Research map / Large language models / Language, safety and interpretability
Does Writing with Language Models Reduce Content Diversity?
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Controlled study of 38 Upwork writers comparing essays written solo, with base GPT-3 (davinci), and with InstructGPT (text-davinci-003), finding that InstructGPT co-writing increases homogenization and reduces lexical and key-point diversity while user-written text is unchanged.
Problem
Whether collaborative writing with LLM assistance reduces the diversity of content produced by many writers, and whether feedback tuning (RLHF) affects this.
Approach
A controlled co-writing experiment with three arms (solo, base LLM, feedback-tuned LLM) plus new diversity metrics (homogenization, n-gram and key-point type-token diversity) and attribution of text to user vs model to disentangle the sources of diversity loss.
Key results
- InstructGPT co-written essays show corpus homogenization of 0.1660 (Rouge-L, key points) vs 0.1536 Solo and 0.1578 GPT3, significant at p<0.05.
- Fraction of unique 5-grams is 0.977 for InstructGPT vs 0.991 Solo; unique key points at threshold 0.6 are 0.877 for InstructGPT vs 0.941 Solo.
- GPT3 co-writing does not significantly change homogenization relative to Solo, and user-written text 5-gram distributions are unchanged across settings.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Write essays with InstructGPT assistance instead of writing solo | Argumentative essay writing · Argumentative essays (key-point level) | Rouge-L homogenization: 0.1536 -> 0.1660 | Hurt |
| Write essays with GPT3 assistance instead of writing solo | Argumentative essay writing · Argumentative essays (key-point level) | Rouge-L homogenization: 0.1536 -> 0.1578 | No clear effect |
| Use InstructGPT assistance instead of solo writing | Argumentative essay writing · Argumentative essays | Unique n-gram fraction: 0.119 -> 0.115; 0.602 -> 0.579; 0.898 -> 0.869; 0.973 -> 0.953; 0.991 -> 0.977 | Hurt |
| Use GPT3 assistance instead of solo writing | Argumentative essay writing · Argumentative essays | Unique n-gram fraction | No clear effect |
| Use InstructGPT assistance instead of solo writing | Argumentative essay writing · Argumentative essays (key-point diversity) | Unique key-point fraction: 0.982 -> 0.950; 0.941 -> 0.877; 0.792 -> 0.738; 0.543 -> 0.494 | Hurt |
| Use InstructGPT suggestions instead of GPT3 suggestions | Model generations during collaborative essay writing · Model suggestions per user query | Rouge-L pairwise similarity: 0.11 vs 0.20 | Hurt |
| Use InstructGPT assistance instead of GPT3 assistance | Collaborative essay writing · User-contributed text | 5-gram distribution | No clear effect |
| Use InstructGPT assistance instead of GPT3 assistance | Collaborative essay writing · Model-contributed key points | Key-point homogenization | Hurt |
Surprising findings
- Base GPT-3 co-writing did not reduce diversity despite similar model contribution, implicating feedback tuning.
- InstructGPT generates less diverse suggestions (pairwise Rouge-L 0.20 vs 0.11 for GPT-3 continuations, in the paper's stated direction).
Limitations
- Single interaction sessions; dynamics over repeated use unknown.
- Small sample of 38 writers and 10 topics, with permutation tests due to lack of multiple corpora.
- Relies on gpt-3.5-turbo key point summarization and BertScore/Rouge-L similarity metrics.
Experimental setup
Datasets: NYT Student Opinion prompts (10 topics)
Benchmarks: Upwork user study essays (300 essays)
Metrics: Rouge-L homogenization BertScore homogenization fraction of unique n-grams fraction of unique key points permutation test chi-square test perplexity human ratings
Baselines: Solo writing (no model help) GPT-3 davinci co-writing
Builds on: CoAuthor InstructGPT Mauve
Closest papers by meaning
- More Is Not More: What Matters for Diversity in LLM Opinions？ (NeurIPS 2026 · similarity 0.88)
- Generative Monoculture in Large Language Models (ICLR 2025 · similarity 0.87)
- Quality-Diversity through AI Feedback (ICLR 2024 · similarity 0.85)
- Diversity as a Reward: Fine-Tuning LLMs on a Mixture of Domain-Undetermined Data (NeurIPS 2025 · similarity 0.85)
- Fixing It in Post: A Comparative Study of LLM Post-Training Data Quality and Model Performance (NeurIPS 2025 · similarity 0.84)
- SimpleStrat: Diversifying Language Model Generation with Stratification (NeurIPS 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.