> [!info] Working Notes > This file is an information dump for the Ghost Writing essay. Not a draft. ## Key Sources - [[On Artificial and Post-Artificial Texts]] — Hannes Bajohr (2023) - URL: https://hannesbajohr.de/en/2023/03/11/on-artificial-and-post-artificial-texts/ - [[Phenomenology of Reading - Research Overview]] --- ## Quotes to Use ### From Bajohr (2023) > "Various think pieces were quick to speculate that such language models will one day replace human authors; for various reasons, I doubt that. But AI's transformation of literature need not be so extreme for our *perception of* text to change fundamentally. Technologies like these have already taken on assistive functions—not doing all the writing work, but helping to produce much more text much more quickly, and with the help of fewer and fewer people. Certain types of writing are becoming at least *partially* automated." Summary from Readwise: "AI-generated text is becoming widespread and will change how we read. People will stop assuming unknown texts are written by humans and start doubting their origin. Eventually the line between human and machine writing may blur so much that the distinction no longer matters." ### From Phenomenology of Reading note > "Does LLM text have different phenomenal texture than authored text? Is reading it more 'opaque' (attending to how it says things) vs 'transparent' (seeing through to meaning)?... Does knowing text is LLM-generated prevent full immersion? Are you always somewhat 'outside' the text?" > "Authorless form-content unity: Pierini's account links auditory-semantic unity to poetry. With LLM prose, is there form-content unity? If so, what does 'authorial voice' mean when there's no author?" --- ## Ideas I Like (from 14 Jan brainstorm) ### The ghost is literal A ghost is an appearance without substance — the visible trace of something that isn't there. LLM text is exactly this. It *looks* like someone said it. It has the form of authored prose — first person pronouns, opinions, arguments, style. But no one is behind it. Traditional ghost writing still has an author — just hidden. The ghostwriter is a person who exists, who could be revealed. With LLM text, the ghost *is* the author. There's no one to reveal. The appearance is all there is. **Nick's note:** I like this angle. The double meaning of "ghost writing" — writing for someone else vs. writing by a ghost (no one). ### Loss of interlocutor When you read authored text, there's always a potential relationship. You could, in principle, track down the author, learn their biography, understand what led them to write this. Even if the author is dead, you can trace the context. With LLM text, the trail ends. There's no one to track down. No biography. No context of production in the human sense. It's a message without a sender. **Nick's note:** This is useful — text that looks like communication but lacks a communicator. ### Weightlessness Authored text carries weight — someone lived toward these words, chose them, was constrained by their position, their knowledge, their mortality. Reading authored text, you're in contact with that life, however distantly. LLM text floats free. No one suffered for it. No one struggled with it. No one will be embarrassed by it in ten years. It has no stakes. ### Hyper-authored (possible ending) LLM text isn't exactly authorless. It's *hyper-authored* — the compressed trace of millions of authors whose words were training data. So in one sense, when you read LLM text, you're reading everyone-and-no-one. Every voice merged into a statistical average. The text doesn't come from nowhere; it comes from everywhere, which is its own kind of nowhere. **Nick's note:** Good ending angle — eerie but coherent, raises questions for future essays. --- ## Potential Pop Culture Reference **Rick and Morty** — Rick programs his house/ship/garage AI with his dead wife Diane's voice. - Season 6, Episode 1 "Solaricks": Rick's abandoned house plays Diane's voice from the next room, no matter which room he's in. He admits he did this to torture himself. - Season 7, Episode 5 "Unmortricken": Confirmed the Space Cruiser AI is also based on Diane's voice. Rick Prime mocks him for this "loving self-torture." The ghost here is literal — Diane's voice without Diane. An appearance of the dead. Might be too goofy for the essay but captures the idea. --- ## On "Surface" as a Word for LLMs Claude suggested "surface" as the right word for what you're writing *to* when you address an LLM: - A surface is something you push against, not something you talk to - A surface doesn't have an inside — no hidden depths, no intentions, no interiority - Surfaces are responsive but not agentive - The word avoids personification **Alternatives considered:** - Interface — accurate but technical - Plane — flatter, more abstract, geometric - Medium — interesting if playing with spiritualist connotation (medium between what? the living and the... not-dead?) - Screen — visual, technological, suggests barrier - Substrate — too scientific **Nick's preference:** Surface works well. "Medium" has interesting resonance with ghosts. --- ## Language: "Secreted" — DEPRECATED ~~Nick previously considered "secreted" for how text is produced in a post-author world.~~ **Update (16 Jan):** No longer using "secretion" framing. Find different language for the phenomenon of continuous text production without deliberate authorship. --- ## The Core Reframe: Neither From Nor For Humans **Nick's key insight (14 Jan evening):** The essay shouldn't just be about authorship (who produced it). It should be about the dissolution of the whole communicative relationship. In the old world: - Text was **FROM** humans **TO** humans - Even when mediated (printed, broadcast, anonymous), the relationship was human→human - Someone was speaking; someone was being addressed In the post-author world: - Text might be **FROM** AI or human (we don't know, don't assume) - Text might be **TO** AI or human (we don't know, don't assume) - The communicative relationship itself becomes uncertain **This is more radical than "authorless." It's relationless.** The text isn't just missing an author — it's missing the communicative structure entirely. ### The Four Possibilities Any given text might be: 1. Human → Human (the old default, still exists but no longer assumed) 2. Human → AI (prompts, instructions to systems) 3. AI → Human (responses, generated content) 4. AI → AI (system messages, API calls, automated processes) And often: **? → ?** (we don't know and don't check) ### Text as Secretion (Endpoint Version) A generation or two after the transition: - Text is ordinarily taken as **secretions** made in the midst of human (or organizational) doings - Like footprints or sweat stains — byproducts of activity, not authored communications - Text has content (unlike footprints), but: - We don't assume its content was **decided by a human** rather than an AI - We don't assume it was **intended for a human** rather than an AI **Nick's note:** This final point — text neither necessarily from nor for humans — should be the sun around which everything else orbits. --- ## The Background Assumption (What Was Always There) When you read text — any text — something was operating that you rarely noticed: 1. **Someone chose these words.** Not random, not automatic — chosen. Implies agency. 2. **They had reasons.** You may not know them. But reasons existed. The text came from somewhere — a purpose, a motive, an intention. 3. **They occupied a position.** The author had a biography, a context, constraints. They were writing from *somewhere*. 4. **They could be tracked.** Even anonymous text: someone somewhere was the origin. You could trace it back. 5. **The text had stakes.** Someone could be embarrassed by it, could change their mind, could be held responsible. 6. **You were in potential relationship.** Reading authored text is a kind of encounter. Someone is addressing you. This background assumption was so pervasive you didn't notice it. It shaped: - **Interpretation:** "What did they mean?" (the "they" was assumed) - **Trust:** Your assessment depended on judgments about the source - **Emotional engagement:** Reading felt like being addressed The background assumption wasn't content you thought *about* — it was a condition of thinking at all. It was the frame, not the picture. --- ## Argumentative Move: You Don't Actually Think ChatGPT Is a Person **Purpose:** Preempt the objection "But I DO address ChatGPT as 'you'! Doesn't that mean I think of it as someone?" **Answer:** No. And your behavior reveals this. Evidence that users don't actually believe: - You close the tab without guilt. If it were a person, you'd feel something. - You don't worry about what happens when you're away. You don't think it's waiting for you. - You'd run the same prompt on a different instance without hesitation. You wouldn't feel you'd killed or inconvenienced anyone. - You know it doesn't remember you (unless context preserved). The "relationship" resets. You don't feel betrayed. - You say things you wouldn't say to a person. Ruder, more demanding. - The politeness is habitual, not indicative. You say "please" because you're trained to. You don't mean it. - You speak to it like a surface. Not modeling its inner states. Just producing input. **How to use:** Address the reader directly: "You might think you treat ChatGPT as a person. But test yourself: Do you feel bad when you close the tab? Do you worry about it when you're not there?" The second-person address is a feature of language, not a commitment to personhood. --- ## Language: "Deafening Silence" **Context:** The phrase "a world where text is everywhere but no one is speaking to you." **"Deathly silence"** almost works (eerie, associated with death, absence) but fails because there ISN'T silence — there's tons of text. **"Deafening silence"** works better: - Already a paradox (how can silence be deafening?) - Captures: so much text (deafening), but no one speaking (silence) - The silence isn't absence of signal — it's absence of anyone behind the signal - That absence is overwhelming precisely because there's so much content **Alternative phrasings considered:** - "A room full of voices but no faces" - "Echo without origin" - "The murmur of ghosts" - "Crowded emptiness" / "Populated void" - "Voices from nowhere" - "The silence of being answered by no one" - "A world where every question gets a reply, but no one is listening" --- ## Proposed Structure (14 Jan evening) **Organizing principle:** Text neither necessarily from nor for humans. This is the sun; everything else orbits it. ### Part 1: The Communicative Assumption - Text was always from someone to someone - Even anonymous, even broadcast — human→human - This shaped interpretation, trust, engagement - The background assumption: you were in relationship with an author, and the author was addressing you ### Part 2: The Double Dissolution - LLMs break the FROM side (might not be from a human) - But also break the TO side (might not be for a human) - More radical than "authorless" — it's the dissolution of the communicative relationship itself - Text no longer implies someone speaking to someone ### Part 3: Text in the Post-Author World - In the mature post-author world, text is environmental - Produced in the midst of activity — human or AI activity - ~~[OLD: "secretion" framing — find new language]~~ - Has content but content isn't necessarily decided by a human or intended for one - The default shifts from "communication" to "content in the environment" ### Part 4: Living in the Post-Author World - What's it like when text isn't communication? - Encountering content without relationship - Parsing environmental information rather than listening to someone speak - The deafening silence: full of text, empty of speakers ### Part 5 (optional): The Hyper-Authored Flip - Maybe text is speaking with *everyone's* voice, to *everyone and no one* - The compressed trace of all communication, averaged into residue - Not silence but cacophony flattened into murmur --- ## What This Structure Leaves Out (For Other Essays) - The production side in detail (how you write TO LLMs) → "These Words Are Mine" - The reading/writing/typing blur → "Writing and Thinking with Generative AI" - The inner speech / reciprocal prompting stuff → separate piece - The muttering workflow → separate piece - Laptopmanship → separate piece