Key Insights from a Multidisciplinary Review of Trust Literature
Understanding trust between humans and AI systems is integral to promoting the development and deployment of socially beneficial and responsible AI. Successfully doing so warrants multidisciplinary collaboration.
In order to better understand trust between humans and artificially intelligent systems, the Partnership on AI (PAI), supported by members of its Collaborations Between People and AI Systems (CPAIS) Expert Group, conducted an initial survey and analysis of the multidisciplinary literature on AI, humans, and trust. This project includes a thematically-tagged Bibliography with 78 aggregated research articles, as well as an overview document presenting seven key insights.
These key insights, themes, and aggregated texts can serve as fruitful entry points for those investigating the nuances in the literature on humans, trust, and AI, and can help align understandings related to trust between people and AI systems. This work can also inform future research, which should investigate gaps in the research and our bibliography to improve our understanding of how human-AI trust facilitates, or sometimes hinders, the responsible implementation and application of AI technologies.
Several high-level insights emerged when reflecting on the bibliography of submitted articles:
- There is a presupposition that trust in AI is a good thing, with limited consideration of distrust’s value.
The original project proposal emphasized a need to understand the literature on humans, AI, and trust in order to eventually determine appropriate levels of trust and distrust between AI and humans in different contexts. However, the articles included in the bibliography are largely framed with the need and motivation towards trust – not distrust – between AI systems and humans. While certain instances may warrant facilitated trust between humans and AI, others may actually enable more socially beneficial outcomes if they prompt distrust or caution. Future literature should explore distrust as related, but not necessarily directly opposite, to the concept of trust. For example, an AI system that helps doctors detect cancer cells is only useful if the human doctor and patient trust that information. In contrast, individuals should remain skeptical of AI systems designed to induce trust for malevolent purposes, such as AI-generated malware that may use data to more realistically mimic the conversational style of a target’s closest friends.
- Many of the articles were published before the Internet’s ubiquity/the social implications of AI became a central research focus.
It is important to contextualize recent literature on intelligent systems and humans with literature focused on social and cognitive mechanisms undergirding human to human, or human to organizational, trust. Future work can put many of the foundational, conceptual articles that were written before the 21st century in conversation with those specifically focused on the context of AI systems, and their different use cases. It can also compare foundational, early articles’ exploration of trust with how trust is seen specifically in relation to humans interacting with AI.
- Trust between humans and AI is not monolithic: Context is vital.
Trust is not all or nothing. There often exist varying degrees of trust, and the level of trust sufficient to deploy AI in different contexts is therefore an important question for future exploration. There might also be several layers of trust to secure before someone might trust and perhaps ultimately use an AI tool. For example, one might trust the data upon which an intelligent system was trained, but not the organization using that data, or one might trust a recommender system or algorithm’s ability to provide useful information, but not the specific platform upon which it is delivered. The implications of this multifaceted trust between human and AI systems, as well as its implications on adoption and use, should be explored in future research.
- Promoting trust is often presented simplistically in the literature.
The majority of the literature appears to assert that not only are AI systems inherently deserving of trust, but also that people need guidance in order to trust the systems. The basic formula is that explanation will demonstrate trustworthiness, and once understood to be deserving of trust, people will use AI. Both of these conceptual leaps are contestable. While explaining the internal logic of AI systems does, in some instances, improve confidence for expert users, in general, providing simplified models of the internal workings of AI has not been shown to be helpful or to increase trust.
- Articles make different assumptions about why trust matters.
Within our corpus, we found a range of implicit assumptions about why fostering and maintaining trust is important and valuable. The dominant stance is that trust is necessary to ensure that people will use AI. The link between trust and adoption is tenuous at best, as people often use technologies without trusting them. What is largely consistent across the corpus – with the exception of some papers concerned about the dangers of overtrust in AI – is the goal of fostering more trust in AI, or stated differently, that more trust is inherently better than less trust. This premise needs challenging. A more reasonable goal would be that people are able to make individual assessments about which AI they ought to trust and which they ought not trust, in the service of their goals for what specifically and in which circumstances. This connects to insight 1: There is a presupposition that trust in AI is a good thing. It is important to think about context, person-level motivations and preferences, as well as instances in which trust might not be a precondition for use or adoption.
- AI definitions differ between publications.
The lack of consistent definitions for AI within our corpus makes it difficult to compare findings. Most articles do not present a formal definition of AI, as they are concerned with a particular intelligent system applied in a specific domain. The systems in question differ in significant ways, in terms of the types of users who may need to trust the system, the types of outputs that a person may need to trust, and the contexts in which the AI is operating (e.g., high- vs. low-stakes environments). It is likely that these entail different strategies as they relate to trust. There is a need to develop a framework for understanding how these different contributions relate to each other, potentially looking not at trust in AI, but at trust in different facets and applications of AI. For a more detailed analysis of what questions to ask to differentiate particular types of human-AI collaboration, see the PAI CPAI Human-AI Collaboration Framework.
- Institutional trust is underrepresented.
Institutional trust might be especially relevant in the context of AI, where there is often a competence or knowledge gap between everyday users and those developing the AI technologies. Everyday users, lacking high levels of technical capital and knowledge, may find it difficult to make informed judgments of particular AI technologies; in the absence of this knowledge, they may rely on generalized feelings of institutional trust.