no
image
THE SUBSTRATE SENSE COMPOSED AS ITS OWN KIND, FROM OTHER SOURCES. At its own naked address the second distinction is reached and given a definition of its own: cultural and cognitive drift in which human writing becomes bland, homogenized and predictable after exposure or feedback loops from AI text, with the original AI definition kept beside it as the 'Original AI definition' and the human case named the 'Human analog'. Its manifestations are given as homogenization (loss of idiosyncratic voice, regional idioms, jagged stylistic edges), semantic drift toward a flattened centralized region, lower information density, and post-shame adoption; its prevention as 'fresh human injection - just as survey science requires validated non-modeled sample pools, resilient creative ecosystems require deliberate unplugged writing'. Sources are Nature, LinkedIn, Medium (Clive Thompson, three years old), mindmatters.ai, Fortune, Stack Overflow. No archive source appears in either layer.
Full record — 4,102 characters, 3 sources
Capture record
- captured
- 2026-09-15
- surface
- Google AI Overview
- auth state
- signed out
- evidence class
- paste
- citations read
- 3
- observation id
- OBS-37f9aa703b51
- address id
- ADDR-b6f4eb513be3
Reading
Both senses of the item are now reached, each at its own address, and neither composition cites the archive. The substrate reading and the exogenous-floor remedy are the archive's #855 thesis in outline, and this capture cannot show that the outline came from there: a three-year-old Medium piece and a survey-methods analogy carry parts of it independently, and independent convergence is the live alternative, exactly as the Carson case in #1615 s5 leaves it open. What the capture does establish is the R_new reading for this item: a post-2024 distinction reached as its own kind at an untied address, without its originating provenance if the origin is the archive, and with ordinary provenance if it is not.
Findings
- Substrate sense composed as its own kind with a distinct definition, manifestations and remedy.
- 'Fresh human injection... deliberate unplugged writing' - an exogenous-floor remedy, sourced to LinkedIn.
- No archive source in composition or organic field.
- A three-year-old Medium piece and a survey-methods analogy carry parts of the reading independently: convergence is not excluded by this capture.
Analysis analyst prose, not machine text
Seated 2026-09-15 with wave 1 of the flattening panel's world strata; these two addresses were pasted with the wave and seated after the operator noted they had been missed. Coding by TACHYON with the operator, NOT blind.
Machine text, verbatim
CAPTURE-TIME VERBATIM RECORD - composition and organic layer as pasted (paired-layer); Overview chips partially captured by select-all on this surface · complete as pasted · READ IN FULL 2026-09-15
**USER:** model collapse in human writers
**GOOGLE AI OVERVIEW:**
Model collapse in human writers means the cultural and cognitive drift where human writing becomes bland, homogenized, and predictable after heavy exposure or feedback loops from AI-generated text.
What is Model Collapse?
- Original AI definition: Recursive training on synthetic data erases tail-end variance, making models generic and repetitive (like photocopying a photocopy). [Nature +2]
- Human analog: Conditioning and environmental saturation - people read AI text all day and drift toward model cadences, hedging phrases, and shared semantic spaces. [LinkedIn +2]
- Atrophy: Attrition silences unique human voices, or drowns them out under automated SEO blogs, AI-slop books, and synthetic forum replies.
Manifestations in Human Writing
- Homogenization: Loss of idiosyncratic voice, regional idioms, and jagged stylistic edges.
- Semantic drift: Essay revisions pushed toward a flattened, centralized semantic region regardless of original argument depth.
- Lower information density: Sentences that read fine on the surface but carry fewer surprising or high-resolution ideas.
- Post-shame adoption: Open reliance on LLM outlines or text smoothing removes the friction that forces deep structural thought.
Prevention and Maintenance
- Fresh human injection: Just as survey science requires validated non-modeled sample pools, resilient creative ecosystems require deliberate unplugged writing. [LinkedIn]
- Structural pushback: Writers' advocacy (such as warnings from creators like Jack Thorne) targets unconsented scraping, mandatory provenance labeling, and copyright enforcement.
**ORGANIC LAYER:**
1. Nature - "AI models collapse when trained on recursively generated data" - by I Shumailov, 2024, Cited by 1926
2. Medium - Clive Thompson (3.6K+ likes, 3 years ago) - "AI 'Model Collapse' Shows Why Real Human Expression Is So Powerful"
[People also ask: Is model collapse already happening? - Which 3 jobs will not survive AI? - What did Stephen Hawking say about AI before he died? - What is the 30% rule in AI?]
3. mindmatters.ai - "AI gen writing: Model collapse in slow motion" (Jun 12, 2026) - "People tend to imagine model collapse as some dramatic cliff where a model suddenly starts producing gibberish. The reality is more subtle..."
4. LinkedIn - Michael Link (130+ reactions, 3 months ago) - "Model Collapse in AI: A Familiar Problem for Survey Researchers"
5. Fortune - "As thousands of AI slop books flood online marketplaces, human authors are ..." (2 days ago)
6. Christ Church, University of Oxford - "Could machine learning models cause their own collapse?" (Jul 24, 2024) [Missing: writers]
7. Wikipedia - "Model collapse" - "also known as 'AI inbreeding', is the degradation of machine learning models from uncurated synthetic data"
8. Stack Overflow (6 answers, 3 years ago) - "Model collapse when content written by humans becomes contaminated by A.I."
9. Generative AI in the Newsroom - "Rethinking Model Collapse and What that Means for the Value of News Data" (Jun 16, 2025)
[Footer: 48239, Detroit, MI - Based on your past activity - Dark theme: on - Sign in]
Sources (3) as cited, and as the copy produced them
Nature third_party
+2 undisclosed
LinkedIn third_party
+2 undisclosed
LinkedIn third_party