IC-812InstructGPT (text-davinci-003) reduces content diversity in co-written essays while GPT-3 (davinci) does not, and the effect is attributable to the model's own less diverse text contributions

Vishakh Padmakumar, He He

SourceDoes Writing with Language Models Reduce Content Diversity?

In a controlled experiment with 38 writers producing 100 argumentative essays per condition (solo, GPT-3, InstructGPT) across 10 topics, essays co-written with InstructGPT show significantly higher homogenization (0.1660 vs 0.1536 solo, 0.1578 GPT-3 at key-point level with ROUGE-L) and lower lexical and content diversity. GPT-3 does not produce a statistically significant difference from solo writing. The mechanism is that InstructGPT's own generated text is less diverse (pairwise suggestion similarity 0.20 vs 0.11 ROUGE-L), while user-contributed text remains statistically unchanged regardless of which model assists.

Evidence
correlational
Key metric
Corpus homogenization (key point, ROUGE-L): solo 0.1536, gpt3 0.1578, instructgpt 0.1660; fraction of unique 5-grams: solo 0.991, gpt3 0.988, instructgpt 0.977; pairwise suggestion similarity (ROUGE-L): instructgpt 0.20 vs gpt3 0.11; all instructgpt vs solo/gpt3 differences significant at p < 0.05
Caveat
Experiments use two limited-access OpenAI models; the authors note other models may exhibit different behavior given the RLHF pipeline is highly customized. The study focuses on single interactions; dynamics may change through repeated interactions over time. Participants are native English speakers from the US recruited via Upwork.
Model
InstructGPT, GPT-3 / GPT base
Concepts
Failure mode
Methods
ROUGE-L [eval], BERTScore [eval], Agglomerative Clustering [eval], Self-BLEU [eval]
Related work
Training a Helpful and Harmless Assistant with RLHF (Bai et al. 2022) [context], Training Language Models to Follow Instructions with Human Feedback (Ouyang et al. 2022) [context]
Extraction
automatic-extraction