Coach not crutch: Evidence that AI can improve writing skill despite reducing effort

Synthesis: Lira, Rogers, Goldstein, Ungar & Duckworth (2025) test the intuition that using AI inevitably hinders learning by sparing effort. Across pre-registered studies they find the opposite is possible: AI can reduce effort while improving the learning environment, so learners can "work less and learn more." A nationally representative Gallup survey of young adults (N = 2,472) confirmed that the public largely holds the "crutch" intuition (73% negative about AI's impact on capability). But in experiments, participants who practiced writing cover letters with an AI tool wrote higher-quality no-AI cover letters than those who practiced alone — and AI beat both googling examples and receiving personalized feedback from experienced human editors. A third experiment showed AI teaches by example: merely viewing an AI-revised letter (no further practice) produced the same gain as practicing with the tool.

Key Findings

• "Work less, learn more" is empirically possible. Across pre-registered studies, AI access reduced practice effort but improved (or did not harm) learning outcomes — effort and learning rate can move in opposite directions.
• Public belief in the crutch intuition. In the Gallup survey (N = 2,472, ages 18–28), 73% scored above the midpoint on negative AI-impact attitudes: 79% agreed AI makes people lazier, 63% that it makes people less smart, 62% that it reduces ability to learn.
• Practice with AI beat practice without AI. Random-assignment experiment: those who practiced cover letters with an AI tool wrote better no-AI cover letters than those who practiced on their own.
• AI beat Google Search and human editors. Practicing with AI improved writing more than googling examples/tips (test d = 0.46) and more than personalized feedback from experienced human editors (test d = 0.20).
• No illusion-of-mastery effect. AI-practice participants did not over-report learning or skill vs. editor-feedback participants, mitigating concern that gains reflect inflated self-assessment.
• AI teaches by example. Participants who only viewed an AI-revised cover letter improved as much as those who practiced with the AI tool — the Scaffolding/example function, not just practice, drives the gain.
• Two-channel framework. AI simultaneously (a) reduces effort (crutch risk) and (b) improves the learning environment (coach benefit, e.g., worked examples). Skill outcomes hinge on the net effect.

What this means for practice

• Learners. Practice the target skill with an AI tool rather than avoiding it on principle: participants who practiced cover letters with AI wrote better unaided letters than those who practiced alone (d = .38) or not at all (d = .46), and the advantage persisted to the one-day follow-up (d = .29).
• Learners. Use the tool to obtain a worked example of the principle, not only a finished product: merely viewing one AI-revised letter, with no further practice, improved unaided writing as much as practicing with the tool did (d = .02 between those two conditions, p = .830).
• Learners. Attempt the task on your own before consulting the tool — the paper notes that trying first has been shown to improve learning more than using AI first — and treat the gain as a single-exposure effect rather than license for habitual use.
• Learners. Prefer AI over a search engine when you want guidance: practicing with AI beat googling tips and examples (d = .46) and personalized feedback from professional editors (d = .20) on the unaided test, and the AI-generated examples were themselves rated higher in quality (d = 1.66).
• Learners. Judge the tool by what it shows you rather than by how hard the session felt: AI users spent less time and fewer keystrokes yet reported similar subjective effort, and did not over-report how much they had learned.

Limitations

• All experiments ran on online Prolific samples of working adults rather than students in a course — Study 2 N = 2,238, Study 3 N = 2,997, and Study 4 N = 2,003, mostly college-educated, ages 18 to 95 — and the nationally representative survey (N = 2,472) covered only young adults aged 18–28.
• The skill measured is one short, structured genre — a cover letter, scored by raters plus a hypothetical interview judgment. The authors themselves question whether gains from observing an AI example would transfer to math or programming, where the final answer often does not reveal the process that produced it.
• Retention was probed only one day out, by recontacting a subsample of 800 participants of whom 633 responded (17–24% attrition, not differing by condition), and participants interacted with AI only once in the paradigm, so repeated use is untested.
• Effort outcomes rest on behavioral proxies (time on task and keystrokes), and the keystrokes-per-minute metric was added after preregistration; the human-editor comparison used 49 editors who, on the authors' own check, got faster without getting worse as they worked, so the AI advantage is not an artifact of editor fatigue.

Citation

Lira, B., Rogers, T., Goldstein, D. G., Ungar, L., & Duckworth, A. L. (2025). Coach not crutch: Evidence that AI can improve writing skill despite reducing effort.