AI Self-Efficacy Is the New Digital Divide

Compeau and Higgins published their computer self-efficacy measure in 1995. Twenty seven years later, Compeau, Correia, and Thatcher (2022) argued that the construct might be obsolete. Not because self-efficacy stopped mattering, but because the technology it referred to had changed. Their primary example walked through updating computer self-efficacy to IT self-efficacy, and the paper gave the field a systematic method for evaluating when any construct has outlived its original context. I read this paper three times this semester. The first time it was about measurement. The second time it was about how the IS core evolves through construct reformulation. The third time it was about the digital divide.

Bandura (1977) defined self-efficacy as the belief in one's ability to execute a behavior required to produce a desired outcome. The idea is deceptively simple. If you do not believe you can do something, you are less likely to try, less likely to persist when it gets hard, and less likely to learn from the attempt. Social cognitive theory (Bandura 1986) made this reciprocal: personal beliefs, behavior, and environment shape each other in a continuous loop. Compeau and Higgins (1995) brought this into IS by developing and validating a measure of computer self-efficacy, the belief that one can successfully use a computer to accomplish a task. The construct became foundational. It showed up in PMT as part of coping appraisal, in training research as both an input and an outcome, and across adoption models as a determinant of ease of use. The logic was always the same: if you doubt your ability to use the tool, you will not persist, learn, or perform well no matter how useful the tool looks on paper.

That logic has not changed. The tool has.

AI self-efficacy is the belief that you can use AI tools effectively to accomplish your goals. It is computer self-efficacy for a world where the interface is conversational, the output is stochastic, and the human role has shifted from operator to delegator. I think it is replacing computer self-efficacy as the construct that actually predicts who benefits from technology. The reason is not that computers have disappeared. It is that the nature of the interaction has changed. Using a spreadsheet or a CRM system requires a predictable sequence of actions where the user controls the input and the system produces deterministic output. Using an LLM requires something different: crafting prompts, evaluating outputs, iterating on failures, knowing when the model is hallucinating, and deciding what to delegate versus what to verify. These are skills that depend heavily on whether you believe you can do them.

Compeau and Higgins predicted this structure without needing to name AI. Their framework, grounded in social cognitive theory, argued that self-efficacy shapes effort, persistence, learning, and effective performance. People with high computer self-efficacy invest more effort, persist longer through difficulty, and use tools more effectively. People with low computer self-efficacy expect frustration, put in less effort, and disengage at the first sign of trouble. Now apply this to the same LLM. Two people open ChatGPT at the same time. One has high AI self-efficacy. They experiment with prompts, try different phrasings, test the model's boundaries, and integrate the output into their work. The other has low AI self-efficacy. They type one question, get an answer that is fine but not great, assume that is all the tool can do, and close the tab. Same tool. Same access. Radically different outcome. I wrote about what effective use means for AI tools, and the mechanism here is the same: the quality of use depends on the user's engagement with the deep structure, and self-efficacy determines whether that engagement happens.

This is the new digital divide. It is not about who has a subscription to ChatGPT or Claude or Gemini. Those subscriptions cost twenty dollars a month. Access is not the bottleneck. The bottleneck is whether you believe you can use the tool well enough for it to matter.

Consider prompt engineering tutorials. I see them everywhere now. Write a system prompt with role, context, and output format. Use chain of thought. Break complex tasks into subtasks. These are genuinely useful techniques. But who do they help? The person who already believes they can learn to prompt effectively. The person who reads a tutorial, tries the technique, fails the first time, adjusts, and tries again. That is high self-efficacy behavior. The person with low AI self-efficacy reads the same tutorial and feels overwhelmed. There are too many rules. The examples are too specific. The failure on the first attempt confirms what they already suspected: they are not good at this. The tutorial did not create the gap. It widened a gap that already existed.

I see the same pattern in the generational divide around AI. The common narrative is that younger people adopt AI faster because they grew up with technology. I think the mechanism is more specific to self-efficacy. Younger professionals have had more opportunities to build AI self-efficacy through repeated low stakes experimentation. They tried ChatGPT in college for an assignment. They used it to draft a message. They asked it to explain something they needed to understand. Each successful interaction raised their self-efficacy slightly. Each failure was low cost. Older professionals have had fewer of these opportunities. Their first interaction with AI might be in a high visibility context where failure matters, which makes the stakes higher, the anxiety greater, and the self-efficacy lower. The generational gap in AI adoption is not about digital nativity. It is about accumulated self-efficacy building experiences.

I think this is where the IS field needs to pay attention. Compeau, Correia, and Thatcher (2022) gave us the method for evaluating construct obsolescence. The question their paper forces is whether computer self-efficacy still captures what matters when the interface is generative, the output is probabilistic, and the human role involves delegation, verification, and judgment. I wrote recently about why delegation has replaced use for agentic systems, and AI self-efficacy is the parallel move on the individual capability side. The relevant self-efficacy is not about operating a system. It is about managing a relationship with an agentic artifact, evaluating its outputs, and deciding when to trust versus when to override. When I wrote about social cognitive theory and training, the same principle applied: enactive mastery experiences are the strongest source of self-efficacy, and training programs that skip them will not shift beliefs.

The policy implication is uncomfortable. If the digital divide in the AI era is driven by self-efficacy rather than access, the usual solutions will not work. Giving everyone a free AI subscription does not close the gap. Training programs that teach prompt engineering as a technical skill miss the point if trainees do not believe they can learn it. The intervention that works for low self-efficacy users is not more information. It is structured success experiences: small, scaffolded tasks where the user succeeds repeatedly until the belief shifts. Bandura (1977) called these enactive mastery experiences, and they were always the strongest source of self-efficacy, stronger than verbal persuasion, vicarious experience, or physiological state. The field has known this for almost fifty years. We just forgot to apply it to AI.

I do not think the gap between high and low AI self-efficacy users will close on its own. I think it will widen. The people who already believe they can use AI effectively will keep accumulating positive experiences, building skills, and compounding their advantage. The people who do not believe it will avoid the tool, fall further behind, and interpret every AI advancement as confirmation that the world moved without them. The defining inequity of the next decade is not who has access to AI. It is who believes they can use it effectively.