Better Technology, Worse Motivation: Generative AI's Mediocrity Trap (ADB EWP 862)
A · Strong evidence An Asian Development Bank and University of Hong Kong team ran a randomised experiment with 219 artists, recorded every work session on video and had blinded expert juries score quality at 15-minute intervals.
Summary based on the full text·Original: about 40 minutes to read
In short
ADB Economics Working Paper 862 (September 2026) by Yvonne Jie Chen, Jie Gong, Jin Li and Zibo Zhao. Illustrators drew one picture without AI and one with free access to an image tool. AI lifted quality at every point in time, but the gains flattened early, so many stopped sooner. For more than a third, the shorter effort outweighed the tool's boost and their final work got worse. The authors call this a mediocrity trap: the tool lowers the return to sustained effort.
Key numbers
- 219professional and student artists in the experiment (168 treatment, 51 control)
- 77%illustrators whose work sessions got shorter with generative AI
- 38%illustrators whose final quality was lower with AI, despite higher quality at every point in time
- 60-minute marknew peak in time spent with AI, where the AI quality curve flattens (vs the 150-minute cap without AI)
As reported by the source.
English·Web page·opens d.repec.org
What it reports
- Generative AI raises the whole quality curve: at any minute, AI-assisted work scores higher than unaided work.
- The AI curve rises fast and flattens soon, so the marginal return to extra time drops earlier.
- 77% of illustrators worked for less time with AI. Without AI, the time distribution peaks at the 150-minute limit; with AI it becomes bimodal, with a second peak around 60 minutes.
- For 38% of illustrators, final quality fell with AI; most of them (32 of the 38 points) also shortened their sessions, which the authors read as a deliberate stop.
- Quality gains were directionally larger for lower-skilled creators, not statistically significant. With AI, the correlation between a person's first and second scores dropped to nearly zero, suggesting AI compresses execution-skill differences and shifts variation toward ideation.
How they got it
219 participants with art degrees or full-time design jobs were recruited in China, August-September 2024, and worked remotely on their own equipment. Within-subject design with a randomised 1:3 control-to-treatment split to absorb order effects; 12 novel-excerpt topics, 2.5-hour cap per task, prize money tied to quality. Each session was recorded; snapshots every 15 minutes and final outputs were scored 0-100 by at least five blinded judges (professors and industry professionals), normalised within judge and topic.
Limits
- About 74% of participants were art and design students, average age 23; results for senior professionals may differ.
- One creative task with a tournament prize; real commissions have different deadlines and pay.
- Short familiarisation (30 minutes) with the AI tool before the task.
- Extending the finding to writing and business reports is the authors' argument, not tested here.
Why it matters for companies
Time saved by AI can come back as lower quality, because people stop when extra effort stops paying visibly. Firms that need top-end work must reward the final stretch explicitly, since the tool makes "good enough" arrive fast. Firms that standardise on "good enough" should redesign jobs so people do not spend time on the flat part of the curve. Measuring AI value by time saved alone will miss this quality drop.
Summary based on the full text. Read the original for the full argument and data.Read the original on d.repec.org
- Tags
- Work and AI, Experiments