Why it is worth making an effort with GenAI

Yvonne Rogers

REVIEW 4 major objections 5 minor 8 references

Why it is worth making an effort with GenAI

T0 review · 4 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read Requiring students to use generative AI more effortfully—writing rationales before feedback, critiquing output, taking handwritten notes—may deepen learning and give a greater sense of achievement.

desk verdict A worthwhile design provocation that applies the effort paradox to GenAI learning, with the empirical support a little thinner than the summary lets on. read the letter →

arxiv 2509.00852 v1 pith:M6XOGJ2J submitted 2025-08-31 cs.HC cs.AI

Yvonne Rogers This is my paper · ORCID

classification cs.HCcs.AI

keywords generativeAIineducationeffortparadoxIKEAeffectcriticalthinkingmetacognitionhuman-AIinteractionlearningdesigncognitive

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper argues that the widespread student habit of letting ChatGPT produce essays and homework at minimal effort is quietly eroding the thinking skills that writing itself builds. To counter that, it introduces the effort paradox—people avoid effort yet also value it, as in the IKEA effect—and applies it to learning: AI tools should sometimes make students work harder, not always easier. The paper proposes designing GenAI tools that stage effort across a task, letting AI help with getting started and then requiring students to evaluate, critique, verify, and build on what it generated. It supports this with early evidence from studies of AI assistants that scaffold reflection, a comparison of an advisory AI with one that demands users write their own rationale first, and a survey showing knowledge workers already shift their effort to verifying AI output. The payoff, if the argument holds, is a way to keep the convenience of GenAI without sacrificing critical thinking, metacognition, or the sense of accomplishment that comes from hard-won understanding.

What carries the argument

The paper's central mechanism is the "effort paradox"—the finding that effort is simultaneously costly and valued—deployed through a design pattern of staging cognitive effort across task phases: low effort when starting with GenAI (getting ideas, plans, drafts) and higher effort afterward (verifying, critiquing, iterating, note-taking). Named vehicles include "Proberbots" (chatbots that nudge reflection), "ExtendAI" (an AI that requires users to write their rationale before receiving feedback), and "SelVReflect" (a voice/VR tool for guided reflection), plus the "IKEA effect" as the motivational analogy explaining why invested labor increases felt value.

What would settle it

A randomized classroom experiment: one group writes a rationale before receiving AI feedback, another gets AI recommendations directly; if the rationale group does not score higher on delayed tests of understanding or reports more frustration, the claim that required effort deepens learning is weakened.

Watch

Extended reading notes

Core claim

The central claim is that the effort students avoid when they outsource homework to ChatGPT is the same effort that makes learning feel worthwhile, so the aim should not be to ban generative AI but to design it to deliberately require more effort at the right moments. The paper names this the effort paradox—effort is costly and also valued—and uses the IKEA effect, where labor invested in an object increases its value, to explain why effortful interaction could make learning more rewarding. Concretely, it proposes staging cognitive effort across a task: use GenAI for low-effort starts, plans, and explanations, then require students to invest more effort when evaluating, critiquing, iterating

Load-bearing premise

The load-bearing premise—which the paper states explicitly as "Extrapolating from these findings into the domain of learning suggests that it can be beneficial"—is that extra effort imposed by AI tool design converts into learning and achievement rather than frustration, a link drawn from small decision-making studies and a survey rather than from direct learning-outcome experiments.

Editorial extensions

If this is right

Students can learn to use GenAI with less effort at the start of a task and more effort later—checking, critiquing, and iterating—without losing the time-saving benefits.
Combining GenAI use with traditional methods such as handwritten note-taking can make complex material accessible while preserving the memory benefits of effortful note-making.
Educators can assess the process as well as the product by asking students to document their prompts, disagreements, revisions, and acceptance decisions.
AI assistants that ask users to articulate their own rationale before giving feedback can make decisions more reflective, at the cost of being burdensome—a cost students may nonetheless value.
Dialogic, persona-based uses of GenAI can provoke critical questioning and follow-up research in classrooms, rather than passive acceptance of output.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

One extension the author leaves implicit: the effort-staging design implies an inverted-U curve between imposed effort and learning, with too much required effort pushing students toward disengagement; experiments could identify the point at which effortful AI use stops being rewarding.
The argument suggests a product-design direction only gestured at in the paper: GenAI tools could offer an explicit "deep mode" that withholds recommendations until the user has written a rationale or made a prediction, turning effort into a feature rather than a bug.
If the effort paradox generalizes, it predicts that students who use AI to produce a draft will later value the final work more when they have revised it substantially themselves—an education-specific instance of the IKEA effect that could be measured with ownership and achievement surveys.
The paper's evidence comes from decision-making and short-term studies; an unstated corollary is that the metacognitive benefits depend on students noticing the value of the effort, so simply forcing effort without explanation or reflection may fail.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Here is my take. This is a position paper, not a research report, and it should be read as a design provocation. The fresh idea is to apply the effort paradox to GenAI learning: effort is both costly and valued, so we might design GenAI tools to deliberately shift students between low-effort and high-effort modes. The five design directions—tools for thought, slowing thinking down, externalizing half-baked ideas, constraining tasks, and building metacognition—are concrete enough to spark empirical work. The paper draws on a reasonable literature base, including Kreijkes et al., Sharples, Tang et al., and the author's own earlier tools (Proberbots, SelVReflect, ExtendAI).

The main soft spot is the evidence bridge. ExtendAI had 20 participants and measured investment decisions, not learning. SelVReflect measured reflection during a VR experience, not learning gains. Kreijkes et al. compared note-taking with LLM use; the effortful condition was traditional note-taking, not a GenAI tool that imposes effort. The knowledge-worker survey is self-report. None of these directly support the conclusion that imposed effort in GenAI interactions improves learning outcomes. To her credit, Rogers mostly hedges with 'suggests' and 'could', and even says the key question remains open. But the summary's 'I have shown' overstates the case; that sentence needs softening. A related gap is the lack of attention to failure modes: forced effort can frustrate students, push them to use other AI tools covertly, or disadvantage those with weaker prior knowledge. These deserve a paragraph in any revision.

The stress-test note is right about the thinness of the empirical anchors, but it slightly overreaches if it expects a position paper to supply outcome data. The value here is framing and agenda-setting. As such, it is a clear, readable piece that would be useful for people designing GenAI learning tools. I would send it to peer review—not to verify any new result, but to sharpen the claims and get the limitations and failure modes on the page.

Referee Report

4 major / 5 minor

Summary. This position paper argues that students' over-reliance on generative AI (GenAI) for homework may undermine critical thinking and writing skills. The author proposes the 'effort paradox' and the IKEA effect as conceptual lenses for understanding why requiring more effort when using GenAI might deepen learning and produce a sense of achievement. The paper sketches five design directions for 'tools for thought' that deliberately impose effort, describes the author's prior empirical work with VoiceViz, SelVReflect, and ExtendAI vs. RecommendAI, and draws on literature by Sharples, Tang et al., Kreijkes et al., and Lui et al. The central claim is that shifting cognitive effort to later stages of a task (e.g., critiquing AI output) or combining GenAI with traditional note-taking can foster critical thinking and metacognition. The paper concludes that the effort paradox provides a useful framework for rethinking learning with GenAI, while acknowledging that more research is needed.

Significance. The paper addresses an important and timely question: how to reconcile the convenience of GenAI with the need to develop students' critical thinking and writing skills. Its strength is in framing the problem through the effort paradox and offering concrete design ideas (e.g., requiring rationale before AI feedback, constraining interactions, combining GenAI with note-taking). The author's own studies, though small, provide proof-of-concept demonstrations that tool design can shift users' reflective behavior. If the central claim were supported by educational outcome data, the framework would meaningfully contribute to the HCI and AI-in-education literature. However, as presented, the claim rests on extrapolation from non-educational decision-making studies and self-report surveys; no study in the manuscript directly tests whether tool-imposed cognitive effort improves learning or metacognitive skills in educational contexts. The paper is therefore best read as a design manifesto or hypothesis-generating essay rather than an evidence-backed finding.

major comments (4)

[Designing new GenAI tools (ExtendAI paragraph)] The central claim that 'the additional effort involved could result in them learning more' (Abstract) is not directly supported by the cited studies. The ExtendAI vs. RecommendAI study involved 20 participants making simulated investment decisions and measured decision outcomes and self-reported reflection, not knowledge retention, transfer, or metacognitive skill. The paper itself acknowledges the leap: 'Extrapolating from these findings into the domain of learning suggests that it can be beneficial.' This extrapolation is load-bearing, so the strength of language in the Abstract and Summary ('have shown') should be tempered or supplemented with educational outcome measures.
[Other opportunities for learning (Kreijkes et al. paragraph)] The Kreijkes et al. RCT is cited as supporting the idea of combining GenAI with note-taking, but that study compared traditional note-taking with LLM use, not a GenAI tool designed to impose effort. Its result actually favored note-taking for comprehension and memory, which says nothing about whether a deliberately effortful GenAI tool—rather than effortful traditional study strategies—yields learning gains. The distinction between effort imposed by tool design and effort from doing the task oneself is confounded, and the manuscript does not address this.
[Designing new GenAI tools (ExtendAI description)] The mechanism attributed to 'extra effort' is confounded with a known learning technique: writing one's own rationale before receiving AI feedback is a form of self-explanation/elaboration. Any benefit of ExtendAI could stem from the content of the reasoning produced, not from the perceived cost of effort. To support the effort-paradox mechanism, an effort-matched control (e.g., requiring the same amount of typing but without self-explanatory content) would be needed. Without this, the paper's central mechanism is not isolated.
[The effort paradox] The analogy to the IKEA effect and effort paradox relies on studies of tangible goods and successful completion (Norton et al. 2012; Inzlicht et al. 2018). Learning outcomes are intangible, delayed, and uncertain; there is a real risk that imposed effort leads to frustration or disengagement rather than a sense of achievement. The paper acknowledges the cost side only in passing ('which they found burdensome') and does not present any evidence on boundary conditions. This is a significant gap for the central claim and should be explicitly discussed as an open empirical question, not only as a design opportunity.

minor comments (5)

[Designing new GenAI tools (item ii)] Typo: 'sometime using AI' should be 'sometimes using AI'.
[Designing new GenAI tools (ExtendAI paragraph)] Typo: 'critiquing and checking the validity of the this' should be 'of this' or 'of the output'.
[Introduction] The phrase 'especially those whose English is not their second language' appears to be an error; it should likely be 'not their first language.'
[Designing new GenAI tools (VoiceViz paragraph)] The typography 'Vo i c e Vi z' and 'S e l V R e f l e c t' is distracting; use regular formatting for tool names.
[References] The Freeman (2025) reference is incomplete: 'Student Generative AI Survey 202' should include the issue number and full title; the HEPI policy note number is listed but not the URL's accessible title. Minor citation formatting issues also appear in the Kreijkes et al. reference, which should include the SSRN preprint status.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper's argument is an explicitly conditional extrapolation from independent empirical work, not a derivation that reduces to its own inputs.

full rationale

The paper is a position/opinion piece, not a formal derivation. Its central claim—that requiring more effort when using GenAI 'could result in them learning more'—is explicitly hedged as a possibility ('What if...', 'could', 'I begin to outline', 'The question remains'). The effort paradox and IKEA effect are taken from external literature (Inzlicht et al. 2018; Norton et al. 2012) and used as analogies, not as premises that already contain the learning conclusion. The author's own prior tools (VoiceViz, SelVReflect, ExtendAI/RecommendAI) are cited as empirical studies of reflection and effort in decision-making tasks; the paper then explicitly labels the transfer to learning as 'Extrapolating from these findings into the domain of learning suggests that it can be beneficial,' which is an acknowledged inference rather than a disguised equivalence. Supporting citations such as Kreijkes et al. (RCT on note-taking vs LLM use), Sharples (2023), Tang et al. (2024), and Lui et al. (2025) are external empirical or conceptual anchors. There is no equation, fitted parameter, uniqueness theorem, or definitional identity connecting the conclusion to the inputs. The self-citations are present but are not load-bearing in the circularity sense: they point to published empirical studies rather than to unverified claims that already assume the target result. At most, the evidence base for the extrapolation is thin and the studies are small-scale, but that is a correctness/validity concern, not circularity. Accordingly, the score is 0.

Assumptions & free parameters 0 free parameters · 3 assumptions · 0 invented entities

No free parameters are used, and no new physical or conceptual entities are introduced. The argument relies on the effort paradox, the IKEA effect, and the transferability of effortful interaction to learning, none of which is established for the GenAI context by data in this paper.

assumptions (3)

domain assumption Effort is both costly and valued (effort paradox), and the IKEA effect transfers to learning.
The entire proposal rests on this psychological foundation, cited from Inzlicht et al. and Norton et al. but not tested in the context of GenAI learning.
ad hoc to paper Cognitive effort induced by tool design yields learning and metacognitive gains rather than disengagement.
The paper's main design thesis; supported only by indirect evidence from small decision-making studies and self-reports.
domain assumption Students can learn to calibrate when to invest more or less effort in different stages of a learning task.
The proposed less-effort/more-effort strategy assumes metacognitive awareness that students may not yet have; stated in the Summary as something to teach.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Why it is worth making an effort with GenAI." pith.science (2026). https://pith.science/paper/M6XOGJ2J

@misc{pith2026250900852,
  author       = {Pith},
  title        = {Pith review of: Why it is worth making an effort with GenAI},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/M6XOGJ2J}},
  note         = {Machine review of arXiv:2509.00852}
}

read the original abstract

Students routinely use ChatGPT and the like now to help them with their homework, such as writing an essay. It takes less effort to complete and is easier to do than by hand. It can even produce as good if not better output than the student's own work. However, there is a growing concern that over-reliance on using GenAI in this way will stifle the development of learning writing and critical thinking skills. How might this trend be reversed? What if students were required to make more effort when using GenAI to do their homework? It might be more challenging, but the additional effort involved could result in them learning more and having a greater sense of achievement. This tension can be viewed as a form of effort paradox; where effort is both viewed as something to be avoided but at the same time is valued. Is it possible to let students learn sometimes with less and other times more effort? Students are already adept at the former but what about the latter? Could we design new kinds of AI tools that deliberately require more effort to use to deepen the learning experience? In this paper, I begin to outline what form these might take, for example, asking students to use a combination of GenAI tools with traditional learning approaches (e.g. note-taking while reading). I also discuss how else to design tools to think with that augments human cognition; where students learn more the skills of metacognition and reflection.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

8 extracted references · 6 canonical work pages

[1]

Shake It Off

Why it is Worth Making An Effort when Learning With GenAI1 Yvonne Rogers UCLIC, UCL, UK Abstract Students routinely use ChatGPT and the like now to help them with their homework, such as writing an essay. It takes less effort to complete and is easier to do than ‘by hand’. It can even produce as good if not better output than the student’s own work. Howev...

work page 2025
[3]

Josh Freeman (2025) Student Generative AI Survey

work page 2025
[7]

National Literacy Trust. https://nlt.cdn.ngo/media/documents/Generative_AI_and_literacy_in_2024_summary.pdf Enkelejda Kasneci, Kathrin Sessler, Stefan Küchemann, Maria Bannert, Daryna Dementieva, Frank Fischer, Urs Gasser, Georg Groh, Stephan Günnemann, Eyke Hüllermeier, Stephan Krusche, Gitta Kutyniok, Tilman Michaeli, Claudia Nerdel, Jürgen Pfeffer, Ole...

work page 2023
[19]

Lev Tankelevitch, Elena L

https://doi.org/10.1080/1554480X.2024.2379774. Lev Tankelevitch, Elena L. Glassman, Jessica He, Majeed Kazemitabaar, Aniket Kittur, Mina Lee, Srishti Palani, Advait Sarkar, Gonzalo Ramos, Yvonne Rogers, and Hari Subramonyam. (2025). Tools for Thought: Research and Design for Understanding, Protecting, and Augmenting Human Cognition with Generative AI. In ...

work page arXiv 2024
[202]

Olivola (2018) The Effort Paradox: Effort Is Both Costly and Valued

HEPI number Policy Note 61 Download https://www.hepi.ac.uk/2025/02/26/student-generative-ai-survey-2025/ Michael Inzlicht, Amitai Shenhav and Christopher Y . Olivola (2018) The Effort Paradox: Effort Is Both Costly and Valued. Trends in Cognitive Sciences, Vo l u m e 2 2 , I s s u e 4 , 3 3 7 –

work page 2025
[349]

Hofman, Abigail Sellen, Sean Rintel, Daniel G

Pia Kreijkes, Viktor Kewenig, Martina Kuvalja, Mina Lee, Sylvia Vitello, Jake M. Hofman, Abigail Sellen, Sean Rintel, Daniel G. Goldstein, David M. Rothschild, Lev Tankelevitch and Tim Oates (2025) Effects of LLM Use and Note-Taking On Reading Comprehension and Memory: A Randomised Experiment in Secondary Schools. Available at SSRN: https://ssrn.com/abstr...

arXiv 2025
[2023]

which they could try doing themselves first using various GenAI tools. On the other hand, there has been much debate about the potential downsides of students becoming overly reliant on GenAI especially for tasks that require critical thinking and problem-solving. There is concern that their widespread use could reduce opportunities to learn and practice ...

2025
[2024]

Do you need a hint for analysing X?

by developing new ‘supertools’ that can expand their minds, daring them to think differently, while at the same time helping them break through the barriers that often stall or prevent creative leaps. For example, the GenAI could be designed to augment what they do by suggesting, prompting, conjuring up, counter-arguing, nudging, probing and even acting a...

work page 2023

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.

[1] [1]

Shake It Off

Why it is Worth Making An Effort when Learning With GenAI1 Yvonne Rogers UCLIC, UCL, UK Abstract Students routinely use ChatGPT and the like now to help them with their homework, such as writing an essay. It takes less effort to complete and is easier to do than ‘by hand’. It can even produce as good if not better output than the student’s own work. Howev...

work page 2025

[2] [3]

Josh Freeman (2025) Student Generative AI Survey

work page 2025

[3] [7]

National Literacy Trust. https://nlt.cdn.ngo/media/documents/Generative_AI_and_literacy_in_2024_summary.pdf Enkelejda Kasneci, Kathrin Sessler, Stefan Küchemann, Maria Bannert, Daryna Dementieva, Frank Fischer, Urs Gasser, Georg Groh, Stephan Günnemann, Eyke Hüllermeier, Stephan Krusche, Gitta Kutyniok, Tilman Michaeli, Claudia Nerdel, Jürgen Pfeffer, Ole...

work page 2023

[4] [19]

Lev Tankelevitch, Elena L

https://doi.org/10.1080/1554480X.2024.2379774. Lev Tankelevitch, Elena L. Glassman, Jessica He, Majeed Kazemitabaar, Aniket Kittur, Mina Lee, Srishti Palani, Advait Sarkar, Gonzalo Ramos, Yvonne Rogers, and Hari Subramonyam. (2025). Tools for Thought: Research and Design for Understanding, Protecting, and Augmenting Human Cognition with Generative AI. In ...

work page arXiv 2024

[5] [202]

Olivola (2018) The Effort Paradox: Effort Is Both Costly and Valued

HEPI number Policy Note 61 Download https://www.hepi.ac.uk/2025/02/26/student-generative-ai-survey-2025/ Michael Inzlicht, Amitai Shenhav and Christopher Y . Olivola (2018) The Effort Paradox: Effort Is Both Costly and Valued. Trends in Cognitive Sciences, Vo l u m e 2 2 , I s s u e 4 , 3 3 7 –

work page 2025

[6] [349]

Hofman, Abigail Sellen, Sean Rintel, Daniel G

Pia Kreijkes, Viktor Kewenig, Martina Kuvalja, Mina Lee, Sylvia Vitello, Jake M. Hofman, Abigail Sellen, Sean Rintel, Daniel G. Goldstein, David M. Rothschild, Lev Tankelevitch and Tim Oates (2025) Effects of LLM Use and Note-Taking On Reading Comprehension and Memory: A Randomised Experiment in Secondary Schools. Available at SSRN: https://ssrn.com/abstr...

arXiv 2025

[7] [2023]

which they could try doing themselves first using various GenAI tools. On the other hand, there has been much debate about the potential downsides of students becoming overly reliant on GenAI especially for tasks that require critical thinking and problem-solving. There is concern that their widespread use could reduce opportunities to learn and practice ...

2025

[8] [2024]

Do you need a hint for analysing X?

by developing new ‘supertools’ that can expand their minds, daring them to think differently, while at the same time helping them break through the barriers that often stall or prevent creative leaps. For example, the GenAI could be designed to augment what they do by suggesting, prompting, conjuring up, counter-arguing, nudging, probing and even acting a...

work page 2023