From Answer to Action: What Changes When an AI System Starts Acting
A system that drafts text and a system that takes steps raise different questions for a psychology course. Telling five easily blurred terms apart is the first step toward asking the right one.
From a page of text to a trail of steps
In a simple drafting exchange, an instructor receives a response, judges it and decides what to do next. Reviewing the content is a central part of that arrangement. When a system starts to act, by opening files, searching, filling in a document, running a check and trying again, the product of a request includes a sequence of steps, and some of those steps have effects outside the conversation. The first question a psychology course should ask is then not whether the text is good. It is what the system was allowed to do, what it did, and who could see it.
That question gets muddy because several different things share one name. It helps to separate five.
Five terms, five questions
A model is the trained system that turns an input into an output. Depending on its design, it may accept or generate text, images, audio or other representations. What it appears to know in a given exchange depends on the model and on the material supplied in that exchange.
An app is the interface built around a model. It decides what gets sent to the model, what is remembered between exchanges, what is displayed and what can be uploaded. Two apps built on the same model can behave differently because of choices in the wrapper.
A tool is a capability a system can call, such as searching, opening a document, running a calculation or editing a file. Every tool has a scope. It may read, write or send, and it may reach some material and not other material.
An agent, in these pages, is a system in which the model itself chooses which tools to use and in what order, toward a goal, and may loop until it decides it is done. A fixed sequence of steps designed in advance is a workflow and is not necessarily agentic. Many real products mix the two, and the boundary is a matter of degree.
An institutionally approved teaching system is an arrangement that an institution has reviewed for data handling, access, accessibility and permitted use. Approval concerns governance and not capability. It can cover a model, an app and a set of tools in particular configurations; a changed configuration may need reassessment under institutional rules. These five categories overlap, and a single product can be several at once.
A lesson prepared from public pages
Consider an invented case, original to this essay and unevaluated, which describes no existing product. An instructor wants a short lesson on observational learning drawn only from public web pages: a textbook chapter, a university teaching page and two public posts by other instructors. She gives a system that goal and leaves for a meeting. When she returns there is a one-page lesson plan, a list of sources and a log of what the system did.
The first thing to settle is permission. Could the system only read, or could it also edit her course page? Could it send a message or post? A read-only arrangement blocks direct edits to the course page, although reading and sending material to a service still have consequences. An arrangement that can edit a shared page can change something students see, and one that can send messages can reach people before she has reviewed anything. Permission is therefore a teaching decision as much as a technical one, because it sets the scale of what a mistake can do.
The second is review time. Reading the final page is only part of the work. Reviewing the plan means opening the four sources, checking that each claim appears where the log says it does, and reading the log for steps she would not have chosen: a source dropped, a page the system could not open, a claim carried over without a source. That is real work. If it takes longer than writing the plan from the same sources, the arrangement has not saved her time, although it may have shown her a source she would not have found. Either outcome is possible, and which one occurs is something to find out in her own practice.
What an action log shows
The log is evidence, but of a particular kind. A reliable tool record can show which sources were opened and which operations were attempted. It may still be incomplete, and an attempted operation is different from a successful one. It does not show that the lesson is accurate, and it says nothing about whether students will learn from it. A well-kept log can make a flawed plan easy to audit, and a missing log can make a good plan harder to inspect. Neither is a statement about learning. Evidence about action helps reconstruct observable operations; it does not expose the complete internal process. Evidence about learning would come from what students can do afterward, ideally on a case they had not seen, and that is a different kind of data.
What this changes in a course
A psychology course meets this change in three places. In preparation, instructors decide what to delegate and what trail to require. In teaching, a course decides which tasks students may delegate and which of their own decisions it wants to keep visible, such as which source to trust. In assessment, it asks what a finished piece of work shows when many of its steps were taken by a system. None of this calls for one policy. It asks that each choice name the permission, the trail and the evidence of learning it depends on.
The teaching decision
When a system acts, review moves from reading a text to inspecting permissions and steps, and a record of what the system did is not a record of what anyone learned.
Read the evidence separately
These sources support conceptual distinctions or document current developments. They do not validate the original exercises in this essay. Source records state the actual inspection scope.