Fostering Critical Thinking in the Age of AI

Synthesis: Aguilar, Nye, Swartout and colleagues (2025) report three studies from the USC Center for Generative AI and Society on how students and teachers are adapting to generative AI. A national survey of over 1,000 U.S. college students separates instrumental help-seeking (using AI to understand a topic) from executive help-seeking (using AI to get quick answers). A pilot of ABE, an LLM-based writing coach built from scaffolded reflection activities, found high technology acceptance and reported gains in metacognitive reflection. A survey of 1,505 in-service teachers in five countries finds awareness outpacing confident classroom adoption.

Key Findings

Executive help dominates. Students sought the most executive help from the internet and GenAI and the least from instructors and tutors, preferring machines to people.
Self-efficacy and Trust pull opposite ways. Higher internet help-seeking self-efficacy went with less GenAI help-seeking of both kinds, while higher trust in GenAI content went with more shortcut-seeking.
Professor encouragement works. Greater encouragement from professors was associated with more instrumental GenAI use, as was avoiding asking peers for help.
A scaffolded writing coach was accepted. ABE's ratings for Ease of Use, Learning Expectancy, Attitude, and Curiosity ranged from 4.69 to 4.92 out of 6, frustration was relatively low (Mean = 3.03, SD = 1.51), and Counter Arguments was most used (89%).
Teachers' awareness outpaces use. Of 1,505 teachers in five countries, 42.1% called themselves "very" or "extremely" familiar with GenAI but only 51.5% were comparably confident using it.
Benefits and risks coexist. 72% said GenAI streamlined routine tasks and 73.1% saw better learning outcomes, while 68.9% saw more plagiarism and 62.4% less student creativity.

Instrumental help, executive help, and what shapes it

Macias and Aguilar surveyed over 1,000 U.S. college students in 2025 about how they seek help from GenAI alongside instructors, tutors, peers, and internet search, separating instrumental help-seeking (using AI to understand) from executive help-seeking (getting a direct answer with minimal effort).

Students sought instrumental help from every source at roughly the same rates, most from the internet and least from tutors. Executive help was different: most from the internet and GenAI, least from instructors and tutors. Four predictors stood out: existing search habits carried over to GenAI; higher internet help-seeking self-efficacy went with less GenAI help-seeking of both kinds; avoiding peers went with more instrumental GenAI use; and professor encouragement went with more instrumental use, evidence that instructors shape how students use GenAI. Lower perceived competence and higher trust in GenAI content both predicted more shortcut-seeking.

An AI coach for argumentative writing

Xing and Aguilar ask what happens when GenAI removes the struggle from writing. Their tool, ABE (AI for Brainstorming and Editing), does not produce essays; it walks students through coach-guided activities inside a live-editable document, including Counter Arguments, Thesis Support, and Vagueness Detection.

A mixed-methods pilot with undergraduates in writing-intensive courses used a technology-acceptance survey. Ratings for Ease of Use, Learning Expectancy, Attitude, and Curiosity ranged from 4.69 to 4.92 out of 6, with low frustration (Mean = 3.03, SD = 1.51). Counter Arguments was most used (89%), then Thesis Support (40%) and Stronger Hook (21%). Among 58 students answering an open-ended question, 43% mentioned Counter Arguments and 40% Perspective Expansion. Students called ABE a learning companion offering personalized Feedback, not a shortcut, though the benefits may not carry over.

Teachers around the world: awareness without confidence

Xiu and Aguilar surveyed 1,505 in-service teachers in the United States, India, Qatar, Colombia, and the Philippines. The sample averaged 13.9 years of teaching experience and 60.6% had received formal training in educational technology tools, excluding AI.

Awareness and confidence were moderately high but uneven across the five countries. 37.0% of teachers reported institutional resources for GenAI, with encouragement ratings reaching 4.5 in Qatar.

Use concentrates in preparation. Among teachers using AI, 53.0% created assignments with it, 14.5% personalized learning activities, and only 0.7% supported special-needs students; 44.3% never or rarely used it for lesson preparation and 53.5% never or rarely used it in live teaching. Most saw benefits (73.1% for improved learning outcomes, 66.4% wanting mandatory GenAI training), but 68.9% saw increases in plagiarism and 62.4% a decrease in creativity and originality. The authors recommend GenAI-specific training and attention to equity in access and use.

What this means for practice

Instructors. Name the help you want: students sought the least executive help from instructors and tutors and the most from the internet and GenAI, so a brief can steer behavior directly.
Instructors. Encourage thoughtfully. Higher professor encouragement predicted more instrumental GenAI use, making instructor messaging a modifiable factor in the help-seeking model.
Instructors. Scaffold the writing process rather than ignoring the tool: Counter Arguments was ABE's most used activity (89%), and students called it a learning companion, not a shortcut.
Administrators. Treat training as the binding constraint: 42.1% of teachers were familiar with GenAI but only 51.5% confident using it, and 66.4% wanted training to be mandatory.
Administrators. Answer the plagiarism and creativity concerns (68.9% and 62.4% agreement) with academic integrity policy, rather than treating adoption as settled.

Limitations

This is a research-center report, not a peer-reviewed journal article: three studies with no shared instrument or unified sample, published as an OSF preprint.
The help-seeking findings are correlational and built on self-reported likelihood ratings, so they cannot establish that professor encouragement causes instrumental use.
The ABE evidence is an acceptance and perception study, with no comparison condition and no outcome measure for critical thinking; the report itself asks whether the benefits transfer.
The teacher survey is a cross-sectional snapshot of 1,505 self-selected respondents in five countries, so country differences are confounded with sampling and context.

Connected Concepts

Connected Articles

Efficiency vs. Effectiveness: Self-Regulated Learning with LLM-Mediated Help-Seeking — instrumental versus executive help-seeking with LLMs
Warning About AI Fallibility Increases Help-Seeking in an Intelligent Tutoring System — help-seeking in an AI-supported tutor
Coach not crutch: Evidence that AI can improve writing skill despite reducing effort — coaching rather than shortcutting in AI-supported writing
AI Adoption Among Teachers: Insights on Concerns, Support, Confidence, and Attitudes — teacher confidence, concerns, and institutional support
Using AI-Generated Feedback to Improve Critical Thinking and Writing Proficiency — AI feedback for critical thinking and writing proficiency

Citation

Aguilar, S. J., Nye, B., Swartout, W. R., Macias, A., Xing, Y., & Xiu, R. (2025). Fostering Critical Thinking in the Age of AI. A report from the USC Center for Generative AI and Society, Summer 2025. Research-center report (not peer-reviewed).

Metadata

Study design	Survey, User Study
Discipline	Writing Education
Level	Higher Ed, K 12
Intended audience	Instructors, Administrators
Learning and instruction	Help-Seeking, Metacognition
Technologies and techniques	Generative AI, Large Language Models (LLMs)
Assessment and measurement	Feedback
Research methods and evaluation	Mixed-Methods Research, Usability Research
Institutions and policy	AI Governance
Equity, ethics, and responsible use	Trust
Foundations of AI in education	Critical Thinking, AI Literacy
Page kind	Synthesis

Source History Edit Chat