In open-ended writing, both single-generalist and role-specialized duo LLM assistance improved essay quality over no-AI control, with no significant difference between them¶

Claim · All claims
Evidence · 1 study · causal r? · q3
Currency · newest study 2026 · not yet checked against its sources · not reviewed by a person

Subclaims¶

q3 i? Both LLM conditions outperformed Control on combined essay quality (Single Δ=0.332, t(244)=6.21, p<.0001; Duo Δ=0.234, t(244)=4.21, p=0.0001), and Single and Duo did not significantly differ (p=0.363). → Harsh Kumar 2026

Evidence¶

Harsh Kumar 2026¶

Harsh Kumar, Zi Kang (Jace) Mu, Jonathan Vincentius, Ashton Anderson. (2026). Human Thinking under Plural LLM Assistance: Mathematical Problem Solving and Open-Ended Writing. Working Paper, July 2026, Toronto, Canada. https://arxiv.org/abs/2604.02677 doi:10.48550/arxiv.2604.02677

q3 · i? · causal · r?

Three-condition between-subjects Prolific experiment (247 participants with two usable essays); essay quality was rubric-based, LLM-annotated, and validated against human raters (Spearman ρ=.73–.77); the omnibus OLS was significant (F(2,244)=21.91, R²=0.152).

"Planned pairwise comparisons with Bonferroni correction showed that both LLM conditions outperformed Control (Control vs. Single: Δ= 0.332,𝑡( 244)= 6.21,𝑝<. 0001; Control vs. Duo: Δ= 0.234,𝑡( 244)= 4.21,𝑝= 0.0001)."

Discussion¶

Learner Variables¶

• Writing Skills — outcome: instruction changes it

Related Claims¶

• Open AI collaboration yielded the highest supported writing quality, while bounded AI support yielded the highest independent no-AI Week 8 performance — related
• Essay quality was highest for AI-generated drafts and lowest with no AI, negatively correlated with ownership, suggesting an ownership-quality tradeoff. — related
• Single-LLM writing assistance increased idea-level homogenization across writers, while a role-specialized duo of architecturally distinct LLMs preserved baseline diversity — related
• Adding a second writing agent increased reported cognitive load, and writing confidence did not increase monotonically with support, with the Single condition reporting the lowest confidence — related