Post
JA EN

Which Work to Hand to AI, and Which to Process Yourself: Sorting Tasks to Avoid Cognitive Debt

Posted
By
12 min read
Which Work to Hand to AI, and Which to Process Yourself: Sorting Tasks to Avoid Cognitive Debt
  • Who this is for: People who use AI daily to produce documents, summaries, or code, and who often need to use that content themselves later
  • Assumed knowledge: Basic experience with AI tools like ChatGPT or Claude
  • Reading time: 10 min

Overview

The sense that “material AI writes for you doesn’t stick” has backing from an MIT Media Lab EEG experiment and from three theories in cognitive science: the levels-of-processing effect, the generation effect, and desirable difficulties. When AI hands you a finished product, the three routes that make memory durable (deep processing, self-generation, and moderate friction) all get bypassed at once. The “why” of this is covered in a separate article.

But knowing the cause runs into a wall in practice: “So I’m supposed to put effort into everything I make with AI? That’s not realistic.” That reaction is correct. There are plenty of tasks where you should not add the effort back. Adding a manual step to format conversions, boilerplate, or a throwaway draft is a waste of time, and in some cases it actively backfires.

So the thing that matters isn’t “always do the work yourself.” It’s sorting. Will you need to recall or explain this content later? Do you have the room (time and background knowledge) to think it through right now? Ask these two questions first, and most tasks fall onto the “fine to hand to AI” side. Only the small remainder need one step of your own processing.

This article lays out how the effect of cognitive debt changes from one type of task to another, and gives you a basis for sorting work into “worth caring about” and “fine to delegate.” Then, for the tasks you decide to keep, it gets concrete about which single step is most worth keeping for yourself. No list of vague good intentions.

First, sort by whether the task is worth caring about

The piece most often missing from discussions of cognitive debt is this sorting step. Trying to apply “process it deeply” and “generate it yourself” to every task breaks down. You end up adding effort back to work that has no memory value, and the whole reason for using AI evaporates.

Start with two questions.

  1. Will you later need to recall, reconstruct, or explain to someone the substance of this work? For work with little substance worth remembering, like format conversion or copy-paste cleanup, or a draft you use once and discard, it doesn’t matter if it doesn’t stick.
  2. Do you have the time and background knowledge to think it through right now? When you lack the background to even check or integrate AI’s output, adding a manual step just spins in place. The same goes when you’re in a hurry and completion is all that matters.

Only tasks that answer “yes” to the first and “yes” to the second are “worth keeping a hand in, and possible to keep a hand in right now.” If either is “no,” you’re better off handing it to AI, or deferring it and revisiting when you have room. As a diagram:

flowchart TB
    Q1["Will you need to recall or<br>explain this content later?"]
    Q1 -->|No| Free["Fine to hand to AI<br>(formatting, boilerplate, throwaway)"]
    Q1 -->|Yes| Q2["Do you have the time and<br>background to think it<br>through right now?"]
    Q2 -->|No| Later["Take the finished output now,<br>revisit it later"]
    Q2 -->|Yes| Keep["Keep one step yourself<br>(generate / paraphrase / retrieval)"]

How the debt lands, task by task

Even within “having AI make it,” the effect on memory shifts with the type of task. The one thing the MIT study directly examined was essay writing, a single task type, so applying it to other tasks means estimating, case by case, which memory mechanism is at work1. What follows is extrapolation from the three cognitive-science theories, not a direct measurement of each task.

Having AI summarize a meeting or document. This is the closest to essay writing. Summarizing is fundamentally a generative act of “compressing information yourself,” so handing it to AI gives up both the generation effect and desirable difficulty. It’s the textbook case of not being able to explain the content the next day. Summaries are also the format where the caveats and exceptions in the source material drop out most easily. High value in keeping a hand in.

Research and looking things up. In the MIT study, the neural connectivity of the group using only a search engine was stronger than the AI group’s, sitting between it and the unaided group12. Searching is active scanning of on-screen information, so more of your own processing survives than when you have AI answer directly. Flip it around: rather than having AI produce “the answer itself,” synthesizing the pieces yourself makes it stick. This is closer to a difference actually observed within the study.

Code generation. Use AI-written code without understanding it and the design and logic never take root. Code has a specific wrinkle, though: running it returns feedback on the spot, and tests can function something like retrieval practice. It affects “can you maintain this later” and “can you respond when it breaks” more than raw memory.

Decision documents and design choices. Here what’s at stake is ownership more than memory. Unless you lay out the options yourself and weigh the trade-offs yourself, the basis for the decision never lodges inside you. Later you can’t reconstruct “why did we decide this,” and accountability gets harder. Given that the AI group’s sense of ownership was the lowest in the MIT study, this isn’t one to shrug off1.

Routine, low-cognitive-value work. Generating boilerplate, converting formats, canned standard text. These have little substance worth retaining in the first place. No need to add effort back here. Better to redirect the time AI frees up toward the “high value to keep a hand in” work above.

TaskValue of keeping a hand inWhat it affects
Summaries (meetings, documents)Highmemory, retention of caveats
Research / looking things upHighmemory, your own synthesis
Decisions / design choicesHighownership, accountability
Code generationMediummaintainability, incident response
Boilerplate / formattingLow(little substance worth retaining)

If you keep one step, which one?

Even for the work you sorted as “high value to keep a hand in,” you don’t have to redo all three bypassed routes (deep processing, self-generation, moderate friction). Insert even one step of your own and the odds of it sticking change substantially. Which single step you choose depends on the nature of the task. For work that means “compressing or synthesizing things yourself,” like summaries and research, generating and paraphrasing work well; for work that turns on ownership, like decisions, using AI as a critic works better.

Put out your own take before asking AI. For a summary, a research question, or a decision, sketch your own answer or structure once, then hold it up against AI’s output. That gets you the benefit of the generation effect while still using AI’s efficiency. A one-line note is fine. Just reversing the order keeps the thinking process on your side.

Restate what you receive in your own words. Take an AI-produced summary or document and try to boil it down into a single sentence yourself. That restating is elaborative rehearsal, and it triggers the deep processing of meaning-making and connection to what you already know3. That’s the decisive difference from passively reading.

Use AI as a critic, not an author. Build your own hypothesis or draft first, then have AI poke holes in it. The initiative for generation stays on your side, and you still get AI’s checking ability. For work where ownership matters, like design choices, this shape fits especially well.

Recall it after some time has passed. Rather than using a document the moment you make it, leave a gap and then try to recall its content from memory. This is a form of spaced repetition and retrieval practice, a “desirable difficulty” that lets you make up after the fact for the “effort of recalling” AI skipped. It also works as a way to insert processing retroactively into work you took as a finished product because you had no room at the time.

What to watch out for here is mistaking the feeling of having understood for understanding itself. Learning research has shown, over and over, that the sense of processing going smoothly (fluency) often correlates inversely with actual retention4. Reading AI’s output fluently doesn’t mean it went in. Measure comprehension not by “does it feel like I got it” but by “can I reconstruct it without looking.” The paraphrasing and delayed recall above double as that measurement.

When not to add effort

Take all of this as “just add a manual step everywhere” and you fail in the opposite direction. “Desirable difficulty” carries an important caveat: difficulty beyond a learner’s background and scaffolding turns into “undesirable difficulty” and actually lowers retention4. Even generating something can be harmful to memory when you lack the background, generate it wrong, and get no feedback5. Effort isn’t good everywhere.

In situations like these, the right move is to take the finished product as is, without adding effort.

Routine work with little substance worth retaining has nothing to keep in the first place. When you’re under heavy time pressure and completion, not memory, is the goal, don’t stop your hands either. When you lack the background and can’t even verify your own generation, adding a step just spins. There’s no point building in a hook for later recall on output you use once and throw away. And above all, pushing effort onto someone who is thinking “I don’t want to think right now, just give me the answer” only erodes their autonomy.

Here’s the principle. An intervention to keep a step for yourself matters only when memory, understanding, or ownership is the goal of the work. If the goal is “just get it done,” handing it to AI is the optimal move. Imposing effort uniformly while mistaking the goal invites undesirable difficulty, which is not what the cognitive-debt argument recommends.

Conclusion

Using AI without running up cognitive debt isn’t about shrinking the scope of what you hand to AI. It’s about sorting which work to do by hand and which to delegate. Pick only the work you’ll need to recall or explain later and have room to think through now (summaries, research, decisions, and the like) for the “keep a hand in” side, and insert one of generation, paraphrasing, or retrieval practice there. Routine work, throwaway output, and no-room situations can go straight to AI.

You don’t need to redo all three memory routes yourself. One is enough. Try to impose effort on everything, in fact, and you invite undesirable difficulty and lose the point of using AI. Sort, narrow down, and keep just one step. That’s the answer the science of cognitive debt offers to everyday practice.

Want the “why” behind AI-produced material not sticking, the neuroscience and cognitive science of it? See the companion piece, Why Material AI Writes for You Doesn’t Stick: The Science of Cognitive Debt. This article is its practical half.

References

References are listed in the order their citation numbers appear in the text.

  1. Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task - Kosmyna, N., Hauptmann, E., Yuan, Y.T., et al. MIT Media Lab, arXiv preprint (June 2025). EEG measured across 54 participants over four months, split into LLM, search, and no-tool groups. A pre-print, not yet peer-reviewed, with a limited sample of 54 (18 in session 4). Limited to a single task (essay writing) and to ChatGPT, so it does not generalize to other tasks or other LLMs. The task-by-task extrapolation in this article includes inference from cognitive-science theory rather than direct results of the study. 【Reliability: Medium】 ↩︎ ↩︎2 ↩︎3

  2. Your Brain on ChatGPT - Armitage, R. British Journal of General Practice, 75(758), 410 (2025). A commentary summarizing Kosmyna et al. from a medical-education angle, published in a peer-reviewed journal (not the original study itself). 【Reliability: Medium-High】 ↩︎

  3. Levels of processing: A framework for memory research - Craik, F.I.M., & Lockhart, R.S. Journal of Verbal Learning and Verbal Behavior, 11(6), 671-684 (1972). The original paper on the levels-of-processing effect, cited and tested as a foundational theory for over 50 years. 【Reliability: High】 ↩︎

  4. Memory and Metamemory Considerations in the Training of Human Beings - Bjork, R.A. In Metcalfe, J., & Shimamura, A. (Eds.), Metacognition: Knowing about Knowing, pp. 185-205. MIT Press (1994). The original paper proposing “desirable difficulties.” The divergence between processing fluency and retention, and the upper bound of desirable difficulty (difficulty beyond one’s background becomes undesirable), are part of this framework. 【Reliability: Medium-High】 ↩︎ ↩︎2

  5. The generation effect: Delineation of a phenomenon - Slamecka, N.J., & Graf, P. Journal of Experimental Psychology: Human Learning and Memory, 4(6), 592-604 (1978). The original paper establishing the generation effect. Later research also documents that generation can be harmful under conditions of failed generation without feedback (the negative generation effect). 【Reliability: High】 ↩︎

AI・Technology
AI cognitive science cognitive debt memory learning
This post is licensed under CC BY 4.0 by the author.