Review of the Factors Affecting Acceptance of AI-Infused Systems

Accurate algorithmic predictions alone are not enough for effective human–AI collaboration, and many published AI studies focus on showing that the technology works. Theories of user acceptance usually explain the adoption of AI devices through behavioral factors (e.g., attitude, perceived usefulness, perceived ease of use, trust), while the effects of technical and human factors are often overlooked, even though user behavior can vary with a system's technical characteristics and with differences between users. Prior reviews had focused mainly on autonomous vehicles or specific populations. According to the authors, the general acceptance of AI-based technologies across application fields, and the combined effects of technological, human and behavioral factors, had not been extensively reviewed.

A systematic literature review following the PRISMA model (Preferred Reporting Items for Systematic Reviews and Meta-Analyses).

• Sources: IEEE Xplore, Springer Link and Google Scholar, searched with keywords such as "artificial intelligence", "technology adoption", "acceptance" and "intention of use"; last search on December 1, 2020; English-language papers on AI technologies addressing adoption factors; studies published before 2000 and duplicates excluded.
• Selection: 464 records (161 IEEE Xplore, 145 Springer Link, 158 Google Scholar); 78 duplicates removed; 386 screened by abstract, of which 249 were excluded; 137 full texts assessed; 85 articles included: 44 on autonomous driving, 25 on robotics and 16 on healthcare.
• Analysis: frequency analysis and qualitative synthesis of three groups of factors: behavioral (attitude, perceived usefulness, perceived ease of use, subjective norms, trust, perceived benefit, perceived risk), technological (transparency, reliability, complexity, compatibility) and human (age, gender, education, cultural background, technical expertise), and of the relationships among them. The summary table of these relationships contains no statistical outcomes.

• Review finding: Behavioral factors have been studied most and were noted to have the strongest influence on adoption; attitude and perceived usefulness were identified in many studies as the most influential factors on adoption intention. Trust and perceived risk also play significant roles, and privacy risk is a particular issue because AI systems make extensive use of personal information.
• Review finding: Technological factors influence behavioral factors. User distrust of AI systems can be addressed by increasing transparency, reliability and compatibility and by reducing complexity. The effects of transparency and compatibility did not vary by application domain, while complexity was not observed as an important factor in autonomous driving.
• Review finding: Age, gender and cultural background were important across application areas. Education mattered in autonomous driving and healthcare but not robotics, and no gender effects were found in healthcare. Older adults, women, people with low education, people inexperienced with technology and members of relatively collectivistic cultures were found to have more difficulty adopting AI technology.
• Review finding: No significant differences were found in the impact of behavioral factors across the three application areas. However, no effects of perceived benefit or perceived risk were observed for robotics, which the authors note may reflect the relatively few studies on risk in that field.
• Conceptual contribution / recommendations: The review suggests extending traditional TAMs to include perceived risk and to integrate technological and human factors. It lists design considerations, such as privacy protection and ease of use for older adults, more functionality for younger users, increasing trust among women, and attention to cultural background.

The review gives a comprehensive overview of factors affecting the acceptance of AI-infused systems. It concludes that intention to use AI technologies is influenced by technological and human factors as well as by factors from behavior theories, and that including them in integrated TAMs might predict acceptance more precisely. The authors recommend considering compatibility and transparency in every application area, reliability in healthcare and autonomous driving, and complexity in robotics. The stated limitations are the limited number of studies, which came from selective keywords, and the resulting coverage of relatively few application areas. The authors note that the findings and implications need to be validated through experiments or surveys, and that future research should develop an enhanced TAM integrating the reviewed factors.