For two years I have watched the AI-widens-contribution story get built into executive decisions. The story has three steps. One, AI lowers the writing cost. Two, more people contribute. Three, more voices get heard. Nobody tested step three.
Last week someone tested step three on 926,739 users. Step three broke.
The participation dashboard in Rotterdam
A VP of engineering at a B2B SaaS firm in Rotterdam, in a cohort session last week. Her team had rolled out an AI-assisted internal Q&A platform six months earlier. The dashboard she showed me had the shape every rollout memo wants. Submission volume up. Contributor headcount tripled. Newcomer-to-veteran ratio healthier than it had been in years.
She was proud of it. Fair enough. Those numbers were real.
Then I asked for the acceptance rate split by cohort. She did not have that view. Her platform lead pulled it in twenty minutes. The newcomers who had tripled the contributor count were getting their answers marked accepted at roughly half the rate of the veterans. The veteran group had answered fewer questions since the rollout. Their acceptance share had gone up.
She stared at the screen for a long time. The platform looked flatter on the dashboard and more vertical in the data.
What the Michigan team measured on Stack Overflow
Ji Eun Kim, Léa Vitale, Libby Hemphill and Yulin Yu ran an interrupted time-series on Stack Overflow across the ChatGPT release. Nov 30 2021 to Nov 29 2023. 926,739 users. 1.8M questions. 2.5M answers. Segmented regressions around the release date.
The casual contributors they call consumers sharply increased their answer production (level change b=+0.175, p<0.001). Their acceptance rate fell (b=−0.020, p<0.05). The established producers wrote fewer answers (b=−0.391, p<0.001) and received greater recognition (b=+0.043, p<0.001).
"Consumers shifted from asking questions to producing answers, but their answers were less likely to be accepted relative to the pre-ChatGPT period, suggesting that increased production did not yield equal standing in the community."
Four distinct outcomes moving the same way. The platform did not get flatter after AI arrived. It got more vertical.
This sits in the same shape as the Sep 22 Human Element piece on admissions essays. Wider AI-assisted production, evaluator-side correction. Different domain, same signal.
If your internal Q&A platform, your developer-community investment, or your crowdsourced-contribution programme measures success by submission volume, your dashboard is now shipping a false positive. The volume number can climb while the acceptance share concentrates further at the top. Measure acceptance by cohort before and after rollout. The audit is cheap. The decisions downstream of a wrong assumption are not.