Problem-based learning
Beyond the Right Answer
A model for integrating AI and cognitive bias into problem-based learning
Chris Edmonds
University of Alabama at Birmingham · cte@uab.edu
Act I · The problem
Your students are using AI
In the largest UK student survey, AI use is no longer a trend. It's simply how students study now.
95%
of students now use AI in their studies,
and 94% use it on assessed work
and 94% use it on assessed work
AI-generated text pasted straight into assessed work
Share of students · HEPI surveys 2024 → 2026
2024
3%
2025
8%
2026
12%
4× in two years, and self-reported numbers tend to run conservative.
Source: HEPI / Kortext Student Generative AI Survey 2026 (n=1,054 full-time UK undergraduates, surveyed Dec 2025), hepi.ac.uk/reports/student-generative-ai-survey-2026
The big idea
What we teach shapes what AI gives back
Anthropic studied ~400,000 real AI work sessions. The difference between slop and great work wasn't the tool. It was how deeply the person understood the problem.
🎓
Our students
+ AI
+ AI
IF WE ONLY TEACH THE AI TOOL
AI does the work for them
→
Domain knowledge never develops
→
No way to judge or fix the output
→
AI slop
Novices: 15% verified success · give up when stuck ~4× as often
IF WE TEACH CRITICAL THINKING + DOMAIN KNOWLEDGE
Students understand the problem first (PBL)
→
Judgment and domain knowledge grow
→
They direct and evaluate the AI
→
Great work
Experts: ~2× the success rate · direct 12-action AI chains vs 5
Critical thinking and domain knowledge aren't competing with AI. They're the multiplier.
Source: Anthropic, “How Claude Code is used in practice” (2026), ~400K sessions / ~235K users, anthropic.com/research/claude-code-expertise
Act I · The problem
How students actually use AI
Anthropic classified ~574,000 real student conversations. Four patterns emerged, and about half the time students asked for the answer rather than the thinking.
PROBLEM SOLVING
OUTPUT CREATION
DIRECT
≈47%
“Solve this for me”
Direct answers to homework-style problems, minimal engagement
“Write this for me”
Complete essays, summaries, and materials produced on request
COLLABORATIVE
≈53%
“Help me work through it”
Guided, back-and-forth problem solving
“Help me improve it”
Iterative feedback and refinement of the student's own work
All four patterns occur at similar rates (23–29% of conversations each); Anthropic doesn't publish exact per-cell splits.
~half
of conversations are Direct: seeking finished answers or content with minimal engagement.
That's the half where we can make the biggest difference.
Source: Anthropic Education Report, How University Students Use Claude (2025), anthropic.com/news/anthropic-education-report-how-university-students-use-claude
Act I · The problem
Your brain on ChatGPT: cognitive debt
MIT Media Lab put EEG caps on 54 students writing essays: with an LLM, with search, or with brain only. Connectivity scaled down with every layer of help.
Neural connectivity while writing (strongest → weakest)
Brain only
Search engine
ChatGPT
Schematic of the study's ordering: “brain connectivity systematically scaled down with the amount of external support.”
83%
of LLM users couldn't accurately quote from the essay they had finished writing minutes earlier
The debt compounds
LLM users reported the lowest ownership of their work, and when later asked to write without AI, their neural engagement stayed depressed.
Source: Kosmyna et al., “Your Brain on ChatGPT: Accumulation of Cognitive Debt…”, MIT Media Lab, 2025, arxiv.org/abs/2506.08872 (preprint; n=54, 18 in session 4)
Act I · The problem
I'm seeing it in my intro classes
Same students, same semester. Maybe this looks familiar.
▲
Participation
on non-proctored work
on non-proctored work
Everyone's suddenly doing the homework
▲
Scores
on non-proctored exams
on non-proctored exams
And the homework looks great
▼
Scores
on proctored exams
on proctored exams
But close the laptop, and the learning isn't there
The work looks better than ever. The learning isn't keeping up.
Act I · The problem
Good news: our students' AI habits aren't set yet
Share of student AI conversations vs. share of U.S. bachelor's degrees. STEM over-indexes heavily, while business students are still early in their AI journey.
Share of Claude conversations
Share of U.S. bachelor's degrees
Computer science: 38.6% of conversations from 5.4% of degrees. Business: 8.9% from 18.6%. We arrive early enough to help shape how our students use it.
Source: Anthropic Education Report (2025), ~574K anonymized student conversations mapped to NCES degree data
Act II · The stakes
Students are offloading exactly the thinking their profession is about to pay a premium for.
Act II · The stakes
Meanwhile, the profession is already moving
AI adoption in tax and accounting firms nearly doubled in a single year, and it's already reshaping what firms bill for.
Tax firms using AI-powered research weekly
2025
33%
2026
60%
Top uses: advisory projects (44%), tax planning (40%), compliance research (39%). Judgment work, not data entry.
84%
of tax professionals agree AI saves them time, reallocated to client response, quality, and work-life balance
69%
expect to move off hourly billing as AI commodifies routine knowledge; value shifts to advice
Source: Blue J & CPA.com survey of tax firms, June 2026, cpa.com/news/blue-j-and-cpacom-survey-finds-ai-adoption-among-tax-firms-has-nearly-doubled-one-year
Act II · The stakes
AI eats the routine. Judgment survives.
The tasks we drill hardest in intro accounting are the ones most exposed to automation.
TASKEXPOSURE TO AI REPLACEMENT
Most of us inherited courses built around the top three rows. The opportunity: shift weight toward the bottom two, where our graduates' careers will live.
Framework: Abbas et al. 2025; Wang et al. 2025: AI pushing the profession toward interpretation, judgment, and advisory work
Act III · A solution
Reframe accounting as a judgment profession
…and teach it that way.
HOW IT'S OFTEN TAUGHT
🧮
Number cruncher
Rules to follow
Mechanics to practice
Terms to memorize
WHAT THE WORK REALLY IS
🧭
Trusted advisor
Interpretation of messy situations
Judgment under uncertainty
Ethics clients can rely on
✅It's what we actually do
🛡️It's harder to automate
✨It's more interesting
Act IV · Live case
Let's try one. Together.
You're the students. I'll give you a problem. Solve it however you like.
Then we'll come back to the slides and unpack what happened.
The debrief
Why that just worked
Across 25 studies, the average PBL student out-scores ~86% of traditionally-taught peers on critical thinking
1
Understanding must come before terminologyTeach terms first and memorization is the student's only option. Teach the problem first and critical thinking is the only option.
2
A problem students understand is a problem they can think critically aboutThe emphasis moves from memorizing to reasoning on day one.
3
Group problem-solving makes cognitive bias visibleWhen a whole class works one problem, you can point at anchoring and overconfidence as they happen, and teach them.
4
Domain language layers on easily afterwardOnce students truly understand the problem, accounting terminology is labels for things they already know.
5
Deep understanding is what makes AI useful instead of dangerousYou can only evaluate AI output on problems you genuinely understand.
Source: Li, Mustakim & Muhamad, “A Meta-analysis of the Effectiveness of Problem-based Learning on Critical Thinking,” European Journal of Educational Research (2025), eu-jer.com
The debrief
Field notes: making PBL land
1
Choose student problems over business problemsA problem they own beats a problem they're assigned.
2
Get a student to explain the topic. Excitement is contagiousThe moment a peer explains it, the room leans in.
3
Let students identify the problem themselvesIf all they do is correctly name the problem, they've already achieved a lot.
4
Let them be wrong, loudlyAny solution is a good solution for learning. Mistakes are the raw material.
5
Show how cognitive bias shaped their answersThe reveal is the lesson: their own reasoning becomes the case study.
The debrief
The class session
One repeatable structure, droppable into any course, any topic.
🧩
1 · Open with the PBL case
A student-life problem, no jargon. Groups propose solutions cold.
→
🏷️
2 · Add the language
Introduce accounting terminology and the business version of the same problem.
→
✏️
3 · Work problems
In-class accounting problems, now anchored to a problem they understand.
Thank you
Reach out.
Want the case materials, the slides, or to swap notes on what's working in your classroom? I'd love to hear from you.
Chris Edmonds
University of Alabama at Birmingham
cte@uab.edu
www.edmondshub.com
Scan to connect
Appendix
Sources
HEPI / Kortext Student Generative AI Survey 2026. Higher Education Policy Institute. 95% of students use AI; 94% on assessed work; 12% submit AI-generated text (3% in 2024). hepi.ac.uk/reports/student-generative-ai-survey-2026
Anthropic Education Report (2025). "How University Students Use Claude." ~574K anonymized conversations; four interaction patterns; ~47% Direct; CS 38.6% of conversations vs 5.4% of degrees. anthropic.com/news/anthropic-education-report-how-university-students-use-claude
Kosmyna et al. (2025). "Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task." MIT Media Lab preprint. n=54; EEG connectivity scaled down with external support; 83% of LLM users failed to produce a correct quote from their own essay in Session 1. arxiv.org/abs/2506.08872
Blue J & CPA.com (2026). AI adoption among tax firms survey: weekly AI research use 33%→60% in one year; 84% report time savings; 69% expect billing model shifts. cpa.com
Li, Mustakim & Muhamad (2025). "A Meta-analysis of the Effectiveness of Problem-based Learning on Critical Thinking." European Journal of Educational Research: 25 studies, overall effect size g = 1.08 (large), meaning the average PBL student scores higher than roughly 86% of traditionally-taught students. eu-jer.com
Anthropic (2026). "How Claude Code is used in practice." ~400K sessions, ~235K users; domain experts ~2× verified success vs novices (28–33% vs 15%); novices abandon nearly 4× as often (19% vs 5–7%). anthropic.com/research/claude-code-expertise
Abbas et al. (2025); Wang et al. (2025). AI and the transformation of accounting toward interpretation, judgment, and advisory work.