**Algorithmic Law - Project/ You said: Are we all going to start sounding like this?** Are we all going to start sounding like this? Probably some of us already do AI tools are everywhere. Gmail now finishes your sentences. Your boss drafts a memo with it. Students paste essay prompts through AI. This probably sounds like AI to you! – It’s not, sadly. Humans copy the language they see The "delve" spike is real — and bigger than the illustrative chart Kobak et al., writing in Science Advances last July, went through 15 million PubMed abstracts from 2010 to 2024 arXiv. They found that "delves" showed up in roughly 25 times as many 2024 papers as the pre-LLM trend line would have predicted Language Log. "Showcasing" and "underscores" — close to ninefold. Scientific American ran the same query in raw counts, which makes the thing concrete. In 2020, "delve" appears in 349 papers. In 2023, 2,847. In just the opening months of 2024, 2,630 — a 654 percent increase Scientific American, and most of the year still to come. Kousha and Thelwall pulled from six different scholarly databases and picked twelve LLM-associated words. Between 2022 and 2024, by their count: delve up about 1,500%, underscore about 1,000%, intricate about 700% Springer. Full-text data is wilder still. In PMC, the share of papers using "underscore" six or more times grew by more than 10,000% from 2022 to 2025. Springer [Make a chart for this] "Most articles are written by AI now" Graphite (October 2025) analyzed 65,000 English-language articles published between January 2020 and May 2025 Dataconomy. Before ChatGPT, roughly 10% of new online content was AI-generated. By late 2024, the share had climbed over 40% and briefly surpassed human-written articles before settling near 50-50 Dataconomy; the most recent figure is 52% AI-generated, 48% human Barrett Media. Most of us are reading AI-assisted articles. Only 14% of content ranking in Google searches was AI-generated, and ChatGPT cited human-written articles 82% of the time AI slop is being produced in volume but not read in proportion. Make a chart for this At the very least, since we see it a lot, we can mimic it [insert video] Make a prompt for thisShow moreApr 22 **Claude responded: Here's a prompt you can feed to a video model (Sora, Veo, Runway, etc.**Architected video prompt illustrating linguistic contagion through visual metaphorArchitected video prompt illustrating linguistic contagion through visual metaphor Here's a prompt you can feed to a video model (Sora, Veo, Runway, etc.). The concept: show the contagion visually by having AI-associated words briefly glow as a person writes them, and having that glow fade as the words become normal — i.e., the contagion completes itself on camera. Prompt — "The Glow Fades" Slow push-in on a woman in her mid-30s at a kitchen table, morning light, typing on a laptop with a coffee mug beside her. Over-the-shoulder angle — we see her email draft on screen. She types naturally, small pauses, slight smile. Every few seconds, a word briefly lights up the instant it's written: delve, underscore, crucial, pivotal, tapestry, showcase. The glow is subtle — a thin cyan outline that fades in half a second. She doesn't notice. She's pleased with her phrasing. Match-cut on the glow: the same cyan flicker on a student's handwritten notebook in a library. Cut: a Slack message on a phone on a bus. Cut: subtitles running across a podcast transcript on a tablet. With each cut the glow is dimmer — the words are becoming ordinary. Final shot: back to the woman's laptop. Her paragraph is full of those same words now, but nothing glows. She hits send. The whole paragraph pulses once, faintly, and the screen goes dark. Style: warm, documentary-natural lighting; shallow depth of field; handheld but steady; realistic, not stylized. Sound: keyboard clicks, distant traffic, a kettle ticking. No music. Length: 12–15 seconds. Shorter alt if you want pure typography (no actors): A blank page on screen. Prose types itself out at a natural human pace. Ordinary words appear in soft black; AI-associated words (delve, underscore, tapestry, pivotal, crucial, showcase) flicker briefly in cyan as they land, then settle to black with the rest. The passage grows. The flickers get shorter, weaker, rarer. By the end of the paragraph, an AI-associated word appears and doesn't flicker at all — it just looks normal. Hold. Cursor blinks. 8–10 seconds. Ambient keyboard sound only. Two notes on why this framing: (1) it doesn't tell the viewer contagion is happening — the fading glow makes them feel it; (2) it avoids the lazy visual of a person-turning-into-a-robot, which would undercut the essay's argument that this is subtle, invisible, and already done. You said: A promp for chatgpt codex or claude cowork A promp for chatgpt codex or claude coworkApr 22 **Claude responded: Got it — you want a prompt you can hand to a coding agent to build this whole section into your existing site.**Architected comprehensive coding agent prompt for website expansionArchitected comprehensive coding agent prompt for website expansion Got it — you want a prompt you can hand to a coding agent to build this whole section into your existing site. Here it is, structured so Codex/Cowork has everything it needs to finish without coming back with questions:
`# Task: expand the "Future" tab of the Signs of AI Writing field guide
<section id="future"> tabdrawChart() into #delve-chart.chart-card, .future-body, .future-pullquote,
.future-two-col, .section-head, CSS vars (--accent, --ink, --muted,
--rule, --bg, --card)Replace the current opening paragraphs with:
"Probably some of us already are. The honest answer is: partly yes, partly no, and which part wins depends on choices we make now."
"The yes side first. AI tools are now wedged into every layer of written work. Gmail finishes your sentences. Your boss drafts a memo with a chatbot. Students paste essay prompts through a rewrite-for- clarity tool. Each interaction pushes the prose a little closer to the average. The nudges are tiny. They compound."
Then a short italic muted aside (new .future-aside class, style like
.quiz-intro):
"This paragraph probably sounds like AI to you. It isn't. Humans copy the language they see."
Title: "The 'delve' spike is real — and bigger than the illustrative version suggested" Sub: "Relative frequency of 'delve' in academic writing vs. pre-LLM trend. Multiple studies, multiple corpora."
In demos.js, update DELVE_DATA so the post-2022 spike peaks around 25× (not 2.8×). Rough shape — keep flat ≤2022, then: 2023 ≈ 8×, 2024 ≈ 25× (peak), 2025 ≈ 15× (decline). Update y-axis max to 30, gridlines at 0/10/20/30. Update callout at the 2024 peak to "≈ 25× predicted baseline". Keep the ChatGPT-launch vertical dashed marker.
Below the chart, add a sourced prose graf:
"Kobak et al. (Science Advances, 2024) analyzed 15 million PubMed abstracts from 2010–2024 and found 'delve' appearing in roughly 25 times as many 2024 papers as the pre-LLM trend predicted; 'showcasing' and 'underscores' were close to ninefold. Kousha and Thelwall, across six scholarly databases, put 'delve' up about 1,500% and 'underscore' about 1,000% between 2022 and 2024. In PubMed Central full text, the share of papers using 'underscore' six or more times grew by more than 10,000% from 2022 to 2025."
Insert a second .chart-card after the sourced paragraph above.
Title: "Most of what's being published is now AI-assisted" Sub: "Share of English-language online articles, Jan 2020 – May 2025. Source: Graphite (Oct 2025), 65,000 articles sampled."
Render as two lines (AI-generated in var(--accent), Human-written in
var(--ink) at 0.55 opacity), crossing near late 2024:
Jan 2020 → AI 5, Human 95 Jan 2022 → AI 10, Human 90 Nov 2022 → AI 10, Human 90 ← mark ChatGPT launch w/ dashed accent line Jan 2024 → AI 30, Human 70 Oct 2024 → AI 42, Human 58 May 2025 → AI 52, Human 48
Y-axis 0–100 with "%" suffix. Two legend keys: "AI-generated" and
"Human-written". New function drawAIShareChart() in demos.js, called
from initAll(). Reuse every visual idiom of drawChart() — gridlines
in var(--rule), tick labels in JetBrains Mono 10px var(--muted),
line width 2.4, dots as var(--bg) fill with accent stroke.
Below, add:
"Production is not reading, though. Only 14% of articles ranking in Google results were AI-generated, and ChatGPT itself cites human- written articles about 82% of the time. AI slop is produced in volume but not read in proportion."
After both charts, new .section-head (roman numeral "ii."):
"Contagion — on camera"
Short lead-in: "At the very least, since we read AI prose constantly, we start producing it. Here's what that looks like, sped up."
Then a placeholder container: <div class="video-embed" aria-label="Video placeholder"> <!-- <video src="assets/glow-fades.mp4" autoplay muted loop playsinline></video> --> <span class="video-placeholder-label">VIDEO — THE GLOW FADES · 12s</span> </div>
CSS for .video-embed: 16:9 aspect-ratio box, var(--card) bg,
var(--rule) outer border, 1px dashed var(--accent) inner outline
inset 6px, centered mono label in var(--muted) at the style of
.chart-legend.
drawChart() and drawAIShareChart() once
on load and again on tab-switch to Future (in case widths measured 0
while the tab was hidden)..future-two-col "Two futures"
block untouched below.Edit Signs_of_AI_Writing.html, demos.js, and styles.css in place. Two notes on why it's shaped like this: Agents do much better when the prompt is **structured as an edit plan against real files** than as a brief. So I've named the files, told it which functions to reuse, and given it the exact copy — so it's not paraphrasing your sources or inventing numbers. I've also told it explicitly what *not* to touch (pullquote, two-futures grid) and what *not* to introduce (new fonts, new colors). Agents tend to over-add; a constraint list tames that. If you want me to also pre-generate the drawAIShareChart function and .video-embed` CSS so the agent just has to drop them in, say the word.
Claude is AI and can make mistakes. Please double-check responses.