The Doing Is the Thinking
Generative AI can produce a polished answer before a learner has developed the ability to produce, evaluate, or defend it. This investigation asks when AI assistance extends human agency and when it transfers the appearance of capability from the system to the user.
| Level | Estimated time | Textbook connection |
| Intermediate | 75 to 100 minutes | Chapter 2 pp. 26 to 39 |
Critical Moral Reasoning skills used facts and judgments, stakeholders, moral principles, duties, conflict resolution, independent transfer
Connection to the textbook This public module demonstrates one application of the method developed in Critical Moral Reasoning. The textbook provides the fuller explanations of the concepts and ethical frameworks used here.
The learner must make the first cognitive move.
Inquiry Focus
When is it ethically permissible for a learner to delegate cognitive work to generative AI
Learning Objectives
- Distinguish assisted output from independent capability and accurate self-knowledge.
- Classify claims about AI and learning as facts, judgments, or preferences.
- Identify stakeholders, competing principles, duties, and risks in an AI-use decision.
- Apply the Attempt Assist Audit Transfer protocol to a learning task.
- Defend a capability-based boundary and test it on a new case without AI assistance.
Textbook Connection
Chapter 2 distinguishes facts, judgments, and preferences and explains why a moral conclusion cannot be derived by jumping directly from what is to what ought to be. The Critical Moral Reasoning process moves from relevant facts through parties of interest, moral principles, duties, conflict resolution, and conclusion (Holcombe, 2025, pp. 26 to 39).
Inquiry Case The Polished Argument
Maya is a tenth-grade student. Her assignment is to write a 900-word argument about whether schools should restrict student phone use during class. The teacher permits generative AI if students disclose how they used it. The rubric grades the thesis, evidence, counterargument, organization, and mechanics.
Maya asks an AI system to summarize two assigned sources, propose a thesis, select quotations, outline the argument, draft the essay, and improve the wording. She reads the result, changes several phrases, checks that the quotations appear in the sources, and submits the essay with an AI-use statement. The essay earns a high score.
During a conference, Maya can state her conclusion but cannot explain why one source is more credible, identify the warrant connecting a quotation to her thesis, or answer a new counterexample without returning to the AI system.
Initial Position
Is Maya’s submission ethically acceptable under the teacher’s rule? Select one position and identify the strongest reason and one fact that could change it.
- Yes, disclosure and source checking make the use acceptable.
- No, the AI performed too much of the assessed reasoning.
- Only if Maya completes an independent oral or written check.
- Undecided, because additional facts are needed.
____________________________________________________________________________________________
____________________________________________________________________________________________
____________________________________________________________________________________________
Evidence Brief
The research record does not support a universal claim that AI either harms or improves learning. It supports conditional conclusions that depend on the task, learner, design, and measure of learning.
| Evidence | Finding | Reasoning relevance | Limit |
| Programming experiment | AI use improved some task performance but reduced later conceptual understanding and debugging skill. | Productivity and skill formation can diverge. | Short experiment with an unfamiliar programming library and a preprint report. |
| Reasoning experiments | AI improved objective performance while participants overestimated how well they had performed. | Output quality and calibrated self-knowledge can diverge. | Specific tasks do not establish permanent cognitive decline. |
| Learning meta-analysis | Across 35 experiments structured AI use had a positive overall effect with substantial variation across studies. | Instructional design can preserve learning while using AI. | Effects varied by discipline duration and use pattern. |
| Structured mathematics tutoring | A guided system emphasized scaffolding questions and was associated with improved mathematics performance. | AI can support learning when it preserves learner work. | Program context vendor involvement and uneven subgroup effects limit generalization. |
- Which finding most strongly challenges your initial position?
- Which limitation blocks a blanket rule about all AI use?
- What additional fact about Maya’s assignment would matter most?
Output Capability and Calibration
| Concept | Meaning | Maya case |
| Assisted output | The quality of an artifact produced with assistance | The polished essay |
| Independent capability | What the person can perform explain and adapt without the original assistance | Maya’s ability to construct and defend an argument |
| Calibration | How accurately the person judges the limits of that capability | Maya’s awareness of what she cannot yet explain |
False cognitive power transfer occurs when observers attribute the quality of an AI-assisted artifact to the user’s independent ability even though the user cannot reproduce, explain, evaluate, or transfer the relevant operation. A grade can describe the essay accurately while describing the learner’s capability inaccurately (Fernandes et al., 2025; Shen & Tamkin, 2026).
Exercise 1 Facts Judgments and Preferences
| Claim | Fact judgment or preference | Reason |
| Maya used AI to draft the essay | ||
| Maya’s essay earned a high score | ||
| The teacher should prohibit all generative AI | ||
| I dislike writing without AI | ||
| Disclosed AI use is always ethically acceptable | ||
| Maya could not explain the warrant during the conference | ||
| An independent transfer task is a better measure of capability than the assisted essay alone |
Exercise 2 Stakeholders and Principles
| Stakeholder | Interests | Possible benefit | Possible harm |
| Maya | Learning autonomy fair evaluation support | Feedback and improved expression | Dependency inflated self-assessment or mismeasurement |
| Classmates | Comparable standards and access | Shared assistance | Unequal advantage or uneven skill development |
| Teacher | Valid assessment and workable instruction | Faster feedback | Unreliable evidence of mastery |
| Future educators or employers | Accurate evidence of capability | AI-fluent learner | Decisions based on false signals |
| School | Equity privacy and learning outcomes | Scalable support | Widened gaps data risk or weak accountability |
Analyze the conflict among autonomy, fairness, beneficence, nonmaleficence, and integrity. Which principle has priority when learner choice conflicts with valid assessment? Would the answer change if AI supplied language support for a documented disability without generating the argument? Identify the precise cognitive operation that changes.
Attempt Assist Audit Transfer
Use the following four-stage protocol to distinguish assistance from substitution. The protected operation changes by subject, but the sequence remains stable.
| Stage | Learner responsibility | Permissible AI role | Evidence retained |
| Attempt | Make a prediction draft sketch source analysis or attempted solution before AI use. | None except approved access support that does not perform the protected operation. | First attempt and confidence estimate |
| Assist | Request a hint critique counterexample alternative or explanation. | Respond to the learner’s work without producing the final assessed artifact. | Prompt and response |
| Audit | Verify claims sources logic omissions and fit. Accept reject or revise with reasons. | Expose weaknesses and answer questions without deciding for the learner. | Decision trace |
| Transfer | Perform a novel comparable task without the original AI output. | No AI during the transfer check. | Independent explanation or performance |
Exercise 3 Apply the Protocol
- Name the protected cognitive operation in Maya’s assignment.
- Redesign Maya’s AI use so that she makes the first cognitive move.
- Specify what her decision trace must show.
- Design a five-minute unaided transfer check.
- Classify the redesigned AI use as required permissible or prohibited and justify the classification.
____________________________________________________________________________________________
____________________________________________________________________________________________
____________________________________________________________________________________________
____________________________________________________________________________________________
Capability Based Permission
| Learner stage | AI boundary | Evidence needed before broader delegation |
| Novice | Restrict substitution and protect the first attempt. | Explain core concepts and complete a basic operation. |
| Apprentice | Permit coaching hints and comparison after an attempt. | Diagnose errors and justify revisions. |
| Independent | Permit partnership with verification and disclosure. | Transfer the capability to a novel task. |
| Expert | Permit selective delegation with continuing accountability. | Audit exceptions limitations and consequences. |
Reasoned Conclusion
Write 500 to 700 words. Answer the central ethical question by defining the protected operation, identifying the learner’s stage, using at least one empirical finding and its limitation, analyzing stakeholders and competing principles, and designing evidence of independent transfer. Address one accessibility or expert-user counterexample.
Reflection and Unaided Transfer
Return to your initial judgment about Maya. State what changed, what remains uncertain, and your revised one-sentence rule for ethically acceptable AI assistance.
Transfer Case
A first-year accounting employee uses generative AI to reconcile a client’s records. The AI flags discrepancies, proposes adjusting entries, and drafts an explanation for a supervisor. The employee checks that totals balance but cannot explain why two entries were classified as liabilities rather than expenses. The firm has no rule against the use, and the supervisor values speed.
Without using AI, apply the Critical Moral Reasoning sequence and design an Attempt Assist Audit Transfer protocol for the employee’s next reconciliation.
References
Alnemrat, A. A., Aldamen, H. A., Almashour, M. A., Al-Deaibes, M. A., & AlSharefeen, R. A. (2025). Comparing AI and teacher feedback in EFL argumentative writing. Frontiers in Education, 10, 1614673. https://doi.org/10.3389/feduc.2025.1614673
Fernandes, D., Villa, S., Nicholls, S., Haavisto, O., Buschek, D., Schmidt, A., Kosch, T., Shen, C., & Welsch, R. (2025). AI makes you smarter, but none the wiser. Proceedings of CHI 2025. https://arxiv.org/abs/2409.16708
Holcombe, M. T. (2025). Critical moral reasoning: An applied empirical ethics approach.
Holcombe, M. T. (2026). The doing is the thinking: Why let AI do the doing fails novices [Presentation].
LearnLM Team, & Fab AI. (2026). Teaching with Gemini: Measuring the impact of Guided Learning on student mathematics progress in Sierra Leone. Google. https://storage.googleapis.com/deepmind-media/LearnLM/learnLM_sierraleone_may26.pdf
Shen, J. H., & Tamkin, A. (2026). How AI impacts skill formation [Preprint]. arXiv. https://arxiv.org/abs/2601.20245
Wu, X., Zhu, P., Zhang, J., Yin, M., & Wang, Y. (2026). ChatGPT’s impact on student learning outcomes: A meta-analysis of 35 experimental studies. Humanities and Social Sciences Communications, 13, 684. https://doi.org/10.1057/s41599-026-07019-z