When a student hands work to a machine, what kind of agency is that?
The field keeps asking whether AI helps or harms learning. That's the wrong question. The answer depends on a belief we don't yet measure.
Sixty years of Bandura's social-cognitive theory give us the vocabulary for it. This is the construct at the center of my research, AI Proxy Efficacy, and why "does AI help learning?" has no answer until you know a student's calibration.
You can act three ways.
Bandura distinguishes three modes of agency, and each runs on its own efficacy belief. Act yourself, act through others, or act together. Tap each:
Self-efficacy asks can I? Proxy efficacy asks can it?
A student's task-specific confidence that a generative-AI system can perform a delegated academic task effectively on their behalf.
Defined here, building on Bray et al. (2001, p. 426) and Hanham et al.'s technological proxy efficacy (2014, p. 4).It is not trust, not perceived usefulness, not confidence in yourself. It is a capability belief about a third party, pointed at a task you have delegated.
Proxy is a relationship, not a piece of software.
The same ChatGPT is a tool when you check spelling, a proxy when it writes the essay you submit, a partner when you argue with it to sharpen a thesis you then write yourself. What moves it along the line is how much of the work (and the authorship) you delegate.
It's not the level of the belief.
It's the calibration.
A high proxy-efficacy belief is neither good nor bad in itself. What matters is whether it tracks what the AI can actually do on this task. Pick a task; drag your belief against the AI's real capability.
The same belief can grow you or erode you.
Bandura never resolved this. In 1982 he warned that leaning on a proxy "reduces opportunities to build the requisite skills." By 1997 he allowed the opposite: a proxy can "free time and effort to enhance personal efficacy in other areas." Which branch a student lands on is not fixed. It turns on conditions we can measure and change:
The central, testable claim (H1). APE → self-efficacy is positive or negative depending on measurable moderators, which is exactly why the belief is worth measuring rather than assuming. The same construct has opposite downstream signs; interventions are predicted to move students from the substitutive branch to the complementary one.
What proxy efficacy is not.
A reviewer's first objection: isn't this just trust, or perceived usefulness, relabeled? No. The difference is structural, not cosmetic.
A learner-centered construct AIED doesn't yet have.
The GenAI-in-education literature is rich in trust, usefulness, literacy, and dependency, but has no validated construct for a student's belief in the AI's capability to act on their behalf, and no model of how that belief builds or erodes their own competence. APE is that missing piece: purpose-built, domain-specific, embedded in a moderated model of the substitution–complementarity fork.
That turns a vibe ("does AI help learning?") into something decidable and intervenable. AI literacy, calibration training, capability transparency, and task design become the levers predicted to move students from the substitutive branch to the complementary one.
This is new. I'm building the instrument and the model.
Proxy efficacy for a generative-AI proxy: defined, measured across the domains students actually delegate, and tested with real learners. If you work on human–AI interaction, AI literacy, or learning at scale, I'd like to talk.