Most adoption models treat resistance as the absence of adoption. But resistance is its own construct with its own causes, and ignoring it means your model is doing half the job.
Venkatesh et al. (2003) consolidated eight acceptance models into UTAUT and explained 70 percent of the variance in behavioral intention. I have read that sentence dozens of times across papers and textbooks and study notes. The number is impressive. But the thing that kept bothering me each time I revisited the model was who gets left out of that 70 percent. The model predicts who will adopt. It does not explain who will not, or why some people never come around, or why resistance stubbornly persists even when the technology is free, fast, and easy to use.
UTAUT identifies four core determinants of intention and use: performance expectancy, effort expectancy, social influence, and facilitating conditions, with age, gender, experience, and voluntariness as moderators. As I wrote about when counting users stopped being enough, these predictors explain intention, but they do not explain what happens after people start using a system, let alone why some people never start. The logic is straightforward. If you believe a technology helps you perform, find it easy to use, see people around you using it, and have the resources to do so, you will adopt it. When someone does not adopt, the model treats that as low scores on one or more of these predictors. The non-adoption is a gap to be filled, a deficit to be corrected. More training, better interface design, stronger social norms, more organizational support. The assumption built into the framework is that adoption is the goal and resistance is deviation from that goal.
Parasuraman (2000) offered a richer vocabulary for the negative side through the Technology Readiness Index. His framework splits technology readiness into two drivers and two inhibitors. Optimism and innovativeness push people toward technology. Discomfort and insecurity pull them away. Discomfort is feeling overwhelmed by technological complexity and lacking a sense of control over it. Insecurity is distrust, skepticism about whether the technology works as promised, and worry about data loss or malfunction. These are not the same thing, and neither of them maps neatly onto low performance expectancy. A person can believe that an AI assistant would make them more productive, and still refuse to use it because they do not trust the output. The adoption model would register high performance expectancy and low intention, and treat that as a puzzle. Parasuraman's framework would say: that is insecurity, and it has its own logic.
I keep thinking about elderly users and smart home devices. The resistance there is not usually about inability. Many older adults operate complex remote controls, drive cars, manage medical regimens that would confuse someone half their age. What they resist about a voice assistant or a connected thermostat is often the discomfort dimension. They feel the device is always listening. They feel an erosion of control over their own living space. They worry, and this is the insecurity dimension, about what happens when the system fails and there is no person to reach. UTAUT can model their low intention scores, but it cannot name the mechanism. The model sees a flat score. Parasuraman sees a shape with distinct causes.
The same pattern shows up with knowledge workers resisting AI assistants. I am not thinking of people who find the tools confusing. I am thinking of senior professionals who understand the tools perfectly well, who could use them if they chose to, and who choose not to. When a senior analyst refuses to use a self-service analytics dashboard because they do not trust the numbers, the problem is not effort expectancy. It resonates more with what Strich et al. (2021) identified about AI and professional identity: when technology reshapes what counts as expertise, people whose identity is built around that expertise experience the change as a threat, not an opportunity. The resistance is not about the features of the tool. It is about what using the tool means for who they believe they are, a point I explored in how identity shapes technology adoption.
The study-hub's innovation resistance materials formalize this idea through Ram and Sheth's (1989) framework, which identifies functional barriers like usage difficulty and value misalignment alongside psychological barriers like tradition and image. The IS field's behavioral adoption chain, running from TRA through TAM through UTAUT, is good at modeling the functional side of why people adopt. Bagozzi (2007) and Silva (2007) already critiqued the field's overreliance on narrow acceptance models. The chain needs extending, because resistance is not just low adoption. It has its own antecedents and its own structure.
Bhattacherjee (2001) takes a different path but arrives at an adjacent insight. His expectation-confirmation model for IS continuance draws on Oliver's (1977) consumer satisfaction logic: users form expectations, experience the system, compare the experience to the expectations, and the resulting confirmation or disconfirmation drives satisfaction and continuance intention. The model is about what happens after someone adopts. It explains churn and discontinuance. It is not a non-adoption theory. But it reveals something useful. If someone expected a system to be untrustworthy before they ever started, their early experience confirming that expectation does not represent a failure of the model. The model is working as intended. Confirmation cuts both ways. A person whose insecurity tells them the AI will produce unreliable output, and who then encounters an error in that output, has had their resistance confirmed. Expectation-confirmation theory explains why early negative experiences lock in resistance, but it does not explain the initial disposition.
That is where the concept of technology readiness fills the gap. Readiness is not a binary switch. It is a disposition that varies between people, and the variation is not random. People high in optimism about technology believe it gives them more control over their lives. People high in innovativeness like experimenting with new tools. People high in discomfort feel overwhelmed and want a human in the loop. People high in insecurity do not believe the promises the technology makes. A model that measures only positive expectancies will misclassify the last two groups. It will see them as people who just need more information, a better tutorial, or stronger social pressure. Sometimes they need those things. Sometimes they have a reasoned, experience-based judgment that the technology does not fit their needs, and no amount of training will change that because the problem is not knowledge, it is trust. As I argued in trust versus delegation, trust and reliance are separate constructs, and calibrated trust means knowing when not to rely.
A few years after UTAUT, Venkatesh et al. extended the model into UTAUT2 for consumer contexts, adding hedonic motivation, price value, and habit. The habit construct is especially relevant here because it suggests that non-use can also be habitual, an entrenched pattern of doing without a technology. But even UTAUT2 keeps adoption and use as the dependent variables. The outcome being modeled is still movement toward the technology. Resistance remains the unexplained background noise.
I think the IS field has implicitly defined successful adoption research as research that predicts and increases adoption. Papers that model who adopts accumulate citations. Papers that try to understand the structure of resistance are less common and, I suspect, less rewarded by reviewers who see non-adoption as a null result rather than a phenomenon worth explaining on its own terms. But when nurses document patient interactions on paper while an EHR sits open on the screen next to them, when engineers push back on a genuinely superior project management tool, when a senior consultant refuses to delegate summary writing to an AI that could do it in seconds, the resistance is not noise. It is data about trust, identity, and what people believe their work requires of them. If your model only predicts adoption and not resistance, it is doing half the job, and the half it misses might be the half that tells you whether your intervention is pointed at the right problem.
The question worth asking is not only what predicts adoption. It is also what resistance predicts. If it turns out that resistance in some contexts predicts more careful judgment, higher quality control, or better decisions about when not to trust a system, then the field needs to stop treating it as a problem to be eliminated and start treating it as a signal worth interpreting.
About the author
Share
More notes
Related notes