PaperlayerView exact revision on arXiv

Source-linked AI summary

Trust in AI and Its Role in the Acceptance of AI Technologies

Hyesun Choung, Prabu David, Arun Ross

arXiv:2203.12687v1cs.AIcs.HC

TL;DR

Understanding what facilitates trust in AI is a practical need, given trust's role in acceptance. This paper applies trust within the TAM and treats it as multidimensional, finding that perceived ease of use contributes to trust and trust predicts positive attitudes and usefulness associated with greater usage.

Abstract

from arXiv · show

As AI-enhanced technologies become common in a variety of domains, there is an increasing need to define and examine the trust that users have in such technologies. Given the progress in the development of AI, a correspondingly sophisticated understanding of trust in the technology is required. This paper addresses this need by explaining the role of trust on the intention to use AI technologies. Study 1 examined the role of trust in the use of AI voice assistants based on survey responses from college students. A path analysis confirmed that trust had a significant effect on the intention to use AI, which operated through perceived usefulness and participants' attitude toward voice assistants. In study 2, using data from a representative sample of the U.S. population, different dimensions of trust were examined using exploratory factor analysis, which yielded two dimensions: human-like trust and functionality trust. The results of the path analyses from Study 1 were replicated in Study 2, confirming the indirect effect of trust and the effects of perceived usefulness, ease of use, and attitude on intention to use. Further, both dimensions of trust shared a similar pattern of effects within the model, with functionality-related trust exhibiting a greater total impact on usage intention than human-like trust. Overall, the role of trust in the acceptance of AI technologies was significant across both studies. This research contributes to the advancement and application of the TAM in AI-related applications and offers a multidimensional measure of trust that can be utilized in the future study of trustworthy AI.

Trustworthy AI

AI autonomy and opaque processes can create risks and uncertainty for users, making trust important. This section motivates studying trust in consumer AI products and applying it within the TAM framework.

Trustworthy AI

  • AI’s autonomous functioning and black-box nature create user risks and uncertainty, highlighting the importance of trust.The paper contrasts AI with traditional technologies, where users have complete control over functioning.Source: Trustworthy AI · p. 6
  • Trustworthy-AI guidelines recommend building trustworthiness into systems by design, while Shin (2020b) extended FAT to FATE by adding explainability as a determinant of trustworthiness and perceived usefulness.The cited work concerns algorithmic recommendations.Source: Trustworthy AI · pp. 6–7
  • Prior studies found agent embodiment can increase trust through perceived social presence (K. Kim et al., 2018; Shin, 2021d; Waytz et al., 2014), although others reported weak or no anthropomorphism effects on trust.Source: Trustworthy AI · p. 7
  • The study focuses on consumer AI products, where frequent interactions with assistants and smart-home systems allow initial trust to develop, and applies trust within the TAM framework.It tests whether trust can be another key determinant of acceptance of AI technology.Source: Trustworthy AI · pp. 7–8

Trust and Use of AI Technologies: An Extended TAM · [Figure 1 Near Here] · Multidimensional Approach to Trust in AI

This section examines how trust in AI relates to acceptance within an extended TAM, including effects mediated through perceived usefulness and attitude and two proposed dimensions of trust: human-like and functionality trust. It focuses on whether trust predicts use of AI voice assistants and whether the trust dimensions differ in their influence on acceptance.

Trust and Use of AI Technologies: An Extended TAM

  • Trust is proposed to follow perceived ease of use, influence usefulness and attitude, and affect behavioral intention through usefulness, attitude, or both.The model also retains the TAM relationships among ease of use, usefulness, attitude, and intention.Source: Trust and Use of AI Technologies: An… · p. 9

Multidimensional Approach to Trust in AI

  • Because AI can autonomously replace human tasks and decisions, the authors argue that trust in AI draws on both human-like and technology-based trust.They describe trust in technology as dynamic and dependent on context and the characteristics of the trusted agent.Source: Multidimensional Approach to Trust in AI · pp. 11–12
  • The study conceptualizes trust in AI along two dimensions: human-like trust in algorithms’ social, cultural, and ethical values, and functionality trust in technological competence and expertise.The authors note that few studies address social and psychological trust in AI and that varied assessment measures can hinder collaboration and limit research implications.Source: Multidimensional Approach to Trust in AI · pp. 10, 12
  • The second study is framed to compare whether human-like and functionality trust differ in their influence on AI acceptance within the TAM.The research question asks whether the two dimensions have different influence within the model.Source: Multidimensional Approach to Trust in AI · pp. 12–13
  • Study 1 aims to test H1–H5 on the role of trust in AI adoption using voice assistants such as Siri, Alexa, and Google Assistant.The authors select conversational agents as the initial application for examining trust in AI use.Source: Multidimensional Approach to Trust in AI · p. 13
  • Voice assistants are designed to be more human-like.Source: Multidimensional Approach to Trust in AI · p. 13
  • Research explores similarities between how people treat other people and technologies.Source: Multidimensional Approach to Trust in AI · p. 13
  • Because voice assistants are anthropomorphic, the authors considered trust potentially important and focused Study 1 on whether it predicts their use.Source: Multidimensional Approach to Trust in AI · p. 13

Method … Analytic Approach

Study 1 used an online survey of undergraduates to measure trust and technology-acceptance constructs, then tested hypothesized relationships with path analysis. Before structural modeling, the seven constructs were moderately correlated.

Participants and Procedure

  • An online survey collected responses from 312 undergraduates at a large Midwest university between October and December 2020.Ninety-six percent reported prior use of at least one AI voice assistant.Source: Participants and Procedure · pp. 13–14

Measures

  • Ease of use and usefulness each used five items adapted from the original TAM scale, while trust used four newly created items.Behavioral intention to use or continue using an assistant was measured with three items from the TAM 2 scale.Source: Measures · p. 14

Analytic Approach

  • Path analysis tested the hypothesized relationships among variables in the model constructed in Figure 1.The analysis was conducted in R using lavaan.Source: Analytic Approach · p. 14
  • Model fit was assessed with chi-square, CFI, TLI, RMSEA, and SRMR; good fit required nonsignificant chi-square, CFI and TLI at least .90, and SRMR and RMSEA at most .10.Maximum likelihood estimation was used.Source: Analytic Approach · p. 14
  • Indirect-effect significance was tested using bootstrapped 95% confidence intervals based on 1,000 resamples.Source: Analytic Approach · p. 14
  • The seven constructs were moderately correlated before the structural model was tested.The authors report this in Table 2.Source: Analytic Approach · p. 15

[Table 2 Near Here] … Discussion

Study 1 supported the TAM for AI voice-assistant use: ease of use, usefulness, and attitude predicted intention, while trust operated through mediated paths. The discussion reports that the findings were replicated with a representative U.S. sample and extended to consumer smart technologies.

Results

  • The path model fit the data well: χ2 (1) = .32, p = .58; RMSEA = .00 [.00, .12]; SRMR = .01; CFI = 1.00; TLI = 1.01.Source: Results · p. 15

[Figure 2 About Here]

  • Ease of use, usefulness, and attitude positively predicted usage intention, while trust positively predicted usefulness and the proposed predictors showed direct and indirect effects.The tested model included significant mediated paths involving trust.Source: [Figure 2 About Here] · p. 15
  • Attitude predicted intention (β = .62, p < .001); ease of use predicted usefulness (β = .36, p < .001), trust (β = .39, p < .001), attitude (β = .26, p < .001), and intention (β = .19, p < .05).Source: [Figure 2 About Here] · p. 15
  • Usefulness predicted attitude (β = .42, p < .001) and willingness to use the assistant (β = .20, p < .01).Source: [Figure 2 About Here] · pp. 15–16

[Table 4 Near Here]

  • All depicted mediated paths were significant: trust influenced intention through usefulness, attitude, or usefulness followed by attitude.These pathways supported H4, including the hypothesized H5 paths.Source: [Table 4 Near Here] · p. 16
  • Ease of use had the largest total effect on intention (β = .63; p < .001), followed by usefulness (β = .46; p < .001) and trust (β = .32; p < .001).Usefulness had a greater direct effect than ease of use, but ease of use had the greater total effect.Source: [Table 4 Near Here] · p. 16, Discussion · pp. 16–17

Discussion

  • The validated model explained 52% of variance (R2 = .52), and ease of use contributed to trust, which predicted usefulness and positive attitudes but not intention directly.Source: Discussion · pp. 16–17
  • Study 2 replicated Study 1 with a representative national sample, measured trust multidimensionally, and extended the scope to consumer AI grouped as smart technologies.Source: Discussion · pp. 17–18

Method · Participants and Procedure · Measures

Study 2 used a quota-sampled U.S. survey and retained the Study 1 TAM measures while developing a two-dimension measure of trust in AI.

Participants and Procedure

Measures

  • Study 2 reused Study 1 items for perceived ease of use, perceived usefulness, attitude, and behavioral intention.Source: Measures · p. 18
  • Trust items drew on benevolence/helpfulness, integrity/reliability, and competence/functionality pillars from trust research.Source: Measures · p. 18
  • Exploratory factor analysis grouped benevolence and integrity items as human-like trust and treated competence as functionality trust.The resulting dimensions comprised six and five items, respectively.Source: Measures · p. 18

[Table 5 near here] … Results

Both trust dimensions indirectly predicted intention to use AI smart technologies through TAM constructs, and functionality-related trust had the greater total impact on usage intention. The models also supported the proposed roles of perceived usefulness, ease of use, and attitude.

Analytic Approach

  • The analysis used the Study 1 approach to test separate path models for human-like and functionality-related trust, with construct intercorrelations presented in Table 6.Source: Analytic Approach · p. 19

Results

  • Both path models fit adequately, with CFI = 1.00, TLI = .98, SRMR = .01, and RMSEA = .08 [.02, .15]; the chi-square tests were significant.Source: Results · p. 19
  • Attitude predicted usage intention (β = .65, p < .001), and perceived usefulness predicted attitude (β = .54 and β = .53, p < .001) and usage intention (β = .22, p < .001).Source: Results · pp. 19–20
  • Perceived ease of use positively predicted perceived usefulness, both trust dimensions, attitude, and usage intention in the two models.Source: Results · pp. 19–20
  • Both human-like and functionality-related trust had statistically significant indirect effects on intention to use AI smart technologies.Source: Results · p. 20
  • Trust predicted perceived usefulness and positive attitudes, which in turn predicted usage intention, supporting the proposed indirect pathways.Source: Results · p. 20
  • Perceived ease of use had a total effect of β = .71 in both models, while perceived usefulness had total effects of β = .57 and β = .56 (p < .001).Source: Results · p. 20
  • Functionality-related trust had a greater total effect on usage intention than human-like trust (β = .42 vs. β = .36; p < .001 for both).Source: Results · p. 20

Discussion · General Discussion

Across both studies, trust contributed to AI acceptance within the TAM, alongside perceived ease of use and usefulness. The findings distinguish human-like and functionality-related trust, with functionality trust having the greater total impact on usage intention.

Discussion

  • Both human-like and functionality trust predicted usefulness, positive attitudes, and usage intention, while functionality trust had the greater total impact.Source: Discussion · p. 21
  • Both AI’s functional and social properties matter for trustworthy design, including social graces, privacy protection, fairness, and avoiding bias.Source: Discussion · p. 21

General Discussion

  • Perceived ease of use and usefulness were essential acceptance determinants, with ease of use consistently having greater impact and helping users trust and value AI.The authors argue that straightforward operation is critical for uptake.Source: General Discussion · p. 22
  • Across both studies, trust significantly shaped attitudes and acceptance of AI technologies within the TAM framework.Source: General Discussion · p. 22
  • Trust develops through users’ direct experience across a technology’s lifecycle, so trustworthy AI should address use experience as well as design.Users form judgments about usefulness and ease of use from their experience; the TAM offers a framework for considering that experience.Source: General Discussion · pp. 22–23
  • Future AI acceptance models should include trust and investigate what facilitates it and how human-like and technology-related trust differ in their antecedents and effects.Source: General Discussion · p. 23
  • The study operationalized human-like trust through social cues, intentionality, and moral agency, and found both it and functionality trust predicted technology acceptance.The authors report a clear perceptual distinction between the dimensions and offer their measures for future AI research.Source: General Discussion · p. 23
  • The authors identify multi-level trust as an unexamined approach that may add insight, citing Hoff and Bashir (2015) on dispositional, situational, and learned trust.They note trust will be increasingly critical in high-stakes AI applications such as autonomous vehicles and medical decision-making.Source: General Discussion · p. 23

Conclusion · Theoretical Contributions

The paper argues that AI’s progress and black-box mechanisms make a sophisticated understanding of trust increasingly important. It extends the TAM by treating trust as central to AI acceptance research and by examining distinct dimensions of trust.

Conclusion

  • As AI advances and its influence on everyday life is expected to grow, the paper argues that its black-box mechanisms make a more sophisticated understanding of AI trust increasingly important.The study examined different dimensions of trust within the established TAM framework.Source: Conclusion · pp. 23–24

Theoretical Contributions

  • The study’s core theoretical contribution is extending the classical TAM with AI-specific trust and arguing that trust belongs in predictive models of AI acceptance.The authors also call for research into what facilitates trust and supports the design of trustworthy AI.Source: Theoretical Contributions · p. 24
  • The multidimensional approach offers a lens for representing variability in people’s AI trust perceptions and can guide future research.The authors suggest that specific AI ethical principles may align with specific trust dimensions.Source: Theoretical Contributions · p. 24
  • The authors urge future research to examine AI trust by dimension, since trust-building factors may relate to particular dimensions.They propose that anthropomorphism may increase human-like trust without affecting functionality trust, and that people may differ in trust across the two dimensions.Source: Theoretical Contributions · p. 24

Practical Contributions · Limitations and Future Research Directions · Basis of trust in technology

The paper argues that AI design should combine usability and usefulness with trustworthiness, while future research should test trust across broader applications, trust targets, causal relationships, and populations. It also frames trust in automation through functionality, performance, reliability, process, helpfulness, and purpose.

Practical Contributions

  • AI should be easy to use, useful, and trusted; paired with trust-supporting factors such as FATE or anthropomorphism, these characteristics are expected to encourage acceptance and use.Source: Practical Contributions · p. 25
  • Designers should treat value-embedded trustworthiness as critical, using the multidimensional framework to guide ethical, human-centered AI design, governance, and interventions that encourage appropriate trust.Source: Practical Contributions · p. 25

Limitations and Future Research Directions

  • Because the study focused on lower-risk consumer products, future research should examine high-stakes AI and other applications with human-like features, including robots and embedded algorithms.Such work could examine trust and other critical factors in adoption and use.Source: Limitations and Future Research… · p. 25
  • Trust varies across technology stages, and future work could distinguish initial from ongoing trust and develop measures for trust in technology, companies, and governing bodies.The cited examples are J. B. Kim (2012) and Söllner et al. (2016).Source: Limitations and Future Research… · pp. 25–26
  • Correlational path models cannot establish that trust causes AI acceptance; future studies should test causal relationships among TAM constructs, including the possibility of a reversed relationship.The authors positioned trust before perceived usefulness based on prior TAM findings (K. Wu et al., 2011), while another study found the opposite direction.Source: Limitations and Future Research… · p. 26
  • Online-panel respondents may not represent the general population, because computer and internet access could skew the sample toward greater technology acceptance, access, and experience.Future studies should include more representative samples with varied technology experience.Source: Limitations and Future Research… · p. 26

Basis of trust in technology

Standardized Regression Path Analysis (Study 1) … Results of Path Analysis with Functionality Trust in AI (Study 2)

Study 1’s path analysis shows a strong attitude-to-intention path, while Study 2’s technology-trust analysis reports indirect paths to intention through usefulness and attitude. The supplied passages identify additional analyses and visuals but provide no estimates for them.

Standardized Regression Path Analysis (Study 1)

  • Table 4 reports that attitude predicted behavior intention most strongly (β = .62); trust predicted usefulness (β = .39) and attitude (β = .23), with all three paths p < .001.Perceived ease of use predicted usefulness (β = .36), trust (β = .39), attitude (β = .26), and intention (β = .19); usefulness predicted attitude (β = .42) and intention (β = .20).Source: Standardized Regression Path Analysis… · p. 35

Analysis of Indirect Effects with Technology Trust in AI (Study 2)

  • The reported indirect paths to BI include PU → ATT → BI (.34; 95% CI .28–.43) and six PE-mediated paths, whose effects range from .04 to .15.The PE-mediated effects are .11 (95% CI .06–.17), .10 (.05–.15), .15 (.11–.21), .10 (.06–.15), .04 (.02–.07), and .07 (.05–.10).Source: Analysis of Indirect Effects with… · p. 38
  • Technology trust had indirect effects on intention through usefulness (.08), attitude (.20), and usefulness plus attitude (.13).The reported effects are TTR → PU → BI, TTR → ATT → BI, and TTR → PU → ATT → BI, respectively.Source: Analysis of Indirect Effects with… · p. 39

Hypothesized Path Model

  • The hypothesized path model is presented as Figure 2; no further model details are supplied in the provided passage.Source: Figure 2 · p. 40

Results of Path Analysis with Human-like Trust in AI (Study 2)

  • The human-like trust path-analysis results are identified as Figure 4, but the provided passage gives no figure results.Source: Figure 4 · p. 41

Appendix

The appendix documents the study questionnaires, including trust measures and shared acceptance-model items. It notes that common item wordings were adapted to refer to virtual assistants in Study 1 and smart technologies in Study 2.

Acceptance measures

  • Across both studies, the questionnaires measured ease of use, usefulness, attitude, and behavioral intention, with the technology description varying by study.The common item wording referred to AI virtual assistants in Study 1 and AI smart technologies in Study 2.Source: Appendix · pp. 42–44

Behavioral intention

  • Behavioral-intention items distinguished non-users’ intention to use the technologies in the future from users’ intention to continue using them.Both item sets included intention, prediction, and future-use statements.Source: Appendix · p. 43

Related papers

Loading 2203.12687v1…