prompt → LLM produces response → Response prompts further cognitive activity in you → You produce follow-up → Repeat The LLM plays the role that autonomous (Type 1) processes would play in inner speech—but can generate content genuinely outside your cognitive resources. ### 2. The "Unauthoredness" Parallel Both inner speech and LLM text lack a clear author in the full sense: **Inner speech**: You're both author and audience, but authorship is attenuated. Words often "come to you" rather than being deliberately produced. The "interlocutor" is obscure. **LLM text**: Generated but not "authored"—no person behind it with intentions, beliefs, desires, history, perspective. Both occupy an intermediate zone between fully authored speech (human communication) and random noise. **Implications**: - Similar interpretive stance—interpret without relying on authorial intention - Similar openness—no author constraining meaning - Reduces "social overhead" (no [[Theory of Mind]] processing required) - Creates space that is "cognitively dense like a conversation but socially transparent like inner speech" ### 3. The "Double-Extension" Structure 1. **Internal Extension**: Language is already a cognitive technology ([[Andy Clark]], [[Gary Lupyan]]) 2. **Internalization**: Inner speech is "contracted" outer speech (Frankish, [[Lev Vygotsky]]) 3. **LLM Coupling**: Re-expansion of that loop—coupling with external version of the very medium that constitutes internal thinking Not just extending memory (Otto's notebook)—extending the internalization mechanism itself. ### 4. Different Accommodations of "Language Helps Us Think" Everyone in *[[Inner Speech - New Voices]]* agrees language can be used to work through problems. Different accommodations: - **Format/Vehicle View** ([[Peter Carruthers]], [[José Luis Bermúdez]]): Language provides sensory vehicle making thoughts stable enough for reflection - **Activity/Constitutive View** (Frankish, [[Christopher Gauker]]): Inner speech IS thinking, not expression of it. [[dual-process theory|Type 2 reasoning]] is activity conducted in language - **Internalized Dialogue View** (Vygotsky, [[Charles Fernyhough]]): Inner speech is internalized social dialogue - **Decomposition View** (Frankish): Inner speech breaks complex problems into subproblems solvable by autonomous processes ### 5. Phonological Loop vs Inner Speech - **[[Phonological loop]]** ([[Alan Baddeley]]): Maintenance system—keeps verbal info active through rehearsal - **Inner speech**: Much broader—generation, reasoning, self-regulation, making commitments LLM coupling is about generation/processing, not maintenance—so phonological loop isn't the right comparison class. ## Research Directions ### Path A: Prompter-as-Partner (Operational similarity) LLM as externalized System 1. The LLM's response acts as prompt that your System 1/2 must interpret. ### Path B: Phenomenology of Unauthoredness LLM occupies "sweet spot"—enough otherness to break cognitive ruts, enough nothingness to prevent social anxiety. "Pure linguistic feedback without interpersonal weight." ### Path C: Double-Extension Argument (Structural novelty) LLM coupling is recursion through the medium of thought—adding turbocharger to the internalization mechanism itself. ### Path D: Agency and Attribution Connect to comparator models. If inner speech agency depends on match between intended and heard, what happens when incorporating LLM text? "Benign thought insertion"? ## Working Thesis LLM dialogue is "hybrid inner speech"—has privacy and lack of social overhead of talking to oneself, but novelty and objective resistance of talking to another person. --- ## Theoretical Developments (12 Jan 2026) ### Why Inner Speech Provides Theoretical Foundation for LLM Coupling The inner speech literature answers questions that otherwise seem mysterious about LLM dialogue: **Why does "cognitive traction" help?** Because reasoning is *loopy*. Frankish's cyclical model: Produce → Perceive → Interpret → Respond → Repeat. The LLM participates in that loop. When you prompt, you externalize; when you read the response, you perceive and interpret; that triggers further cognitive activity. The loop needs something to push against—the LLM provides that surface. **Why does non-personhood help?** Because it removes the "social overhead" that comes with talking to another person (no [[Theory of Mind]] processing, no face concerns, no managing the relationship), while preserving the "objective resistance" that comes with genuinely external input. You get the cognitive benefits of dialogue without the interpersonal costs. **Why is LLM dialogue different from just thinking harder?** Because the LLM can generate content *genuinely outside your cognitive resources*. In normal inner speech, you're triggering autonomous (Type 1) processes that draw on what's already in your head. The LLM can produce connections, framings, and articulations that you wouldn't have reached alone. It's not just reorganizing existing resources—it's adding new material to the loop. **Why is it different from writing?** Writing externalizes thought but is: - One-way (you produce, then read your own words) - Fixed once produced - Limited to your own cognitive resources LLM dialogue is: - Bidirectional and responsive - Malleable (you can push back, ask for reformulation) - Capable of generating genuinely novel content **Why is it different from conversation with another person?** Conversation has: - Social overhead (theory of mind, face management, relationship dynamics) - Authorial intention you must model - The other person's agenda LLM dialogue has: - No social overhead - No author to model (unauthoredness) - Pure responsiveness to your cognitive needs ### The "Hybrid Inner Speech" Formulation LLM dialogue is "hybrid inner speech" because it combines: **From inner speech:** - Privacy (no audience, no judgment) - Lack of social overhead - Freedom to think half-formed thoughts - The interpretive stance of not relying on authorial intention **From external dialogue:** - Novelty (content you couldn't have produced alone) - Objective resistance (it's not just telling you what you already think) - The perceivable, stable quality of external language - The responsive, interactive quality of conversation This is the "sweet spot": enough otherness to break cognitive ruts, enough nothingness to prevent social anxiety. "Pure linguistic feedback without interpersonal weight." ### Cognitive Traction (Further Development) The metaphor: Traction is what lets movement happen. Gravel gives tires purchase on the road. The LLM's words give your thinking purchase. **What provides traction:** - The words are external (perceivable, stable, not fleeting like inner speech can be) - They're responsive (unlike writing, which is fixed) - They're novel (unlike inner speech, which is constrained by your own resources) - They're impersonal (no need to manage social dynamics) **What traction enables:** - Forward motion on problems you're stuck on - Articulation of thoughts you couldn't quite express - Discovery of connections you wouldn't have made - Decomposition of complex problems into tractable subproblems **Important clarification:** Traction is not adversarial. The resistance is like gravel giving tires purchase, not like an opponent pushing back. Bad sessions aren't about lack of traction—whenever you're in flow with the LLM, you're getting traction. The difference between good and bad sessions is whether the things you're gripping onto are *useful or interesting content*. ### The Double-Extension Argument (Sharpened) This is potentially the most original contribution: 1. **First Extension**: Language is already a cognitive technology ([[Andy Clark]], [[Gary Lupyan]]). We use words to stabilize thoughts, decompose problems, make commitments. Language augments cognition. 2. **Internalization**: Inner speech is "contracted" outer speech ([[Lev Vygotsky]]). We internalize the linguistic tool. What was once external dialogue becomes internal self-talk. 3. **Second Extension (LLM Coupling)**: We couple with an *external version* of the same medium that constitutes internal thinking. The loop that was externalized → internalized is now re-externalized, but with a responsive, generative partner. **Why this is structurally different from Otto's notebook:** Otto extends *memory*—he uses the notebook to store and retrieve information. LLM coupling extends *the thinking process itself*—or at least, the verbal component of it. It's not storage; it's generation and processing. This is recursion through the medium of thought—adding a turbocharger to the internalization mechanism itself. ### Engaging with AGI Skepticism **[[Benjamin Riley]]'s argument** (The Verge, November 2025, "Large Language Mistake"): - Language ≠ intelligence - LLMs model the *communicative* function of language, not the cognitive process of thinking - Neuroscience shows distinct brain regions for language vs. reasoning - If language *were* thinking, taking it away should take away thought—but it doesn't **The dialectical opportunity:** Riley is *right* about the narrow point: Language isn't intelligence. LLMs modeling language doesn't make them intelligent. But he's *missing* the broader picture: Language *helps us think*, even if it isn't thinking itself. And that's precisely where LLMs become interesting—not as artificial minds, but as cognitive tools that extend the same mechanism by which language already extends cognition. **Positioning:** - Not the AGI hype crowd ("LLMs are thinking!") - Not the dismissive skeptics ("LLMs are just autocomplete!") - Something more nuanced: "LLMs extend human cognition through language, in ways structurally similar to how inner speech does" ### Connection to "LLMs and Understanding" Draft The existing Substack draft ([[LLMs and Understanding]]) argues: - Don't treat LLMs as oracles (fact-dispensers) - Treat them as "comprehension stretchers" - Understanding (not knowing) is the goal - The aesthetics of understanding ([[Elisabeth Schellekens]]): coherence, unity, elegance The inner speech framework provides *theoretical foundation* for why the "comprehension stretcher" approach works: - Why does traction help? → Loopy reasoning needs something to push against - Why does non-personhood help? → Social overhead removed, objective resistance preserved - Why is it different from just thinking? → Content outside your cognitive resources The two pieces are complementary: - Inner speech piece: *Why* does this work? (Theoretical) - Understanding piece: *What* do we do with it? (Practical/aesthetic) --- ## Sources Referenced - [[ChatGPT Extended - Large Language Models and the Extended Mind|Smart, Clowes & Clark (2025) - ChatGPT Extended]] - [[Loops, Constitution, and Cognitive Extension|Palermos (2014)]] - [[Inner Speech and Outer Thought|Frankish (2018)]] - [[Inner Speech - New Voices|Langland-Hassan & Vicente (2018)]] - [[Supersizing the Mind|Clark (2008)]] - Løevenbruck et al. (2018) - A Cognitive Neuroscience View of Inner Language - [[The Centered Mind|Carruthers (2015)]] - Gauker - Anti-Gricean view of inner speech - Lupyan - Language-augmented cognition - [[Benjamin Riley]] - "Large Language Mistake" (The Verge, November 2025) - MIT Technology Review - "The great AI hype correction of 2025" (December 2025)