Early Signals of Problematic AI Use: The Longitudinal Roles of Motivation and Emotional Engagement
Authors: Jina Suh, Constance Hadley, Koustuv Saha
Organizations: CoralMind Institute, University of Washington, Seattle, Washington, USA · Institute for Life at Work, Belmont, Massachusetts, USA · University of Illinois Urbana-Champaign, Urbana, Illinois, USA
Longitudinal evidence about the psychological impact of conversational AI use remains scarce. We report a survey study of 114 adults who reported their level of problematic AI use one year after reporting their motivations (for escape, social, entertainment, and instrumental reasons) for engaging with AI. We found high levels of AI motivation of all types predicted later problematic AI use, especially when participants turned to AI for managing their feelings and overcoming difficulties. Our findings imply that AI motivation could be used as an early warning signal of potential problematic use, especially if individuals engage with AI for emotional purposes over time.
Figures & tables
Construct
Wave(s)
Items
Scale
Rel.
M ( SD )
Role
Primary outcome
Problematic AI use
T2
10
1–5
.87
1.62 (0.68)
Outcome
Focal predictors
Motivation: escape
T0, T1, T2
3
1–4
.94
1.70 (0.93)
Predictor
Motivation: social
T0, T1, T2
3
1–4
.95
1.79 (0.91)
Predictor
Motivation: instrumental
T0, T1, T2
3
1–4
.82
3.44 (0.65)
Predictor
Table 1. Measures used in this paper, including the waves the measure was collected, number of items in the measure, scoring scale, reliability (Cronbach’s α , or the Spearman–Brown coefficient), mean ( M ) and standard deviation ( SD ) at the wave used in the primary analysis (T1 for motivation, T2 otherwise), and the role the measure plays in the analysis. a Usage frequency uses a different instrument at T0/T1 (5 platforms, daily scale) than at T2 (13 platforms, weekly scale); the two are not on a common metric.
Characteristic
n (%)
Gender
Male
54 (47.4%)
Female
60 (52.6%)
Age
18–25
3 (2.6%)
26–35
35 (30.7%)
Table 2. Participant characteristics from the one-year follow-up study ( N =114). Age and education were collected at T2; all other variables were collected at T0.
Figure 1. Path diagram of the main result, with social type of AI motivation as a representative case for the other motivations tested (escape, entertainment, and instrumental). Figure shows standardized coefficients from two single-mediator models. Indirect effects also report 95% bootstrapped confidence intervals ( N =1000). The association between motivation at T1 and problematic AI use at T2 runs mainly through emotional use purpose (blue) rather than usage frequency (red). The direct path c′ is no longer distinguishable from zero once emotional use purpose is in the model (upper triangle) but remains significant with usage frequency as the mediator (lower triangle). * p <.05; ** p <.01; *** p <.001; ns = not significant.
Variables
M
SD
Rel.
1
2
3
4
5
6
7
8
9
10
1
Problematic use (T2)
1.62
0.68
.87
2
Escape motivation (T1)
1.70
0.93
.94
.43
3
Social motivation (T1)
1.79
0.91
.95
.51
.59
4
Entertainment motivation (T1)
2.50
0.89
.89
.42
.52
.50
5
Instrumental motivation (T1)
3.44
0.65
.82
.31
.20
.31
.54
6
Overall motivation (T1)
2.36
0.65
.91
.55
.79
.81
.83
.61
Table 3. Descriptive statistics, reliability, and correlations of the study variables. Reliability is measured using Cronbach’s alpha. Emotional use purpose reliability is Spearman–Brown (2 items); the single-item usage frequency measure has no internal-consistency estimate. The lower panel gives each individual characteristic’s correlation with variables 1–10. Correlations are Pearson and uncorrected, with ∣r∣≥.19 is significant at p<.05 , ∣r∣≥.24 at p<.01 , and ∣r∣≥.31 at p<.001 .
Figure 2. Motivation at T1 and T2 as a predictor of problematic AI use at T2. Points are standardized coefficients from separate ordinary least squares models, with age, education, and baseline mental-health status as covariates. Vertical bars indicate 95% confidence intervals. Prospective estimates (T1) are shown in blue; the same-wave estimate (T2) in red. The prospective association is weaker than the concurrent one but is available a year in advance.
Figure 3. Indirect effect of each T1 motivation measure on problematic AI use at T2 through each T2 mediator, from single-mediator bootstrap models with age, education, and baseline mental-health status as covariates. Bars indicate 95% confidence intervals ( N =1000). Colors indicate emotional use purpose (dark blue), functional purpose (light blue), external purpose (gray), and usage frequency (red). Channels whose interval includes zero are marked ns.
Motivation (T1)
Mediator (T2)
a
b
c′
Indirect [95% CI]
% med.
Overall
Emotional purpose
.58
.55
.22*
.32 [.21, .44]
60%
Overall
Usage frequency
.33
.29
.44***
.10 [.04, .16]
18%
Overall
Functional purpose
.37
.19
.47***
.07 [.02, .13]
13%
Overall
External purpose
.05
.18
.53***
.01 [ − .03, .05] ns
—
Escape
Emotional purpose
.41
.59
.22**
.24 [.13, .36]
52%
Escape
Usage frequency
.22
.36
.38***
.08 [.02, .13]
17%
Table 4. Single-mediator models (95% confidence intervals; N =1000): each T1 motivation measured through each T2 mediator to problematic AI use at T2, with age, education, and baseline mental-health status as covariates in both stages. a = motivation → mediator; b = mediator → outcome; c′ = direct effect with the mediator included. % med. is the indirect effect as a share of the total, shown only where the indirect interval excludes zero; it is a derived ratio that is unstable at this sample size and is reported for ordering, not as an effect size. Mediator and outcome are measured in the same wave. †p <.10; * p <.05; ** p <.01; *** p <.001.
Appendix figures & tables17 assets
Supplementary material from the paper’s appendix.
Appendix
Instrument
Waves
Source
In analyses
Motivation for AI use
T0, T1, T2
( Huang et al., 2024 )
Yes
AI dependence
T0, T1, T2
( Huang et al., 2024 )
Yes
Problematic AI use a
T2
adapted from ( Yu et al., 2024 )
Yes
AI use purpose a
T2
adapted from ( Wolf and Maier, 2024 )
Yes
Usage frequency a
T0, T1, T2
this study
Yes
General attitudes toward AI (GAAIS)
T0, T1, T2
( Schepman and Rodway, 2020 )
Yes
Appendix
Table A1. Instruments administered across the three waves. a Introduced or adapted in this paper.
1. I constantly have thoughts related to AI conversational agents lingering in my mind.
2. I frequently find myself opening AI conversational agents even when I had no initial intention to use one.
3. I progressively spend more time with AI conversational agents.
4. I have lost interest in previously enjoyable activities due to using AI conversational agents.
5. My use of AI conversational agents causes me to procrastinate and delay completing necessary tasks.
6. I become fully absorbed when using AI conversational agents.
7. I devote excessive time to AI conversational agents despite problems.
Appendix
Table A2. Problematic AI use: items as administered (T2 only).
Item
Scale
To overcome certain difficulties.
Emotional
To enhance my learning and knowledge.
Functional
For the pleasure it gives me to know more about it.
(unassigned)
To manage my feelings.
Emotional
Please select neutral for this statement.
(attention check)
Because I am encouraged or required to.
External
Appendix
Table A3. AI use purpose items and scale assignment (T2 only). (R) = reverse-scored.
Model
χ2 ( df )
χ2/df
CFI
TLI
RMSEA
SRMR
Motivation 4-factor (T2, 12 items)
78.6 (48)
1.64
.973
.963
.075
.059
Use purpose 3-factor (T2, 13 items)
133.4 (62)
2.15
.899
.873
.101
.073
Use purpose 2-factor (Emotional+Functional, 8 items)
65.9 (19)
3.47
.915
.874
.148
.060
Appendix
Table A4. Confirmatory factor analyses of motivation.
M
Per-item M
SD
Median
80th pct.
90th pct.
Skewness
This study (10-item sum)
16.2
1.62
6.8
14
21
26
1.39
This study (9-item sum, source items only)
13.8
1.53
5.9
12
17
23
1.61
Yu et al. (11-item sum)
15.9
1.44
6.9
13
21
26
1.71
Appendix
Table A5. Problematic AI use on the summed metric compared with the source scale’s published distribution ( Yu et al., 2024 ) . Our sums span the 10 administered items and the 9 source-scale items (dropping the investigator-added item 6) on a 5-point scale (possible ranges 10–50 and 9–45); the source spans 11 items on a 4-point scale (range 11–44). Per-item M = sum divided by the number of items.
Motivation (T1)
No controls
Primary
+ baseline usage
Correlation-screened
Overall
0.373***
0.367***
0.359***
0.335***
Escape
0.290***
0.312***
0.298***
0.281***
Social
0.343***
0.328***
0.315***
0.282***
Entertainment
0.288***
0.287***
0.271***
0.242***
Instrumental
0.207***
0.201***
0.186***
0.200***
Appendix
Table A6. Control-specification sensitivity for the T1 motivation models ( n =114). *** p <.001.
Outcome (T2)
M treat (AU)
M control (BU)
t
p
d
q
Positive AI attitudes (T2)
3.71
3.55
1.03
.308
.19
.410
Negative AI attitudes (T2)
3.21
3.11
0.59
.556
.12
.635
AI dependence (T2)
7.06
6.58
1.05
.295
.20
.410
Problematic use (T2)
1.68
1.52
1.22
.225
.23
.410
Social motivation (T2)
1.70
1.50
1.20
.232
.22
.410
Escape motivation (T2)
1.48
1.44
0.32
.748
.06
.748
Appendix
Table A7. Experimental arm differences at T2 (Welch two-sample t -tests; AU arm n =69 vs. BU arm n =45). Positive d indicates that AU arm is higher. q from Benjamini–Hochberg correction within this 8-test family; 0/8 survive.
Motivation (T0)
β
Adjusted R2
Overall
0.346***
.341
Escape
0.323***
.316
Social
0.289***
.268
Entertainment
0.233***
.207
Instrumental
0.163**
.146
Appendix
Table A8. Motivation measured at T0 as a predictor of problematic use at T2 ( n =114). ** p <.01; *** p <.001.
Motivation (T1)
β OLS (reported)
β rank
β log
OR logistic
Overall
.367
.564
.206
2.89
Escape
.312
.463
.172
1.87
Social
.328
.479
.179
2.56
Instrumental
.201
.353
.121
2.40
Entertainment
.287
.468
.164
2.52
Appendix
Table A9. Distributional robustness of the primary T1 motivation models ( n =114), with age, education, and baseline mental-health status as covariates. Rank = OLS on rank-transformed outcome and predictor; log = log-transformed outcome; logistic = P (problematic use > 2.0), n =24 above the cut. All estimates significant (largest p =.018).
Subscale
% at floor
% at ceiling
SD
Escape
52
4
0.93
Social
45
4
0.91
Instrumental
2
38
0.65
Entertainment
12
5
0.89
Appendix
Table A10. Distributions of the T1 motivation subscales ( n =114).
Test
Comparison
Statistic
p
pairwise
Escape vs Social
-0.11
.914
pairwise
Escape vs Instrumental
0.89
.372
pairwise
Escape vs Entertainment
0.45
.650
pairwise
Social vs Instrumental
0.91
.362
pairwise
Social vs Entertainment
0.64
.521
pairwise
Instrumental vs Entertainment
-0.31
.753
Appendix
Table A11. Coefficient-equality tests among the four T1 motivation subscales in a joint model of problematic use at T2 ( n =114; Wald tests on the difference of standardized coefficients, HC3).
Characteristic
β
SE
95% CI
p
q
Age
.236
.062
[.115, .357]
< .001
.001
Education
.054
.068
[ − .080, .188]
.427
.702
Gender (female)
.014
.137
[ − .254, .282]
.920
.920
Mental-health condition (T0)
.210
.133
[ − .052, .471]
.116
.347
Loneliness (T0)
.010
.071
[ − .129, .148]
.891
.920
Impulsivity (T0)
.050
.069
[ − .085, .185]
.468
.702
Appendix
Table A12. Individual characteristics as joint predictors of problematic AI use at T2 ( n =114; one OLS model, standardized continuous predictors, HC3 standard errors; gender, mental-health condition, and race entered as 0/1). Adjusted R2 = .137. q from Benjamini–Hochberg correction within this 9-test family.
Characteristic
β (T1)
β (T2)
β (change)
Age
.066
.132
.066
Education
-.082
-.027
.055
Gender (female)
-.120
-.173
-.052
Mental-health condition (T0)
-.007
.045
.052
Loneliness (T0)
-.021
.039
.060
Impulsivity (T0)
.097
.054
-.043
Appendix
Table A13. Individual characteristics as joint predictors of overall motivation ( n =114; one OLS model per column). Adjusted R2 = .059 (T1), .089 (T2), .025 (T1 to T2 change).
Motivation (T1)
β solo
β with frequency
% removed
Overall
.367
.301
18%
Escape
.312
.260
17%
Social
.328
.257
22%
Instrumental
.201
.122
39%
Entertainment
.287
.231
19%
Appendix
Table A14. Attenuation of the T1 motivation coefficients when T2 usage frequency enters the model ( n =114). All motivation coefficients remain significant after correction.
Motivation (T1)
Quantity
Estimate [95% CI]
Overall
specific indirect: EMO
.28 [.17, .41]
Overall
specific indirect: FREQ
.05 [.01, .10]
Overall
specific indirect: FUNC
-.03 [-.08, .03]
ns
Overall
specific indirect: EXT
.01 [-.01, .04]
ns
Overall
contrast EMO - FREQ
.23 [.11, .37]
Overall
contrast EMO - FUNC
.31 [.18, .46]
Appendix
Table A15. Parallel four-mediator model ( n =114): all four T2 mediators entered simultaneously; specific indirect effects (each channel controlling the other three) and pairwise emotional-minus-rival contrasts, 95% bootstrap CIs ( N =1000). EMO = emotional use purpose; FREQ = usage frequency; FUNC = functional use purpose; EXT = external use purpose. * = CI excludes zero.
Measure
Interval
M before
M after
Δ
dz
q
%↑
AI dependence
T0 to T1
6.01
6.33
+0.32
.18
.084
32
AI dependence
T1 to T2
6.33
6.87
+0.54
.25
.011
39
AI dependence
T0 to T2
6.01
6.87
+0.86
.39
< .001
41
Overall motivation
T0 to T1
2.13
2.36
+0.23
.44
< .001
61
Overall motivation
T1 to T2
2.36
2.17
-0.19
-.43
< .001
29
Overall motivation
T0 to T2
2.13
2.17
+0.04
.09
.445
52
Appendix
Table A16. Paired within-person shifts in AI dependence, overall motivation, and AI attitudes ( n =114; paired t -tests, dz = mean change / SD of change, q = Benjamini–Hochberg within the shift family). Dependence is the 5–20 sum composite; motivation is on its 1–4 scale; attitudes on their 1–5 scale. %↑ = share of participants whose score increased.
Figure A1 . Standardized within-person change from T1 to T2 (Cohen’s dz with 95% confidence intervals, n =114). Self-assessed AI dependence and its excessive-usage and over-reliance facets rose, while stated motivation (all four subscales and the composite) and both positive and negative attitudes toward AI declined. Stars reflect q (Benjamini–Hochberg within the shift family): †q <.10; * q <.05; ** q <.01; *** q <.001.
Public discourse and emerging policy typically assume that AI emotional support is a deliberate act: a lonely user consciously seeking comfort from a dedicated companion chatbot. In this paper, we draw on emerging empirical evidence and argue that this picture is inaccurate on two accounts, both in how AI emotional support arises and how it shapes future behavior. First, AI emotional support commonly emerges incidentally within task-oriented interactions on general-purpose platforms, much as workplace friendships deepen through collaboration. Second, these incidental encounters are path-dependent: positive experiences of AI emotional support update people's beliefs about AI's emotional capabilities and redirect their choices for future emotional support, increasing preference for AI and decreasing preference for humans. We review recent evidence, including a large-scale longitudinal study conducted in collaboration with OpenAI, showing that daily five-minute conversations with an AI about personal issues over 28 days led to a 10.3% decrease in the preference for seeking support from humans and an 11.6% increase in the preference for AI. These findings suggest that current policy, focused on companion apps and isolated interactions, cannot adequately protect human connection. Instead, effective regulations should extend to general-purpose AI systems and address cumulative, trajectory-level changes in how people seek support. Recognizing how people stumble into AI emotional support and how those encounters redirect human connections over time is essential to safeguarding human well-being.
Yaoxi Shi, Cathy Mengying Fang, Pattie Maez +1
Imperial College Business School, Imperial College London; London, UK. · Harvard Business School AI Institute, Harvard University; Cambridge, USA. · MIT Media Lab, Massachusetts Institute of Technology; Cambridge, USA. +2
As conversational AI systems become increasingly integrated into daily life, their potential effects on user well-being require ongoing attention. While consumer-facing generalist models can provide benefits, including improved access to information, learning, productivity, self-reflection, and companionship, they also introduce risks, such as emotional entanglement, unhealthy dependence, and the amplification of psychological vulnerabilities. Drawing on prior research and empirical observations of AI chatbot behavior, we propose a set of aspirational directions for guiding the behavior of general-purpose AI systems in ways that may reduce potential psychological harms and support user well-being. We acknowledge the difficulty of systematically assessing the long-term impacts of AI chatbot use and frame these directions as hypotheses for studying how AI behavior may influence users across general interactions, role-playing scenarios, and contexts that could be characterized as providing psychological support. While some proposed directions are supported by existing research and expert insights, others identify open questions and areas requiring deeper study. We hope that this formulation and these hypotheses encourage further discussion, empirical investigation, and exploration of interactive design approaches aimed at better accommodating users' psychological needs and promoting their well-being.
Jina Suh, Mihaela Vorvoreanu, Forough Poursabzi-Sangdeh +5
Microsoft, USA · University of Washington · City University of New York, USA +1
As conversational AI systems are designed for repeated use, a central question is how a series of interactions becomes a relationship. We present a longitudinal multimodal study of a memory-augmented conversational agent (24 participants x 10 sessions), in which participants rated five relational constructs -- familiarity, self-disclosure, perceived memory, conversational quality, and enjoyment -- after each session. Two complementary dynamics emerge. First, conversational quality strongly shapes how enjoyable a session feels in the moment but does not carry forward across sessions, whereas perceived memory is relationally conditioned -- predicted by prior relational state rather than reflecting system capability alone -- and it shapes later enjoyment indirectly, via subsequent self-disclosure. Second, relationships are punctuated by discrete turning points -- crashes and surges -- that are partially traceable in multimodal behavior and open different intervention windows: surges are more behaviorally detectable in the moment, enjoyment surges persist more reliably than enjoyment crashes recover, and some crashes are better forecast from person-specific behavioral drift than detected after they have already occurred. Together, the findings suggest that longitudinal human-AI relationships are built through both slow accumulation and abrupt turning points.
Ryuichi Sumida, Mao Saeki, Masaki Eguchi +4
Graduate School of Informatics, Kyoto University Kyoto, Japan · Equmenopolis, Inc. Tokyo, Japan · Waseda University Tokyo, Japan +1