v42d — Which task binds? Where did the AI surplus go?

三个 Nobel-track question，DDD 给三个 clean answer

Question: 在 (upstream AI firms / platforms / new entrants / incumbents / consumers) 这 5 个 channel 里，AI shock 的 surplus 最终累积到谁？

Test: DDD across (substitutable × non-substitutable bottleneck) on 3 outcomes (cohort / incumbent / total reviews)。

不在 entrants 那里。 Substitutable bottleneck 的新 entrants β_DDD = −0.45 (p=0.025).
不在 incumbents 那里。 β_DDD = −0.10 (p=0.57).
不在 total market pie 里。 β_DDD = −0.11 (p=0.63).
→ 这 3 个 within-iOS channel 都 null/negative → AI surplus 不在 iOS marketplace 里被 captured。最 likely 流向 upstream foundation model providers 或 unmeasured consumer surplus。

Question: Acemoglu framework 说生产是 task bundle。哪个 task 是 binding factor，AI 替代它就会显著影响 output。哪个 task 不 binding，AI 替代它就 no effect。这是 empirical question that theory cannot pin down.

Test: 比较 AI-substitutable bottlenecks (self_contained_code) vs AI-non-substitutable (commerce_physical, network_effects, regulated, physical_device, enterprise) 的 differential effect。

Code 不是 binding task at marketplace level.
AI 极大放松了 code-production task constraint (paper +48% entry)，但在同一批可替代 bottleneck 里: new entrants 失败了 (β=−0.45), incumbents 没受伤 (β=−0.10 null)。
→ Binding task 在 marketplace level 不是 code production，而是 consumer discovery / attention allocation / trust。 AI 让 supply side 暴涨但 demand-side matching 没跟上。

Question: Classical Schumpeter: 新 entrants 替代 incumbents (zero-sum). Akcigit-Kerr (REStud 2010): heterogeneous — some entrants displace, others coexist as fringe.

Test: DDD on incumbent reviews — 如果新 AI-enabled entrants displace incumbents in substitutable markets，应该看到 incumbent β_DDD < 0 显著负。

No Schumpeterian destruction. β_{incumbent_DDD} = −0.10 (p=0.57) — incumbents 在 substitutable bottleneck 里完全没受 differential effect。
这是 Akcigit-Kerr "fringe entrants" 的极端 case: 大量进场，但 zero displacement of incumbents。 Creative addition without destruction — 跟 strict Schumpeterian framework 矛盾。

DDD identification — 为什么这个是 clean 的

Spec

y_bwc ~ β · (treated_c × post_w × code_sufficient_b) | bottleneck×cycle + bottleneck×event_week + cycle×event_week

三个 two-way FE 吸收所有可能的 confounder: cycle-wide YoY drift in iOS reviews (通过 cycle×event_week)， bottleneck-level structural differences (通过 bottleneck×cycle)， bottleneck-specific seasonal patterns (通过 bottleneck×event_week)。只有 triple-interaction 残余，identification clean。

Placebo passes

把 2024 当 fake treated vs 2023 placebo control 跑同样 DDD spec:

log_total β_DDD = +0.042 (p=0.72) null ✓
log_cohort β_DDD = −0.022 (p=0.81) null ✓
log_incumbent β_DDD = +0.058 (p=0.64) null ✓

Parallel-trends across bottleneck × cycle holds → DDD identification 干净，唯一显著 β 是真 shock effect。

主表 — DDD 系数全集

Panel	Outcome	Contrast	β_DDD	SE	p	Verdict
Cell	log_total_rev	STACKED (T2025)	−0.110	0.213	0.63	null
Cell	log_cohort_rev	STACKED (T2025)	−0.450	0.151	0.025	显著负
Cell	log_incumbent_rev	STACKED (T2025)	−0.097	0.163	0.57	null
Cell	log_cohort_rev	Placebo C2024 vs C2023	−0.022	0.087	0.81	null ✓
Cell	log_total_rev	Placebo C2024 vs C2023	+0.042	0.114	0.72	null ✓
Cell	log_incumbent_rev	Placebo C2024 vs C2023	+0.058	0.117	0.64	null ✓
App	log_rev	STACKED Nov 24 (T2025)	−0.0071	0.0029	0.014	显著负
App	log_rev	Placebo Nov 24, C2024 vs C2023	+0.0014	0.0064	0.83	null ✓
App	log_rev	STACKED May 22 (T2025)	+0.0005	0.0018	0.79	null

CS = self_contained_code (AI-substitutable); non-CS = {commerce_physical_local, network_effects, regulated_trust_sensitive, physical_device_integration, enterprise_integration}; excluded: content_rights_curation, other_mixed (ambiguous). Nov 24 design: cohort Jun-Jul 2025, track Aug-Apr, cutoff Nov 24. May 22 design: cohort Oct-Nov 2024, track Jan-Sep, cutoff May 22.

Implication — Literature 怎么 read 这个

Direct contribution to Acemoglu 2024 framework. Acemoglu task-bottleneck theory 说 "binding task identification 是 empirical question that theory cannot pin down"。 v42d 提供干净的 marketplace-scale 实证: 在 iOS App Store，code-production 不是 binding task。 AI 让 code-substitutable 市场入场 +48%，但同 market 的 cohort 自身 review 显著少 (−45%)，incumbents 完全没动。 Binding 在 demand-side matching (discovery, attention, trust)，不在 supply-side code production。

Direct contribution to Tirole 2014 surplus migration question. 所有 3 个 within-iOS surplus channel (新 entrants / incumbents / total market) 都 null 或显著负。 AI productivity dividend 没在 iOS marketplace 里被 capture。 剩下两个 channel 是 upstream foundation model providers (Anthropic, OpenAI, Google) 或 consumer (lower prices, more variety)。这给后续 paper 一个清晰的 follow-up: 用 model-provider API revenue 或 consumer-side WTP 实证补完 surplus migration map。

Direct exculpation of foundation-model-provider competitive-harm concern. Getty / Athey-Morton "Double Harm" 假设: AI 训练数据 → AI-generated competing supply → 损害 incumbent revenue。 v42d 在 iOS App Store 这个 setting 上直接 reject 这个 mechanism: incumbents 在 substitutable bottlenecks 里完全没受 differential effect (β=−0.10, p=0.57)。显著负 effect 在 新 AI-enabled entrants 自身 (β=−0.45) — 是新供给质量不足以 capture attention，不是抢 incumbent 的 attention。这是直接 marketplace evidence。Caveats: reviews ≠ revenue, iOS only, 6-month window.

Tension with strict Schumpeter / failed creative destruction. Schumpeter: 新 entrants 替代 incumbents (creative destruction)。v42d: 大量 entrants (+48%), zero displacement of incumbents (incumbent β null)。跟 Akcigit-Kerr (REStud 2010) "heterogeneous creative destruction" 的 fringe-entrant 极端 case 一致。需要 follow-up: 在何种 condition 下 entrant 会 displace？(retention / quality threshold / network effect / regulatory moat)

Visual — Placebo passes, 唯一显著 β 是 cohort underperformance

蓝 = T2025 stacked DDD. 灰 = within-control placebo (C2024 vs C2023) DDD. 所有 3 个 outcome 的 placebo 都接近 0 且 not significant → parallel-trends across bottleneck × cycle 成立。唯一显著 effect 是新 entrants 的 review 累积 (cohort, p=0.025) — 不是 incumbents 被抢 (null)，不是市场扩张 (null)。

AI 把 code 的 task constraint 拿掉了。但 discovery 还 binds — 所以 surplus 没动。