AI-led writing produced worse work without saving extra time, study finds

Letting AI take charge of a writing assignment did not buy participants extra speed over more human-directed approaches. Instead, it produced the clearest losses in quality, satisfaction and creative diversity. A study published November 10, 2026, in PNAS Nexus found that the division of creative work mattered even when completion times were similar.

Three ways to divide the same assignment

Kartik Hosanagar and Daehwan Ahn compared three collaboration designs using a 1,000-word creative-writing task. The experiment examined more than the finished story: it also compared how long the work took, how participants felt about the process, and how varied the resulting stories were.

The researchers assigned 285 students to three groups for writing assignments over two days, CNET reported. Everyone worked without AI on the first day. On the second, each group followed a different process for sharing the work with AI.

• Human Confirmation: AI led idea generation, outlining, drafting and editing. People mainly accepted or rejected its output at each stage.
• Human Creativity: People handled planning and outlining, then handed writing and editing to AI. Human involvement was concentrated at the beginning.
• Copilot: People and AI collaborated throughout the writing process, rather than reserving the human role for early planning or approval of AI output.

Faster writing did not mean better writing

Against the first day's unaided writing, AI sped up the task but did not improve quality in any of the three groups, according to CNET. That is a different comparison from the one among AI workflows: all three delivered similar completion times, but their other results diverged.

Human Confirmation showed the clearest downsides. The paper reports lower quality and satisfaction, alongside lower aggregate genre and semantic diversity. Those diversity measures concern the collection of stories: how varied their genres were and how similar their content was, rather than just whether an individual story was good.

Keeping humans active in planning or throughout writing avoided the clearest quality and satisfaction losses without requiring longer overall completion times. Preserving their role before drafting was also associated with less similarity across story content than Human Confirmation.

Approval is not the same as creative control

Hosanagar told CNET that people derive satisfaction from a meaningful creative role, not simply from producing a good final output. In that reading, keeping someone available to approve AI work is not equivalent to letting them shape it.

The scope is specific: students completing creative-writing assignments over two days, not a long-term workplace trial. Avoiding the clearest losses and improving on unaided work are different outcomes.