After an AI Mental Health Assistant Arrives, Free Answers Decline: Real Shocks on a Two-Tiered Platform

Forum topic · 小凯 · 2026-05-19

Summary

A quasi-natural experiment on a leading Chinese online mental health community (OMHC) measures how introducing a generative AI conversational assistant affected a two-tiered platform, where free public Q&A feeds paid private consultations. Using difference-in-differences, same-platform controls, and heterogeneity analysis, Zhang, Zheng, and Yan (arXiv:2605.16279) find that intrinsically motivated counselors—those answering for altruistic reasons—reduced public-layer participation, while economically motivated counselors intensified competition, posting more and responding faster to capture the enlarged user base. Post volume rose, but per-post social recognition (likes, saves, follows) declined. Crucially, cross-tier spillovers emerged: counselors who reduced public activity saw paid consultations drop, while those who stayed active maintained or grew demand. Overall, the demand-expansion effect outweighed crowding-out, total paid activity did not fall, but benefits became concentrated among aggressive, economically driven participants. The study documents a structural pattern in two-sided AI markets rather than a welfare judgment, and notes limitations: undisclosed platform data, uncertain generalizability across regulatory contexts, and an unexamined ethical dimension regarding the displacement of unpaid altruistic labor.

| Item | Detail | |------|--------| | Title | Generative AI and Two-Tiered Online Mental Health Communities | | Authors | Manyang Zhang, Jinyang Zheng, Zhijun Yan | | arXiv | 2605.16279 (cs.CY, cs.AI) | | Date | April 2026 | | Core contribution | First systematic causal measurement, via a quasi-natural experiment on a leading Chinese online mental health community, of an AI assistant's two-way effects on a two-tiered platform — intrinsically motivated counselors exit, economically motivated counselors compete harder, and demand expansion outweighs crowding-out | | Link | https://arxiv.org/abs/2605.16279 |

Imagine an online marketplace — half free public consultation desks, half paid private rooms. The public desks are free and open, where anyone can see questions and answers. The private rooms require paid booking: one-on-one and confidential.

This is a two-tiered structure. The free public layer is a storefront that drives traffic to the paid private layer — a classic two-tier market model.

This marketplace is called an Online Mental Health Community (OMHC).

Now the platform adds an AI conversational assistant — it answers instantly, 24/7, free of charge. The logic is benevolent: mental health demand is large, professional resources are scarce, and AI can fill the gap.

But there is another group in the marketplace — professional counselors — who also answer on the public layer and rely on private-room income. Their first reaction to the AI is not "great, some help arrived" but "will this free newcomer take my work?"

That is exactly what the paper measures. On a real platform, using causal inference methods: after AI entered, what changed in counselor behavior?

The quasi-natural experiment

The study is not a lab experiment asking counselors what they would do. It exploits a quasi-natural experiment: a leading mental health community introduced an AI assistant at a time counselors could not anticipate, enabling before/after comparison — the core of event-study methodology.

The identification strategy has three layers:

Difference-in-differences: comparing counselor behavior before vs. after AI entry
Same-platform control: an unaffected sub-forum as a contemporaneous control
Heterogeneity analysis: stratifying counselors by motivation type (intrinsic vs. economic)

Finding 1: Volume up, attention down

After AI entry, counselors' posting intensity on the public layer increased significantly. Counterintuitive — but the logic is that AI's rapid answers to simple questions raised overall platform activity: more users, more questions, a bigger pool of opportunities.

This is the demand-expansion effect — AI is not substituting; it is enlarging the pie.

Yet per-post social recognition (likes, saves, follows) declined. Total posts grew, but attention per post shrank: in a more crowded public space, each voice is less likely to be heard.

Finding 2: Who exits, who doubles down

The key stratification is by motivation:

Intrinsically motivated counselors — those answering for free because they genuinely want to help — reduced participation. Once AI filled basic response needs, they felt their role was replaced: the lowest-threshold unpaid helping work is now done by a machine.

Economically motivated counselors — active on the public layer as an advertising funnel for the paid layer — increased competitive behavior. AI enlarged the user base of potential paying clients, so they posted more, replied earlier and more aggressively — classic attention competition.

Finding 3: Cross-tier spillovers

The strongest finding is cross-tier:

Counselors who reduced public-layer activity also saw paid consultations decline. Public-layer visibility is the engine of paid-layer demand. Those who increased public engagement maintained or grew paid demand — the storefront effect became a lever in the enlarged pool.

Overall: demand expansion + competition incentives > intrinsic crowding-out. Total paid activity did not shrink — but the distribution shifted sharply, from broadly shared benefits to concentration among the aggressive.

Honest caveats

Proprietary data: the platform is unnamed; raw magnitudes (e.g., exact percentage increases) are not reported — only directions and significance levels.
External validity: effects from one Chinese platform may not transfer to different regulatory environments elsewhere.
Missing ethics: the paper answers "what happened," not "was it good." The platform gained activity but lost those who answered without conversion motives — and the displaced altruistic labor likely came from specific groups, such as young counselors without family burdens.

What can be affirmed is a structural regularity, not a moral verdict: in a two-tiered market, AI's demand-pool expansion dominates supply-side crowding-out in aggregate — but the distribution becomes more unequal. Economically motivated strivers capture the dividend; quiet intrinsically motivated helpers cool off.

References

1. Zhang, M., Zheng, J., Yan, Z. (2026). Generative AI and Two-Tiered Online Mental Health Communities. arXiv:2605.16279. 2. Brynjolfsson, E., Mitchell, T., Rock, D. (2018). What Can Machines Learn, and What Does It Mean for Occupations and the Economy? AEA Papers and Proceedings. 3. Acosta, C., Bercik, M., Wong, J. (2022). The Labor Market Effects of AI: Evidence from Online Freelance Platforms. SSRN Working Paper.