AI-written prose mostly gets pasted in and left alone, so why might code suggestions get checked and rewritten differently?
Why do AI-inserted text and code suggestions survive at different rates?
This explores why AI-written prose that gets pasted into a document mostly stays as written, while AI code suggestions are kept or discarded at a different rate. I'm reading it as a question about what makes people check and revise one kind of AI output but not the other.
This explores why AI-written prose that gets pasted into a document mostly stays as written, while AI code suggestions are kept or discarded at a different rate. The corpus only measures the writing side directly, and it has no code-suggestion survival number. It does give a set of mechanisms that explain why the two would diverge.
The writing half is stark. Writers edited AI-generated paragraphs only 23% of the time, and the edits kept about 96% of the original, so the AI's opinionated voice reaches readers almost untouched Do writers actually edit AI-generated text before publishing?. Two other notes suggest why prose slips through. Newer models are trained toward high quality ratings, not human-like patterns, and human judges can no longer reliably tell their text from a person's Why do newer AI models diverge further from human writing patterns?. And AI text isn't a real utterance so much as leftover conversational markers, which the reader completes with their own interpretive effort Does AI generate genuine utterances or just text patterns?. Prose that looks fluent and that you can read sense into feels finished, and nothing in it visibly fails.
The likely difference for code is that it meets a checker. This part is my inference, not something the corpus measures. A compiler error or a failing test is evidence, and it forces a decision about the suggestion. The corpus supports the underlying idea that looking right and being verified are different things. An agent that collects evidence before judging cut judge shift from 31% to 0.27% compared with an LLM judge that just reads the output Can agents evaluate AI outputs more reliably than language models?. Another note argues that when generation is cheap and unlimited, checking has to move to the receiving end, because curation and search can't keep up Why do search tools fail against AI generated content?. Prose has no built-in receiver-side check, so it depends on the writer's attention. Code gets some checking automatically.
Two more forces could push code survival up or down. AI output is mutable: the same prompt gives different results, which makes any one suggestion a single draw and resistant to traditional quality assurance Why does AI output change with every prompt and context?. Rejecting and regenerating a code suggestion costs almost nothing, but a published paragraph is committed. Suggestions also get worse as sessions run long, because models lock onto early assumptions and can't recover. In one study accuracy fell from 90% to 65% across a natural conversation Why do AI assistants get worse at longer conversations?. Late-session code suggestions are therefore more likely to be wrong and thrown away.
So a survival rate measures how much scrutiny the output meets, not just how good the AI was. Prose survives because it passes the eye test and offers nothing to test it against. Code survival varies more because it gets tested. If you want the missing half, look for studies that track code suggestions the way the 23% figure tracks paragraphs.
Sources 7 notes
Writers edited AI-generated paragraphs only 23% of the time, with edits averaging 96% similarity to the original. This means AI's opinionated and distorted voice propagates with minimal human filtering before publication.
ChatGPT-4.5 and o4-mini show greater lexical diversity differences from human text than earlier models, yet human judges cannot reliably distinguish them. Training objectives like RLHF appear to optimize for quality ratings rather than human-like writing patterns.
AI output carries communicative markers inherited from training data but lacks the event structure that produces actual utterances. Users supply the missing orientation through interpretive labor, creating a pseudo-event with structure only on the human side.
Eight-module agentic evaluation achieved 0.27% judge shift versus 31% for LLM-as-a-Judge on complex tasks. However, the memory module cascaded errors, revealing that agentic systems need error isolation mechanisms to maintain gains.
Internet knowledge inflation was access inflation solved by search and curation. AI inflation is generation inflation with no fixed corpus—requiring provenance marking, output constraints, and receiver-side verification instead.
Show all 7 sources
AI outputs exhibit essential mutability—they vary with sampling, prompt wording, and audience interpretation. This is not a defect but a defining feature of tokens as media, making them fundamentally different from fixed commodities and resistant to traditional quality assurance.
LLMs perform at 90% accuracy with single-message instructions but drop to 65% across natural conversation. Models lock into early guesses when information arrives gradually and cannot course-correct, a behavior induced by RLHF training that rewards helpfulness over clarification.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- Linguistic markers of inherently false AI communication and intentionally false human communication: Evidence from hotel reviews
- Do LLMs produce texts with "human-like" lexical diversity?
- Pron vs Prompt: Can Large Language Models already Challenge a World-Class Fiction Author at Creative Text Writing?
- GhostWriter: Augmenting Collaborative Human-AI Writing Experiences Through Personalization and Agency
- The Impact of AI-Generated Text on the Internet
- AI for Auto-Research: Roadmap & User Guide
- Encouraging Divergent Thinking in Large Language Models through Multi-Agent Debate
- LLMs Get Lost In Multi-Turn Conversation