AI Can Improve the Essay and Shrink the Learning. Audit the Thinking First.

AI can hand a student a better essay without making that student a better writer. The American Psychological Association said so on September 3. This week two psychologists turned the warning into a test you can run on one assignment before Monday.

The two are Beth M. Schwartz of the APA and Christopher Hakala, a psychology professor who directs the teaching center at Springfield College. In an Inside Higher Ed opinion piece on October 22 and a longer essay in Faculty Focus on October 23, they make an argument that should take some pressure off every faculty member who has been told to wait for the research. Research on generative AI is new. The science of learning is not.

That sounds like a slogan until you see what the science of learning lets you do. It moves the AI decision out of the policy committee and into the assignment, where you already have the expertise to make it.

Why waiting for the AI research is the wrong plan

The honest state of the evidence is thin, and the authors say so first. They were responding to an Inside Higher Ed report that colleges are investing heavily in generative AI while evidence about when, how or whether these tools improve student learning remains limited. In Faculty Focus they go further: much of the current conversation runs on perceptions, anecdotes and survey data, and even large surveys cannot tell us whether students who use AI learn more, learn less or simply learn differently.

That gap is real. It is not a reason to freeze. Students are using these tools this term either way. So the authors change the question. Not “Does AI work?” but “Does a particular use of AI strengthen the cognitive processes necessary for learning, or does it replace them?”

That second question has an answer today, because psychologists have spent decades studying what has to happen in a student’s head for learning to stick.

What the APA report found

The APA expert report on children’s and adolescents’ learning with educational technology was written by fifteen authors, including Richard E. Mayer of the University of California, Santa Barbara and Daniel T. Willingham of the University of Virginia. It covers learners ages 5 to 18 in K to 12 settings. It is not a college study, and it does not pretend to be one.

Read its six broad conclusions anyway, and three of them sound like a first year writing course in 2026:

• Time spent, enjoyment, participation and completion rates are often mistaken for learning.
• The real test of learning happens after the app is closed. Students may perform well within a tool but fail to retain or apply what they learned elsewhere.
• AI may improve performance while reducing learning. Students can produce better work while developing less independent knowledge and skill.

Nicole Barnes, the APA’s executive lead psychologist for education, put the third point in terms every instructor will recognize: “AI can help a student produce a better essay, but the improved final product does not necessarily mean the student has become a better writer.” Then she added the line that matters most for anyone who grades: “Just as we need to separate engagement from learning, we will increasingly need to separate performance from learning.”

Schwartz and Hakala argue that the same principles apply to college students. Hold that as an argument from learning science, not as a new finding about undergraduates. It does line up with what has been showing up in gradebooks. When a Berkeley study this spring found a sharper rise in A grades in courses heavy on writing and coding, its author, Igor Chirikov, told The Wall Street Journal the results suggested students relied on AI to do better in their studies, not that those classes were learning more.

The polished product stopped being proof

For most of the history of college teaching, the finished product stood in for the process. A clear essay meant someone had wrestled the ideas into order. A correct solution meant someone had worked the problem. We graded the product because it was the best evidence we could collect of the thinking behind it.

Generative AI broke that link without announcing it. Faculty Focus puts it in one sentence: “when AI sidesteps effort and leads to a strong product, the process suffers, and that process is where learning happens.”

Three science of learning principles that need no new study

The Faculty Focus essay names three findings the authors call among the most replicated in the learning sciences. None is about AI. All of them decide whether an AI use helps.

Learning takes effort. Students learn more when they actively retrieve information, generate explanations and struggle productively with material. Tools that replace thinking rather than supporting it are likely to impair learning, the authors write, whether they are textbooks, calculators or AI chatbots.

Feedback works when it leads to revision. Feedback is most effective when it is specific, explanatory and connected to a chance to revise. AI produces feedback fast, but speed does not guarantee learning value, and AI often cannot respond to the context of your course the way you can.

Students have to know what they do not know. Metacognition, a student’s awareness of their own understanding, is part of learning. When a tool creates an illusion of fluency, a student can feel confident without having mastered anything.

The five question filter

In Inside Higher Ed the authors turn those principles into a filter. Start with a learning goal, not a tool, and ask:

1. What must students know or be able to do?
2. What thinking must they practice to achieve that goal?
3. Which cognitive work must remain with the student?
4. Is AI extending that thinking or doing it for them?
5. Can students demonstrate the learning independently?

Question three is the one that changes an assignment. Most AI policies are written at the level of the course: allowed, banned, allowed with disclosure. The filter works at the level of the thinking step, which is where the learning actually happens.

What it looks like in one assignment

The authors give two examples. In the first, students produce an explanation or solve a problem on their own, then use AI to generate questions that push their reasoning further. In the second, students finish a first draft, then ask AI to identify an unclear claim, missing evidence or a potential counterargument. In both, the student decides which feedback to accept or reject and explains why.

Notice the order. The student’s own attempt comes first, so the effort happens. The AI comes second, as a challenger rather than an author. The student’s judgment comes last, so the metacognition happens too. The APA report recommends something close to this for younger learners: teach students to question a chatbot’s assumptions, find gaps in its reasoning and compare its answers against other sources, so they treat it less as an answer machine and more as something they are testing.

Then add the APA’s real test. After the AI step, give students a short task with the app closed: explain the argument out loud, or solve a fresh problem of the same kind on paper. If they can do it, the AI extended their thinking. If they cannot, it did the thinking for them, and you found out while there is still time to fix it.

This is the same thread we pulled on Wednesday in AI cheating as a design problem and in September with the reading assignment doom loop. The more fluent the tools get, the more the evidence of learning has to come from what a student can do without them.

What this does not settle

The filter is a design tool, not a verdict on any product. The APA report studied younger learners, and the research on generative AI in college courses is still early. The authors call for smaller, iterative studies that examine student work before and after AI assistance and test whether students can transfer what they learned without AI. Some uses will turn out better than the principles predict and some worse. That is why the after the app is closed check belongs in the assignment itself. It is a small study you run every week on your own students.

Try it on one assignment this week

1. Write the goal as something a student can do. Not “understand the causes of the war” but “explain two causes of the war and defend which mattered more.”

2. List the thinking the assignment asks for. Retrieve, organize, connect, explain, find a gap, correct a mistake, apply to a new case.

3. Mark each step. Student only, AI after a first attempt, or AI is fine. Be strict about student only. That is the learning.

4. Write one sentence for students. Say where AI is welcome and where it stops, in the language of the steps you just marked.

5. Add the check. Design the short task students do with the app closed, and decide now what a passing answer looks like.

Your next step

Take one assignment through The Cognitive Work Audit, by hand, in about 20 minutes.

You will hold a finished AI use plan: the goal, every thinking step marked, the one sentence instruction for students, and the after the app is closed check. Then the next thought arrives on its own. That is one assignment. Your course has a lesson every week, and every one of them now carries the same question about which thinking stays with the student.

Auditing one by hand is how you learn the method. Planning a whole term that way is the slow part. The Lesson Plan Builder builds timed plans and handouts, grounded in your Course Record, the five minute record of your course, your outcomes, your policies and your voice that every tool in the toolkit uses. Bring the plan you build today as the pattern, and do the first one by hand anyway.

If you would rather think it through with other instructors first, the Harvest Kernel Learning Community is free, and this question is the conversation there this week.

Stop asking whether AI works. Ask what your students can do after the app is closed.

Sources

• Beth M. Schwartz and Christopher Hakala, “Research on AI and Learning May Be New. The Science of Learning Is Not.,” Inside Higher Ed, October 22, 2026. insidehighered.com
• Beth Schwartz and Chris Hakala, “AI in Teaching: What the Science of Learning Already Tells Us,” Faculty Focus, October 23, 2026. facultyfocus.com
• American Psychological Association, “Children’s and adolescents’ learning with educational technology,” expert report, September 2026. apa.org
• American Psychological Association, “With education technology, engagement is not the same as learning,” press release, September 3, 2026. apa.org
• The Wall Street Journal, “‘A’ Grades Are Suddenly Everywhere Since the Arrival of ChatGPT,” May 2026. wsj.com