Scale over Preference: The Impact of AI-Generated Content on Online Content Ecology
Authors: Tianhao Shi, Yang Zhang, Xiaoyan Zhao, Fengbin Zhu, Chenyi Lei, Han Li, Wenwu Ou, Tian Yang, Yang Song, Yongdong Zhang, Fuli Feng
Organizations: University of Science and Technology of China, Hefei, China. · National University of Singapore, Singapore. · Kuaishou Technology, Beijing, China. · The Chinese University of Hong Kong, Hong Kong, China.
The rapid proliferation of Artificial Intelligence-Generated Content (AIGC) is fundamentally restructuring online content ecologies, necessitating a rigorous examination of its behavioral and distributional implications. Leveraging a comprehensive longitudinal dataset comprising tens of millions of users from a leading Chinese video-sharing platform, this study elucidated the distinct creation and consumption behaviors characterizing AIGC versus Human-Generated Content (HGC). We identified a prevalent scale-over-preference dynamic, wherein AIGC creators achieve aggregate engagement comparable to HGC creators through high-volume production, despite a marked consumer preference for HGC. Deeper analysis uncovered the ability of the algorithmic content distribution mechanism in moderating these competing interests regarding AIGC. These findings advocated for the implementation of AIGC-sensitive distribution algorithms and precise governance frameworks to ensure the long-term health of the online content platforms.
Figures & tables
Figure 1: The scale-over-preference dynamic of AIGC in content creation and consumption. (a) Integration of AIGC into the studied platform. The studied platform has invested heavily in generative AI, with daily expenditures on AIGC reaching 4 million USD across the entire platform Kuaishou Technology (2025) . Specifically within its Local Life channel, about 35% of creators have adopted AIGC tools, and AIGC videos account for 30% of total uploads on average. (b) Conceptual framework. Schematic illustration of how AIGC reshapes the content ecology through changes in creator production, consumer engagement, and algorithmic distribution. (c) Creator production and engagement returns. Complementary cumulative distribution function (CCDF) plots for matched creators ( n=2,497 pairs) show that AIGC creators exhibit substantially higher productivity (rightward shift in video volume; left panel) while achieving aggregate engagement returns—measured by total valid views (middle) and full views (right)—comparable to those of HGC creators. Valid views denote views exceeding the platform duration threshold, and full views denote videos watched to completion. (d) Consumer engagement preference. Matched interaction pairs ( n=47,288 pairs ) reveal systematically lower engagement with AIGC than HGC. AIGC interactions yield lower mean valid-view and full-view rates (left) and shorter view duration (right). Left subpanels report means with 95% CIs; right boxplots show median (line), mean (triangle), IQR (box), and whiskers at 1.5× IQR. (e) Scale-over-Preference index (SoPI) for each day. Daily observations ( n=61 days) map relative AIGC supply scale ( S ) against relative consumer preference on AIGC ( P ), colored by SoP index SoPI=ln(S/P) . Most days fall in the high-supply/low-preference region above the SoPI=ln(1.5) contour (dashed), indicating a persistent divergence between production scale and observed preference.
Figure 2: Algorithmic content distribution mechanisms moderate the scale-over-preference dynamic. (a) Algorithmic exposure disadvantage and compressed lifecycle of AIGC. CCDFs for matched video pairs ( n=178,854 pairs) show that AIGC receives lower cumulative exposure than HGC (left) and exhibits a more compressed exposure lifecycle (right; days to reach 90% of 31-day exposure), consistent with weaker engagement preference for AIGC. (b) Algorithmic exposure response to AIGC supply expansion. Granger analysis (left) with all negative lag coefficients indicates that increases in daily AIGC supply precede reductions in the median exposure of newly uploaded AIGC ( p=0.022 ; joint F-test). Regressions across exposure tiers (right) show that higher supply is associated with a downward shift in the exposure distribution: a positive association in the lowest tier ( ≤ 10 views) and negative or insignificant associations in higher tiers (11–100, 101–1000, > 1000 views). (c) Algorithmic exposure response under varying SoP dynamics . Dynamic regressions quantify exposure responses to rising tension between scale and preference. Conditional on relative preference for AIGC, AIGC exposure declines with increasing relative supply ( β = − 0.936, 95% CI = [ − 1.089, − 0.794]; left). Conditional on AIGC relative supply, AIGC exposure decreases with lower relative preference ( β = 2.683, 95% CI = [1.618, 3.696]; right). Both indicate the algorithm’s moderation of AIGC visibility as SoP intensifies. (d) Creator outcomes under algorithmic adjustment. Log-log elasticities show that the overall creator engagement return elasticity with respect to AIGC supply is substantially smaller than for HGC, indicating smaller marginal returns to AIGC expansion ( SOPI↑ ). Bars denote coefficients and error bars 95%. (e) Consumer outcomes under algorithmic adjustment. Log-log regression elasticities of overall consumer engagement depth remain near zero as AIGC supply increases, indicating that overall consumer engagement does not decline under algorithmic moderation of AIGC expansion ( SOPI↑ ). Bars denote coefficients and error bars represent 95% CIs, and HGC results are reported for reference. (f) Heterogeneous responses across algorithms. CCDFs of AIGC exposure ratios under two mechanisms show a leftward shift for the population feedback-driven algorithm relative to the individual feedback-driven algorithm, indicating that algorithm design differentially moderates the SoP dynamic.
Content Type
Mean
Std. Dev.
P25
Median
P75
AIGC Video
20.69
7.05
17.80
20.11
22.77
HGC Video
40.75
53.42
15.97
25.80
49.53
Supplementary Table 1: Statistics of video duration (seconds). This table presents the physical duration of the videos in our sample after deduplicating the interaction logs. The duration is measured in seconds.
Group
Feature
AIGC Creator
HGC Creator
SMD
VR
Before Matching
Creator Follower Count
31.33
32.58
-0.0026
0.57
Creator Video Uploads (14 days)
19.96
4.98
0.1401
19.31
Creator Video Uploads (60 days)
24.64
10.57
0.0624
17.61
Creator Total Play Hours (14 days)
21.97
6.92
0.0801
1.31
Creator Total Play Hours (60 days)
35.30
19.71
0.0412
1.36
Creator Total Play Count (14 days)
11044.63
1670.50
0.1927
5.72
Supplementary Table 2: Covariate balance test for creation-side matching (numerical features) . To accurately assess changes in creator behavior, this matching ensures that AIGC creators and HGC creators had comparable baseline productivity (e.g., historical upload volumes) and influence (e.g., follower count, historical play counts) prior to AIGC adoption. The ”AIGC Creator” and ”HGC Creator” columns report the mean values for the respective groups. SMD stands for Standardized Mean Difference. VR stands for Variance Ratio.
Group
Feature
AIGC
HGC
SMD
VR
Before Matching
Creator Follower Count
935.20
1490.10
-0.0731
0.39
Creator Video Uploads (14 days)
196.70
318.51
-0.133
0.33
Creator Video Uploads (60 days)
587.01
1141.83
-0.1629
0.25
Creator Total Play Hours (14 days)
242.32
2177.65
-0.4238
0.04
Creator Total Play Hours (60 days)
686.40
7295.51
-0.4088
0.03
Creator Total Play Count (14 days)
150476.74
1374554.76
-0.3991
0.04
Supplementary Table 3: Covariate balance test for consumption-side matching (numerical features). To accurately measure consumer engagement, this interaction-level matching controls for both the historical viewing preferences of the consumer and the historical performance of the viewed video’s creator. The ”AIGC” and ”HGC” columns report the mean values for the respective interaction groups. SMD stands for Standardized Mean Difference. VR stands for Variance Ratio.
Group
Feature
AIGC Videos
HGC Videos
SMD
VR
Before Matching
Creator Follower Count
594.53
2314.72
-0.2587
0.14
Creator Video Uploads (14 days)
197.94
709.72
-0.3679
0.08
Creator Video Uploads (60 days)
591.63
2642.00
-0.3754
0.05
Creator Total Play Hours (14 days)
178.12
982.94
-0.3098
0.09
Creator Total Play Hours (60 days)
539.79
3887.15
-0.3393
0.07
Creator Total Play Count (14 days)
104990.54
346783.10
-0.2180
0.22
Supplementary Table 4: Covariate balance test for algorithmic distribution comparison (numerical features) . To fairly compare the algorithmic exposure allocated to different content, this video-level matching ensures that AIGC videos and HGC videos shared comparable pre-upload creator characteristics (e.g., historical upload volumes and play counts). The ”AIGC Video” and ”HGC Video” columns report the mean values for the respective groups. SMD stands for Standardized Mean Difference. VR stands for Variance Ratio.
Group
Feature
Indiv. Feed.
Pop. Feed.
SMD
VR
Before Matching
Consumer AIGC Play Hours (14 days)
5.38
4.11
0.0358
0.40
Consumer HGC Play Hours (14 days)
77.08
73.73
0.0153
1.09
Consumer AIGC Play Count (14 days)
4.84
4.54
0.0060
0.06
Consumer HGC Play Count (14 days)
26.01
45.66
-0.2454
0.29
Consumer AIGC Play Hours (60 days)
13.57
12.75
0.0061
0.20
Consumer HGC Play Hours (60 days)
276.38
310.71
-0.0552
0.53
Supplementary Table 5: Covariate balance test for algorithmic mechanism comparison (numerical features) . To robustly compare the two distribution mechanisms, this consumer-level matching ensures that consumers in both algorithm groups had comparable historical consumption behaviors ( e.g. , historical play hours and play counts). The ”Indiv. Feed.” (short for “individual feedback”) and ”Pop. Feed.” (short for “population feedback”) columns report the mean values for the respective consumer groups receiving the algorithm based on individual/population-level feedback. SMD stands for Standardized Mean Difference. VR stands for Variance Ratio.
Outcome Variable
Coef.
RY∼D∣X2
RVq=1
RVq=1,α=0.05
RD∼Zk∣X2
RY∼Zk∣X,D2
Video Creation Volume (log1p)
0.5548 ∗∗∗
0.0862
0.2635
0.2425
0.0488
0.0224
Total Valid Views (log1p)
0.0836
0.0004
0.0198
0.0000
0.0488
0.0105
Total Full Views (log1p)
-0.1484 ∗
0.0013
0.0355
0.0077
0.0488
0.0114
Supplementary Table 6: Sensitivity analysis for creation-side behavioral outcomes. Coef. reports the regression coefficient of the treatment (creator’s adoption of AIGC tools) for each respective outcome (creator behavioral metrics). RY∼D∣X2 denotes the proportion of residual outcome variance explained by the treatment. RVq=1 and RVq=1,α=0.05 represent the robustness values required for an unobserved confounder to reduce the estimate to zero or render it statistically insignificant at the 5% level, respectively. The reference columns report the maximum explanatory power of the observed covariates ( Zk ) on the treatment ( RD∼Zk∣X2 ) and the outcome ( RY∼Zk∣X,D2 ), identified by iterating through all candidates to establish an empirical benchmark. *** p<0.001 , ** p<0.01 , * p<0.05 .
Outcome Variable
Coef.
RY∼D∣X2
RVq=1
RVq=1,α=0.05
RD∼Zk∣X2
RY∼Zk∣X,D2
Valid-View Rate
-0.0767 ∗∗∗
0.0188
0.1292
0.1236
0.0013
0.0010
Full-View Rate
-0.0266 ∗∗∗
0.0065
0.0774
0.0715
0.0013
0.0008
View Duration (s)
-2.2292 ∗∗∗
0.0071
0.0810
0.0751
0.0013
0.0006
Supplementary Table 7: Sensitivity analysis for consumption-side engagement outcomes. Coef. reports the estimated regression coefficient of the treatment (the viewed video being AIGC) for each respective outcome (consumer engagement depth). RY∼D∣X2 denotes the proportion of residual outcome variance explained by the treatment. RVq=1 and RVq=1,α=0.05 represent the robustness values required for an unobserved confounder to reduce the estimate to zero or render it statistically insignificant at the 5% level, respectively. The reference columns report the maximum explanatory power of the observed covariates ( Zk ) on the treatment ( RD∼Zk∣X2 ) and the outcome ( RY∼Zk∣X,D2 ), identified by iterating through all candidates to establish an empirical benchmark. *** p<0.001 , ** p<0.01 , * p<0.05 .
Outcome Variable
Coef.
RY∼D∣X2
RVq=1
RVq=1,α=0.05
RD∼Zk∣X2
RY∼Zk∣X,D2
31-day Cumulative Exposure (log1p)
-1.7910 ∗∗∗
0.1330
0.3224
0.3202
0.0005
0.0021
Days to Reach 90% 31-day Exposure
-5.3077 ∗∗∗
0.1438
0.3344
0.3322
0.0005
0.0005
Supplementary Table 8: Sensitivity analysis for algorithmic distribution outcomes. Coef. reports the regression coefficient of the treatment (the video being AIGC) for each respective outcome(algorithmic exposure metrics). RY∼D∣X2 denotes the proportion of residual outcome variance explained by the treatment. RVq=1 and RVq=1,α=0.05 represent the robustness values required for an unobserved confounder to reduce the estimate to zero or render it statistically insignificant at the 5% level, respectively. The reference columns report the maximum explanatory power of the observed covariates ( Zk ) on the treatment ( RD∼Zk∣X2 ) and the outcome ( RY∼Zk∣X,D2 ), identified by iterating through all candidates to establish an empirical benchmark. *** p<0.001 , ** p<0.01 , * p<0.05 .
Outcome Variable
Coef.
RY∼D∣X2
RVq=1
RVq=1,α=0.05
RD∼Zk∣X2
RY∼Zk∣X,D2
AIGC Exposure Share
0.0202 ∗∗∗
0.0098
0.0948
0.0756
0.0210
0.0050
Supplementary Table 9: Sensitivity analysis for algorithmic mechanism comparison. Coef. reports the regression coefficient of the treatment (distribution algorithm type: 1 = individual feedback-driven, 0 = population feedback-driven) for the outcome (AIGC exposure share). RY∼D∣X2 denotes the proportion of residual outcome variance explained by the treatment. RVq=1 and RVq=1,α=0.05 represent the robustness values required for an unobserved confounder to reduce the estimate to zero or render it statistically insignificant at the 5% level, respectively. The reference columns report the maximum explanatory power of the observed covariates ( Zk ) on the treatment ( RD∼Zk∣X2 ) and the outcome ( RY∼Zk∣X,D2 ), identified by iterating through all candidates to establish an empirical benchmark. *** p<0.001 , ** p<0.01 , * p<0.05 .
Supplementary Fig. 1: Decomposition of creation volume and engagement returns for AIGC creators in creator-side comparison. The complementary cumulative distribution functions (CCDFs) compare the production and engagement metrics between HGC creators (solid blue line) and AIGC creators (solid green line). To isolate the source of AIGC creators’ productivity and engagement, their metrics are further decomposed into their AIGC video contributions (dashed green line) and HGC video contributions (dashed blue line). The panels illustrate (left) Creation Volume, (middle) Valid-View Count, and (right) Full-View Count.
Variables (First Difference)
ADF Test Statistic
KPSS Test Statistic
Median AIGC Exposure (31-day)
-15.76 ( <0.001 )***
0.06 (0.100)
Median HGC Exposure (31-day)
-7.06 ( <0.001 )***
0.04 (0.100)
Median AIGC Exposure (1-day)
-17.23 ( <0.001 )***
0.10 (0.100)
Median HGC Exposure (1-day)
-11.84 ( <0.001 )***
0.05 (0.100)
AIGC Supply Scale
-14.45 ( <0.001 )***
0.07 (0.100)
HGC Supply Scale
-6.49 ( <0.001 )***
0.09 (0.100)
Supplementary Table 10: Unit root tests for variables in first differences. This table presents the Augmented Dickey-Fuller (ADF) and Kwiatkowski-Phillips-Schmidt-Shin (KPSS) test results for all variables after first differencing. The test statistics are reported with p-values in parentheses. For the ADF test, the rejection of the null hypothesis indicates stationarity; for the KPSS test, the failure to reject the null hypothesis indicates stationarity. The results confirm that all variables are stationary in their first differences, satisfying the prerequisite for the subsequent Granger causality analysis. *** p<0.001 , ** p<0.01 , * p<0.05 .
Lag
Scale → Exposure (1d)
Exposure (1d) → Scale
Scale → Exposure (31d)
F-Stat
P-Value
Sig.
F-Stat
P-Value
Sig.
F-Stat
P-Value
Sig.
Panel A: AIGC
1
0.700
0.404
ns
0.018
0.892
ns
2.787
0.105
ns
2
1.423
0.251
ns
0.318
0.693
ns
3.318
0.052
ns
3
6.212
0.012
0.339
0.752
ns
3.383
0.035
4
5.810
0.009
**
0.479
0.691
ns
3.322
0.022
Supplementary Table 11: Granger causality test results for content supply scale and algorithmic exposure. This table presents the Granger causality tests evaluating the directional relationship between content supply scale and algorithmic exposure (1-day and 31-day cumulative) for newly uploaded AIGC (Panel A) and HGC (Panel B) across multiple time lags (1 to 6 days). The results demonstrate a significant forward causality from AIGC supply scale to exposure with no significant reverse causality, whereas HGC shows no significant relationships. *** p<0.001 , ** p<0.01 , * p<0.05 , ns: not significant.
Exposure Tier
Low
Med.-Low
Med.-High
High
Variables
( ≤ 10 views)
(11-100 views)
(101-1K views)
( > 1K views)
AIGC Supply Scale ( ×106 )
1.43***
-0.02
-1.37***
-0.08***
[1.10, 1.76]
[-0.25, 0.20]
[-1.63, -1.12]
[-0.09, -0.06]
HGC Scale Control
Yes
Yes
Yes
Yes
DOLS Lag Order
0
0
1
0
Shin Test Statistic
1.10
1.08
0.33
0.85
Supplementary Table 12: Dynamic OLS (DOLS) regression estimates: associations between AIGC supply scale and exposure tier distribution. This table reports the associations between the AIGC supply scale and the proportion of newly uploaded AIGC videos entering specific algorithmic exposure tiers. For readability, the estimated coefficients and their 95% bias-corrected and accelerated (BCa) confidence intervals (in brackets) for the AIGC supply scale are multiplied by 106 . The HGC supply scale is included to control for platform-level volume dynamics; its coefficients are omitted for brevity. Optimal DOLS lag orders ( p ) were selected via the Akaike Information Criterion (AIC) Akaike (2003) . The Shin test evaluates residual stationarity; the high p-values (failure to reject the null) across all models confirm stationarity, precluding spurious regression. *** p<0.001 , ** p<0.01 , * p<0.05 .
Variables
Relative Median Exposure (AIGC to HGC)
Relative Consumer Preference for AIGC
2.683***
[1.663, 3.725]
Relative Supply Scale of AIGC
-0.936***
[-1.090, -0.794]
DOLS Lag Order (AIC)
4
Shin Test Statistic
0.138
Supplementary Table 13: Dynamic OLS (DOLS) regression estimates: algorithmic exposure response to scale-over-preference dynamics. This table reports the associations of relative consumer preference and relative supply scale with the relative algorithmic exposure of AIGC to HGC. Consistent with Fig. 2(c) in the main text, the results demonstrate a dual moderation mechanism: holding supply scale constant, algorithmic exposure increases significantly with stronger consumer preference; conversely, holding preference constant, exposure decreases significantly as supply scale expands. Estimated coefficients are presented with their 95% bias-corrected and accelerated (BCa) confidence intervals in brackets. The DOLS lag order was optimally selected via the Akaike Information Criterion (AIC). The Shin test evaluates residual stationarity; the high p-value confirms stationarity, precluding spurious regression. *** p<0.001 , ** p<0.01 , * p<0.05 .
Outcome Types
AIGC Creator Outcomes
HGC Creator Outcomes
Full-View
Valid-View
Full-View
Valid-View
Variables (Log)
Count
Count
Count
Count
Own-type Scale Elasticity
0.191
0.132
0.716***
0.508***
[-0.039, 0.364]
[-0.046, 0.268]
[0.325, 1.092]
[0.211, 0.802]
Cross-type Scale Control
Yes
Yes
Yes
Yes
DOLS Lag Order (AIC)
4
4
0
0
Supplementary Table 14: Dynamic OLS (DOLS) regression estimates: elasticities of supply scale on creator engagement returns. This table reports the elasticities derived from a log-log DOLS specification, corresponding to Fig. 2(d) in the main text. The dependent variables are the aggregate full-view and valid-view counts for AIGC and HGC creators, respectively. The coefficient of interest is the ”Own-type Supply Scale,” representing the elasticity of engagement with respect to the supply of the same content type (e.g., AIGC engagement relative to AIGC supply). The cross-type supply scale (e.g., HGC supply in AIGC models) is included as a control in all specifications. Estimated elasticities are presented with 95% bias-corrected and accelerated (BCa) confidence intervals in brackets. The DOLS lag orders were selected via the Akaike Information Criterion (AIC). The high Shin test p-values confirm residual stationarity across all models. *** p<0.001 , ** p<0.01 , * p<0.05 .
(1) Overall Full-View
(2) Overall Valid-View
Variables (Log)
Rate
Rate
AIGC Supply
0.026
0.025
Scale (Elasticity)
[-0.054, 0.117]
[-0.028, 0.086]
HGC Supply
0.256
0.040
Scale (Elasticity)
[-0.114, 0.632]
[-0.198, 0.301]
DOLS Lag Order (AIC)
0
0
Supplementary Table 15: Dynamic OLS (DOLS) regression estimates: elasticities of supply scale on consumer engagement depth. This table reports the elasticities from a log-log DOLS specification. Dependent variables are the overall full-view and valid-view rates. Both AIGC and HGC supply scales are included to examine their respective associations. Consistent with Fig. 2e in the main text, elasticities for both supply scales are near zero and statistically insignificant, indicating that algorithmic distribution stabilizes user engagement despite massive AIGC expansion. Estimated elasticities are presented with 95% BCa confidence intervals in brackets (constant omitted). Optimal DOLS lag orders were selected via AIC. The Shin test evaluates residual stationarity; high p-values confirm stationarity, precluding spurious regression. *** p<0.001 , ** p<0.01 , * p<0.05 .
With the rapid development of Artificial Intelligence Generated Content (AIGC) platforms, users increasingly show cross-platform usage intentions. Existing research focuses on adoption and usage intentions in single-platform AIGC contexts. A theoretical gap still exists in studies on cross-platform usage. This paper constructs and verifies a three-stage multiple mediation model based on the personality trait-perception-behavioral response framework. The model integrates the optimum stimulation level (OSL) theory, complementarity theory, and perceived value theory, and it sets social influence and use experience as control variables to examine users' multi-homing intention. The results show that: (a) OSL significantly enhances users' perceived complementarity; (b) perceived complementarity positively affects perceived epistemic value; (c) perceived epistemic value significantly and positively predicts multi-homing intention; (d) OSL influences multi-homing intention through a chain mediation path of perceived complementarity and perceived epistemic value; and (e) social influence has a significant positive effect on multi-homing intention, while the effect of use experience is not significant.
Xuchao Zhang, Jihye Lee
School of Design, Pusan National University, Busan, South Korea
Search engines traditionally complement online content platforms by directing users seeking information to external websites. The emergence of generative AI search tools that summarize answers directly on the results page may disrupt this relationship by making visits to source platforms optional. We study this question using Google AI Overviews and Reddit, one of the largest online discussion platforms. Our identification exploits Google's content moderation policy: Safe-for-Work (SFW) Reddit communities are indexed by Google organic search and surfaced in Google AI Overviews, while Not-Safe-for-Work (NSFW) communities, though indexed by organic search, are prohibited from being referenced in AI Overview summaries. Using a difference-in-differences design, we find that AI Overviews increase engagement in SFW communities: daily comments rise by 12.0 percent and the number of commenting users by 12.4 percent relative to NSFW communities. The effects are concentrated in experience-based discussions (opinions, advice, and personal experiences) rather than fact-based information. However, the subsequent introduction of Google AI Mode, which allows users to interact conversationally with the AI summary, largely eliminates these gains in experience-based content. These results suggest that the effects of AI search depend critically on interface design and content type.
Peibo Zhang, Ruomeng Cui, Dennis J. Zhang
Goizueta Business School, Emory University · Olin Business School, Washington University in St. Louis
The rapid adoption of large language models (LLMs) creates new opportunities for strategic content generation on online platforms, including potentially harmful forms of manipulation that may undermine platform effectiveness and reshape platform dynamics. However, measuring such activity is difficult because AI-generated content is rarely directly observable. We introduce an empirical approach that leverages discrete LLM supply shocks - abrupt changes in model prices and capabilities, and contrasts verified with non-verified reviews to identify changes in platform activity associated with generative AI supply improvements. We apply this approach to more than 13 million reviews from Trustpilot, one of the leading online platforms for business reviews. A robust finding is that following LLM supply shocks, unverified reviews shift toward greater negativity: more 1-stars, fewer 5-stars, and lower ratings, with effects driven primarily by new model releases and concentrated among firms with the lowest and highest review volumes, suggesting that strategic AI use may reshape platform competition dynamics. We further find that LLM supply shocks trigger short, concentrated bursts of review activity. Together, these findings suggest that generative AI is already reshaping how reputation and competition operate on online platforms.
Valeria Lerman, Oren Rigbi, Yaniv Dover
The Hebrew University Business School, The Hebrew University of Jerusalem, Jerusalem 9190501, Israel · Department of Cognitive and Brain Sciences, Faculty of Humanities, The Hebrew University of Jerusalem, Jerusalem 9190501, Israel · The Federmann Center for the Study of Rationality, The Hebrew University of Jerusalem, Jerusalem 9190501, Israel