Why “Does AI help?” is too broad

AI can improve one result and weaken another. A useful theory has to name what the person and AI each did, and when the outcome was measured.

About 20–30 minutes · 13 named steps · research detail is optional

A student may finish a practice problem faster with AI and remember less a week later. An employee may draft a stronger memo and still be unable to defend its assumptions. Both statements can be true.

The question “Does AI help?” combines different people, tasks, workflows, and outcomes. AIEM separates them so we can ask what kind of engagement helped with which outcome, under which conditions.

The conflation AIEM is designed to remove

Generative AI often improves what a person can produce while the tool is present. Later capability is a different outcome. Some experiments find losses in unaided learning or retention, while guarded use narrows or removes the loss.

These results vary along at least three dimensions: the function the AI performed, the comparison condition, and the delay before measurement. Change any of the three and the answer can change. That pattern is theoretically useful because it points to engagement as a variable that exposure-only studies leave hidden.

AI engagement is not a property of a tool or a single prompt. It is a pattern produced jointly by a person, a tool, a task, and a context. If that is right, then the useful questions become more precise: what cognitive mode did the task invite, what did the person keep hold of, what did the AI perform, and what outcome followed?

Those questions need a mechanism before they need a vocabulary. Cognitive Continuum Theory supplies the first part of that mechanism; Evaluative Agency Theory applies it to AI. They come next.

Quick check

Why can well-run studies of AI assistance reach different conclusions?