Exploring a Behavioral Model of “Positive Friction” in Human-AI Interaction

Self-control is a well-known behavioral challenge across a wide variety of settings, including healthcare, purchasing, and social interactions, in which individuals struggle to overcome “hot state” temptations if even their more reticent “cold state” selves fully know the downsides of impulsive behaviors[34]. As a result, ‘save me from myself’ interventions that intentionally insert logistical, physical, or temporal friction can curtail impulsive tendencies and interrupt undesirable behaviors. These cases of self-control are characterized by clear goals and concrete motivation to change, which can be strategically leveraged by behavioral designers to help individuals avoid temptation. Tactics to help people resist doing things they know they should not may include limiting access to content or engagement, such as reducing product features; imposing timeouts or waiting periods to dampen impulsive actions; adding layers of social accountability or oversight; or introducing self-imposed de-escalation techniques to modulate behavior.

A second form of positive friction is less concerned with promoting self-control over impulsive behavior than overcoming personal ‘cruise control’ that numbs, rather than heightens, impulses in the form of subconscious or autopilot behaviors. While autopilot behaviors can helpfully reduce cognitive effort, they can lead to tuning out important environmental signals or overlooking opportunities to deviate from the norm even in instances where this may be beneficial. In this case, positive friction may be less oriented toward eliciting a specific preferred outcome, and more toward introducing a greater openness and attention to a full range of options. Behavioral tactics to broaden individuals’ active attention and receptivity to inputs include examples such as modal dialog boxes that prompt reflection before quitting applications, strategies for ‘inoculating’ individuals against disinformation by pre-empting their exposure to false news[50] or disfluency techniques that employ intentionally malformed fonts to slow the pace of information processing[3].

In addition to dampening tendencies toward impulsive or autopilot behaviors, positive friction can also be used to stimulate or encourage new behaviors by providing prompts that energize individuals to take action. This can be especially helpful in cases where adopting good behaviors is challenging due to effort aversion or where an abstract promise of logical gains (e.g., “good health,” “sustainability”) does not supply sufficient motivation. In this case, strategies that employ social norms—such as commitment contracts—can serve as forcing mechanisms to stimulate behaviors by increasing external sources of accountability[17]. Behavioral activation can also occur through increasing the concreteness of both actions and potential benefits; for example, where novel or highly experiential activities can increase engagement by motivating action, focusing on specific stories or individuals, known as the identifiable victim (or actor) effect, can help us overcome the malaise of generic data[20, 31].

Finally, broad societal tendencies to valorize efficiency often encourage solutions that generate maximal value for minimal effort. However, this can lead to over-indexing on speed and precision over other valuable outcomes and overlook possibilities to create or recognize new forms of value[53]. Positive friction that disrupts efficiency in productive ways can play an important role in reframing the value of experiences and heightening opportunities to capture these benefits. Behavioral theorists have suggested that amplifying alternative and qualitative aspects of an experience, rather than relying only on quantifiable indicators, can help surface new forms of value and overcome default tendencies toward efficiency. These can include instances of ‘stealth health’ seen in Pokemon Go, in which the motivation to exercise is subsumed into gameplay[29], or in activities that intentionally cater to alternative values, such as slow lanes in grocery stores that convert transactional activities into social opportunities[42].

Many companies now rely on AI as a key component of solutions to craft seamless, personalized experiences and reduce barriers that impede customer journey efficiency[14]. For instance, Facebook’s recent launch of AI-powered ”smart glasses” provides an example of ”ultra-low-friction input” that enables users to remain effortlessly connected[27]. Retail giants like Amazon, Aldi, and Hudson have also embraced frictionless shopping models, allowing customers to bypass traditional checkout lines[35]; similarly, in the transportation sector, Hitachi and Genfare have experimented with frictionless mobility with AI-based traffic management to optimize traffic flow, reduce congestion, and expedite travel times[64, 22].

While the trend of ”frictionless” AI and machine learning solutions brings certain advantages, it also introduces significant risks, related to privacy concerns, potential amplification of algorithmic biases, intellectual property issues, and ethical challenges. Tendencies such as algorithm appreciation bias, which reveals a preference for algorithmic advice over human input in certain situations, regardless of the transparency of the algorithm[35, 48] can lead to overreliance on AI, and underscore the potential pitfalls of excessive appreciation of AI advice in the absence of critical safeguards. These tendencies may be particularly important to consider in light of findings from fields concerned with human cognition that integrate neural and symbolic methods to distinguish between quick, instinctive ’System 1’ human processing of information and the slower, more thoughtful ’System 2’[21]. While current trends in AI development show a preference for the rapid responses characteristic of ’System 1’, favoring technologies that offer immediate, effortless solutions, recent proposals have reintroduced interest in specific types of friction to encourage reflection, conflict, and care in AI systems, essentially integrating slower and more reflective ’System 2’-oriented thinking[49].

Research into algorithm appreciation, cognitive ease, and strain has informed investigations into the intentional use and design of positive friction across various AI-informed domains and challenges. These include examining computational friction on XAI’s interface to calibrate users’ trust in the AI system[43]; the use of friction as a cognitive tool to foster the civility of online discourse against false algorithmic information[25]; how friction can inform ‘micro-moments’ of human-computer interaction in impulsive and reflective behavioral perspectives[8]; and the value of positive friction by Contestational Design for civic engagement[24]. Expanding on these experiments and findings from the perspective of strategic behavioral design can help inform and enrich the design of future AI-empowered products and service systems.

While designers and developers can build on a long tradition of removing and employing friction in products and services, the widespread and embedded nature of AI in consumer offerings has increased both the need and the stakes to consider the role—and intentional incorporation—of friction in AI solutions. Just as instances of positive friction designed into offerings can target end-user or consumer behaviors to encourage self-regulatory behaviors and cultivate beneficial habits[47], AI practitioners can also benefit from behavioral speed bumps during the design and development of AI-enabled solutions by productively slowing down or interrupting design processes or inserting opportunities to identify and rectify biases that algorithms may perpetuate (Table1). However, the specific application of positive friction for both users and developers of AI may differ, depending on the nature of the situation or the intended goal.

In addition, different stakeholders may have diverse interactions, values, and contexts of use that require different strategies and considerations for when, why, and for whom positive friction is used. This parallels findings in the recent AI-related socio-technical research highlighting the significance of clearly defining and considering not only multiple stakeholder groups, including those who contribute to, develop, deploy, and consume or are otherwise affected by AI, but by development stages—such as data collection, training, model evaluation and analysis, and system deployment—when developing AI regulatory oversight[41].

Considering positive friction at multiple levels can yield improved products and experiences, as illustrated through two examples of ‘X’ (formerly known as Twitter). The “Read before you post” feature[58] applies positive friction as a social media safeguard to deter impulsive sharing of misinformation or harmful content. While this feature effectively shields some participants from harmful or deceptive information, its impact is limited to resharing and replying by users. As a result, the use of friction is reactive rather than preventive given that it does not extend to the creators of harmful content or the biased algorithms that promote such content.

In contrast, ‘X’s’ Algorithmic Bias Bounty Challenge[7] represents an example of employing positive friction more proactively by managing misinformation closer to its origination by targeting AI developers[7]. In this case, the Challenge founders shared ‘X’s’ saliency algorithm (also known as Twiitter’s image cropping algorithm), making their code available for participants in order to reproduce and examine bias issues in the algorithms at DEFCON, a prominent hacker conference[7]. The results of this challenge showed that participating hackers succeeded in revealing several hidden biases inside ‘X’s’ AI systems for future debugs, including but not limited to encoded stereotypical beauty standards, algorithmic preferences for certain race features, and linguistic biases[70].

Unlike the purely user-focused “Read before you post” feature, the Bounty Challenge introduced friction within the platform itself; this not only prompted AI developers to reassess and refine machine functions, but also proactively addressed AI algorithmic issues in advance rather than designing positive friction features that users would need to use reactively. While both examples share the objective of fostering a less biased and more inclusive online environment, their differences illustrate how effective use of positive friction in an AI context not only can affect different stakeholders in varying contexts but highlight that different forms of positive friction serve distinct roles. Identifying relevant stakeholder groups can help designers better understand the benefits, harms, and risks presented by different AI applications, thus guiding the development of more effective and precise positive friction strategies and avoiding a one-size-fits-all approach that is inadequate for complex issues surfaced by advances in AI.

Recognizing the stakeholders involved also underscores the interconnectedness between AI technology, AI practitioners and users. Frequently, efforts to integrate AI into society foster a “competitive, human-machine opposition mindset,” either through technical approaches to improve the algorithms themselves outside of human oversight or through governance and oversight that create regulations around algorithms, both of which isolate AI technology from human operators (AI practitioners), and collaborators (AI users)[16]. By introducing behavioral speed bumps into AI development, however, developers can more easily surface hidden issues in algorithms that can subsequently enhance user experiences and promote healthier human-AI interaction behaviors. This positions positive friction less as a choice between “assisting developers for AI” or “improving AI for users” or reinforcing a divide between AI developers and end-users, than fostering a collaborative “practitioner-AI-user” dynamic flow through responsible integration of technology and human experience (Fig.2). In short, incorporating positive friction into both the design and deployment of AI can reframe the antagonistic dynamic between human agents and algorithms into a collaborative one, with the potential for human-machine hybrid intelligence.

Grammarly’s recent research on gender-inclusive grammatical error correction in natural language processing (NLP) systems[37] presents an example of this kind of multi-faceted, multi-stakeholder mutual support problem-solving. For AI users, Grammarly’s AI-driven spell-checker and autocorrection function serves as positive friction through pop-up messages questioning personal pronoun usage (’he’, ’she’, or ’they’) or suggestions to replace ’mankind’ with ’humankind,’ which disrupt writers’ unconscious typing behaviors but encourage them to reconsider pronoun biases, or prompt reflections on gender-neutral language. For AI practitioners, flagging corrections’ biases made by the Autocorrection represents positive friction in the form of a trigger during the evaluation of AI, as when Autocorrections suggest changing “Charlie’s earnrings” into “his earnings” or “her earrings”[69]. Although time-consuming, this use of positive friction to conduct internal checks and ’mine clearance’ is not only vital for debugging and improving the AI systems, but also reinforces organizational values in support of LGBTQIA+ inclusivity and a diverse workforce[69, 15]. Grammarly’s commitment to positive friction not just as a device to improve technology but to bolster social and cultural values can be seen in their engineering team’s Counterfactual Data Augmentation(CDA)-supported model for internal Grammatical Error Correction(GEC) systems testing and training, and its subsequent open-source release to the public[37]. This initiative not only enhances self-control to mitigate gender bias in Grammarly’s AI systems, but also encourages the inclusive development of linguistic AI on a broader scale.

The increased prevalence and use of AI requires considering how a diverse range of stakeholders and situations can benefit from positive friction across AI development and end-user engagement. Recognizing that AI/human hybrid behavior as a complex multi-stakeholder unit underscores that identifying the need for positive friction, selecting strategies to address it, and incorporating it into AI, developer, and user dynamics is not a ’one-size-fits-all’ solution. The case of Grammarly, for example, illustrates how different forms of positive friction, sourced from various quadrants in the model, can benefit different stakeholders to improve the offering as a whole. In addition, given the wide range of stakeholders, community values, and application environments, a singular design approach to positive friction may resonate differently across various groups, suggesting that AI design requires a full understanding of relevant contexts before implementation, and further that designing friction solely with specific groups like AI users or practitioners in mind can inadvertently overlook the needs and responsibilities of still other stakeholders. This suggests not only the value of exploring the model’s dynamic relationships in combination rather than as four individual strategies or a singular audience, but also the importance of recognizing how different forms of positive friction might be naturally intertwined. However, more is not always better, and the effective use of positive friction does not require using strategies from all four quadrants. Instead, the design and development of AI requires a rich understanding of context, the nature of potential challenges that might benefit from positive friction, and a sense of which strategies are best suited to the task.

Seeing positive friction in an AI context as a complex design challenge that requires solving for multiple people and problems simultaneously suggests that employing the Positive Friction Model can take three different forms: 1) as a characterizing lens to systematically identify the nature of existing friction, 2) as a diagnostic lens to analyze how situations within complex challenges that would benefit from positive friction, and 3) as a generative lens to craft targeted positive friction that effectively address these challenges.

While integrating behavioral science and design research approaches like the Positive Friction Model more deeply into HAI practice may enhance HCI approaches to socio-technical studies, merging disciplines faces challenges of methodological and definitional alignment. In addition, advances in AI’s capabilities and use continue to evolve nearly on a daily basis. On the one hand, this research heightened the potential value and urgency of employing a behavioral design lens; on the other, key questions about situations of use and potential applications of behavioral tools, such as positive friction, have yet to be defined or even understood. However, despite this uncertainty, there are several potential research questions and further exploratory directions for positive friction in AI worth considering even at this early stage.

First, the Positive Friction Model should be seen not as a stand-alone tool but as a complementary approach that can be used in conjunction with established design frameworks (e.g., journey maps, POEMS[26]) and behavioral tools (e.g., the COM-B framework[39], the Choice Triad[52]), combining familiar methods in novel ways that can potentially reveal insights not previously considered. Employing mixed methods can balance the reliance on a single discipline, merging data-driven confidence with openness to a broader range of solutions; for instance, the use of journey maps that demonstrate forms of multi-stakeholder involvement may yield insight into postures and attitudes involved in various AI-related activities, providing an experiential perspective that the Positive Friction Model currently lacks. This also suggests the potential value of exploring how HCI frameworks or tools can integrate with the Positive Friction Model, whether using HCI approaches to enhance the implementation of positive friction from strategic principles to practical experiments or how targeted use of positive friction can benefit HCI solutions.

In addition, no methodology is entirely neutral[67]; every designer, framework, and solution inherently carries biases, and even methodologies that are scientific and objective (such as behavioral science), openly participatory (as design strives to be), or data-driven and evidence-based (as HCI studies) are influenced by underlying ideologies that dictate disciplinary norms, what qualifies as evidence, and what constitutes a successful solution. Diversity is itself a form of friction, suggesting that successful collaborations across disciplines can benefit from the insertion of deliberate pause points or stimulations that prompt seeing old ideas in new ways. The Positive Friction Model, therefore, can play a significant role within the design process, in which it can inform and support cross-disciplinary exchange as a dialogic tool that helps transdisciplinary teams to identify tensions or blind spots, foster continual conversation, encourage engagement with diverse viewpoints, and position design as an iterative exchange rather than a final decision-making point.

Finally, effective use of the Positive Friction Model and positive friction strategies may change over time as AI itself evolves. Just as the widespread use of LLMs such as chatGPT and visual engines such as DALL-E and Midjourney in the Fall of 2023 introduced both fascination (of their ‘automagical’ abilities to conjure convincing and well-structured content and media) and fear (of AI taking over jobs or concerns about cheating in educational settings), new and emergent forms of AI will also arise. In addition, the distinction between users and practitioners may blur as AI evolves, requiring new or adaptable design strategies to accommodate new use cases that may be difficult to identify in advance. Given that designers are likely to have a front seat as these new technologies are ushered in, designers will also have the ability to influence how AI is seen and adopted by people and more broadly by society[12]. With this ability comes the responsibility of proficiently employing positive friction across the arc of AI development, implementation, and use; minimizing its potential to create unintended consequences, and updating why, when, and how strategies are used in future contexts or to address future challenges in ethical and equitable ways.