Learn With Jesse
Menu

Group work and discussion

Explain It Yourself: AI-assisted group projects with individual accountability

A design for courses where students may use AI. AI helps students plan. Then every student must explain the material and solve problems of that type alone, in front of teammates and the instructor.

Emerging evidenceBuilt around AIIn personUnder 60 studentsA project to set upProposed design, not yet piloted

What it is

Group projects with AI can look finished while understanding stays thin. One student may run the tool while everyone else reads its output, and a grade on a group presentation does not show what each student can do alone.

Explain It Yourself keeps the group and the AI, and changes what the group is graded on. AI is allowed in planning, and the group must find and correct its errors. Then AI is switched off. Roles rotate, the instructor visits each group and asks individual members questions, and the final round puts randomly chosen members in front of the class. Because any member can be asked any question, the group's result depends on everyone's understanding.

This is a design, not a result. No study has tested this exact sequence, and nothing on this page reports outcomes from it.

What the design is for

The design starts from the result the course needs, following backward design1 and constructive alignment.2 Each element is tied to one of three things.

  • Outcome: each student can solve new problems of this type alone, without notes or AI.
  • Evidence: an individual problem solved during the session, individual answers in the final round, and a closed-book check one week later.
  • Activities: the session sequence below, built to produce that evidence.

Why this problem matters

Lecturing is still the most common format in STEM classrooms.3 Students who never explain material in their own words get less from a course, and a group project that lets students skip explaining can produce a good-looking product with a weak outcome. An AI tool makes a polished result easy to produce without the work that produces learning, and the studies of unrestricted AI use below point the same way.

Evidence and fit at a glance

Emerging evidence

Each part of the design has research behind it, from cooperative learning to guardrails on AI use. The combination has not been tested. It is labeled emerging, and the pilot plan below says how to test it.

How evidence is labeled

  • Set-upYou write the test items first, then a question pool, a rubric, and an AI-use log template. Planning takes several hours the first time.
  • Class timeOne or two class sessions per project, plus a closed-book check one week later.
  • DeliveryIn person
  • Class sizeUnder 60 students. Works best with under 25 students
  • AI roleBuilt around AI

When it works, and when it doesn’t

Best when

  • Your students may use AI for planning and you want each student's own understanding to be visible.
  • You teach in person, ideally with tables and whiteboards, and can circulate among groups.
  • Groups are small, three or four students, and the class is small or medium in size.
  • The project has a core idea or problem type that each student should be able to explain and solve alone, with questions you can write in advance.
  • You can set aside class time for the mid-work visits and a final round.

Weak when

  • The course is fully online and asynchronous. The design depends on live visits, whiteboards, and questioning.
  • The class is large and one instructor has to run every visit and final round without help.
  • The goal is a polished group product, not each student's own skill. This design deliberately spends time on individual checks.
  • Students find public questioning very stressful and there is no written or private alternative.
  • You have no time to write a question pool and rubric before the session. Questions invented on the spot are easy to game and hard to score fairly.

How a session runs

Session sequence. Each phase states what AI use is allowed and what evidence it produces.
PhaseWhat students doAI useEvidence produced
1. PlanningThe group assigns roles and plans the work.Allowed. The typist runs only the prompts the group agrees on.An AI use log: each prompt, the answer, and what the group kept or rejected, with reasons.
2. Error huntThe group finds and corrects at least one error in the AI output.Allowed for checking.The corrected output and a note on what was wrong.
3. Table rehearsalEach member explains a section to a teammate, who asks questions. Whiteboards show the reasoning.Not allowed.Whiteboard work and instructor observation.
4. Mid-work checksThe instructor visits each group and asks two individual members a question.Not allowed.Feedback with a specific next step. Ungraded.
5. Individual problemEach student solves a new problem of the same type.Not allowed.Individual work, scored with a shared rubric.
6. Presentation and final roundThe group presents. Then the instructor asks randomly chosen members questions in front of the class.Not allowed.Individual answers, scored.
7. Delayed checkA closed-book check on the core idea, one week later.Not allowed.Individual score, compared with earlier terms.

Roles

Groups set three roles at the start, in a minute or two. Every member holds every role across the session, so no one can say a part of the work belonged to someone else.

  • Leader. Runs the session and checks each member's understanding. When a member misses a question, the leader sends the group back to teach that member, rather than handing over the answer. The role rotates by section.
  • Typist. Enters only the prompts the group agreed on, so everyone has to go through this person to reach the AI. The person who proposed a prompt reads the answer aloud and explains it.
  • Board writer. Draws the group's visual explanation while narrating it. The group can ask the writer to redraw a section from memory.

Leaders may earn a small bonus. Tie it to the team's results on the individual questions, not to the role itself, so the leader benefits only if every member can answer.

Checks during the work

The instructor visits each group during the work and asks two individual members a question. These checks are ungraded, and groups are expected to be unprepared the first time, since they have not done this before. Their purpose is to reveal gaps while there is still time to close them. Each check ends with a specific next step, such as a section to redo or an explanation to give before the next visit.

A major review concluded that innovations that strengthen frequent feedback to students produce substantial learning gains.4 A later meta-analysis of K-12 studies found that the effect size often credited to formative assessment, about .70, is not supported, and put the weighted mean near .20.5 This design does not rely on a particular effect size.

The final round

After the presentation, the instructor asks randomly chosen members questions in front of the class. Each member answers in their own words. Questions mix explanation ("why does this step work?") with transfer ("what changes if the input is different?"). Every student should answer at least one question, and two or three is better when time allows.

The question pool and scoring rubric are shared before the session, so students know what ready looks like. The pool should be large enough that memorizing answers does not work.

Public questioning can be stressful. Offer a written alternative for any student who needs one, so the design measures understanding and not comfort with speaking in front of a class.

Why each part is here

Design elements and the research behind them
ElementWhySource
Rotated questions and rolesWhen any member can be asked any question, the group's result depends on everyone's understanding. Cooperative learning research describes this as individual accountability combined with positive interdependence.6
Students explain in their own wordsIn experiments on the generation effect, material that people produced themselves was remembered better than the same material when they only read it. The original experiments used words, not explanations, so applying this to explanations is an extension of the finding, not something those experiments tested.7
Teach-back and teaching othersA study of students preparing to teach a computer agent, titled for the protégé effect, concerns effort: students put more effort into learning when they prepared to teach. It involved a computer agent, not a human peer, so treat it as supporting evidence only.8
Peer discussion of conceptsIn ten years of introductory physics courses, students gained in both conceptual understanding and problem solving after peer instruction was adopted.9
Round tables and shared whiteboardsSCALE-UP is a collaborative, hands-on classroom design for large introductory courses. Its authors report gains in conceptual understanding and problem solving in physics at their own institution, so these outcomes come from the project's developers.10
Mid-work checksFeedback during learning, when there is time to act on it, is associated with gains. The size of the gains is disputed.4,5
AI limited in rehearsal and checksIn a randomized field study with high school students, unrestricted access to a general AI tool improved practice scores but lowered later exam scores. A version with tutoring guardrails largely avoided that harm.11
AI allowed in planning, with a log and an error huntA randomized trial in a college physics course found that a purpose-built AI tutor produced more learning than an in-class active learning session, but the tool's design mattered and it was not a general chatbot. In studies of university students, ChatGPT use was linked to polished output without matching gains in transfer, and to lower-quality reasoning than web search. This design therefore allows AI in planning only, and makes students find and correct its errors.12–14
Delayed closed-book checkTaking memory tests, rather than rereading, strengthens long-term retention.15
Active learning as a wholeIn a meta-analysis of 225 studies in science, engineering, and mathematics, active learning raised exam scores by 0.47 standard deviations on average and lowered failure rates compared with traditional lecture.16
Backward designStarting from the outcome and the evidence, then choosing activities, produces tighter alignment between what is taught, practiced, and tested.1,2

How we would know it works

Before the term begins, decide what success means. Then:

  • Pilot the design in some sections of a course, and teach comparison sections the usual way, with the same test.
  • Measure the closed-book individual check one week later, along with the matching items on the final exam.
  • Track process evidence: AI use logs, how many mid-work checks found gaps, and whether the next check showed those gaps closed.
  • Compare course results, including the rate of students not passing, with several earlier terms. One term is noisy.
  • Survey students on confidence and fairness.

Adapting it for your course

  1. Write the test items first.
  2. Decide which parts of the work AI can help with, and which must be done alone.
  3. Build a question pool at the same level as the test.
  4. Keep the mid-work checks ungraded.
  5. Offer a written alternative for any student who needs one.
  6. Decide before the term what evidence will show whether the design worked.

With AI and without

Without AI

If your course does not allow AI, drop phases 1 and 2 and keep the rest. Rotating roles, mid-work checks, the final round, and the delayed check all work without AI, and the page on cooperative groups describes that version.

With AI

AI is allowed only in planning and error checking. The group keeps a log of every prompt, the answer, and what it kept or rejected. Review a few logs for each session to see how groups used the tool, and keep student data out of tools your college has not approved.

Pedagogical justification, ready to paste

Use this in a lesson plan, syllabus, or course-design form. Replace the [bracketed] parts with your own course details. Check the claims against the references before you submit.

One-sentence version

Group projects allow AI only for planning, followed by individual explanation and problem solving without AI, with rotating roles and randomly assigned questions, because unrestricted AI access improved practice scores but lowered later exam scores in a randomized field study,11 and cooperative learning depends on individual accountability combined with positive interdependence.6 This is a proposed design, not yet piloted.

Full rationale

To keep AI-assisted group work from replacing student understanding, group projects follow a design in which AI is allowed only for planning and error checking, and every student must then explain the material and solve a new problem of the same type without AI. Roles rotate, the instructor asks individual members questions during the work, and randomly chosen members answer questions in a final round. Because any member can be asked any question, the group's result depends on everyone's understanding, which is the logic of cooperative learning's individual accountability with positive interdependence.6 AI is limited because unrestricted access to a general AI tool improved practice scores but lowered later no-AI exam scores in a randomized field study of nearly 1,000 high school students, while a version with tutoring guardrails largely avoided that harm,11 and because ChatGPT use for open-ended tasks was linked to polished output without matching gains in transfer or reasoning.13,14 A closed-book check one week later uses retrieval, which strengthens long-term retention more than rereading.15 The design follows backward design: the outcome and evidence are written first, then the activities.1,2 The design is proposed and has not been piloted, so the course compares the closed-book check with earlier terms to judge whether it works.

Learning outcome
Students will be able to [verb + skill] on new problems of [type] alone, without notes or AI, and explain their reasoning in their own words.
Evidence of learning
An individual problem during the session, individual answers in the final round scored with a shared rubric, and a closed-book check one week later, compared with earlier terms.
Learning activity
A group project in which AI is used for planning with a log, followed by table rehearsal, mid-work checks, and a final round of randomly assigned individual questions, with roles that rotate.

The copy buttons paste plain text. The pasted version renumbers the references from 1 and includes them with a source line, so it works on its own. The numbers shown here match the reference list below.

Limits of the evidence

  • Most of the evidence supports active learning and cooperative learning as a whole. No study yet tests this exact sequence.
  • Group work can hide individual gaps. The rotation and the individual questions are designed to address that, and the design needs testing.
  • Less is known about the room itself than about the teaching methods used in it. The room supports the design and does not create the learning. SCALE-UP outcomes are reported by the project's own authors.10
  • The generation-effect and protégé-effect studies are about generating words and teaching a computer agent. This page applies them to student explanations as an extension, not as a tested result.7,8
  • The two randomized AI studies come from a high school math class and a college physics course, and they measured short-term results.11,12
  • The design adds class time and instructor time for the checks. The cost depends on class size.
  • Students differ in how they feel about group work and public questioning. Alternatives need to be built in from the start.

Questions faculty ask

How do I let students use AI in a group project without them skipping the learning?

Allow AI for planning and error checking, with a log, and then require each student to explain the material and solve a new problem without AI. In this proposed design, randomly chosen members answer questions in front of the class, so the group's result depends on everyone's understanding.

Has this design been tested?

No. It is a proposed design and has not been piloted. Its parts have research behind them, including cooperative learning, retrieval, and guardrails on AI use, but the sequence as a whole is untested. The pilot plan above describes how to test it.

What if a student is afraid of speaking in front of the class?

Offer a written alternative to any student who needs one, so the design measures understanding and not comfort with public speaking.

How do I stop the typist from doing all the thinking?

The typist enters only prompts the group agreed on, and the person who proposed a prompt reads the answer aloud and explains it. Every member also holds every role across the session.

References

  1. Wiggins G, McTighe J. Understanding by Design. Expanded 2nd ed. ASCD; 2005.
  2. Biggs J. Enhancing teaching through constructive alignment. High Educ. 1996;32(3):347-364. doi:10.1007/BF00138871
  3. Stains M, Harshman J, Barker MK, et al. Anatomy of STEM teaching in North American universities. Science. 2018;359(6383):1468-1470. doi:10.1126/science.aap8892
  4. Black P, Wiliam D. Assessment and classroom learning. Assess Educ Princ Policy Pract. 1998;5(1):7-74. doi:10.1080/0969595980050102
  5. Kingston N, Nash B. Formative assessment: a meta-analysis and a call for research. Educ Meas Issues Pract. 2011;30(4):28-37. doi:10.1111/j.1745-3992.2011.00220.x
  6. Johnson DW, Johnson RT. An educational psychology success story: social interdependence theory and cooperative learning. Educ Res. 2009;38(5):365-379. doi:10.3102/0013189X09339057
  7. Slamecka NJ, Graf P. The generation effect: delineation of a phenomenon. J Exp Psychol Hum Learn Mem. 1978;4(6):592-604. doi:10.1037/0278-7393.4.6.592
  8. Chase CC, Chin DB, Oppezzo MA, Schwartz DL. Teachable agents and the protégé effect: increasing the effort towards learning. J Sci Educ Technol. 2009;18(4):334-352. doi:10.1007/s10956-009-9180-4
  9. Crouch CH, Mazur E. Peer Instruction: ten years of experience and results. Am J Phys. 2001;69(9):970-977. doi:10.1119/1.1374249
  10. Beichner RJ, Saul JM, Abbott DS, et al. The Student-Centered Activities for Large Enrollment Undergraduate Programs (SCALE-UP) project. In: Redish EF, Cooney PJ, eds. Research-Based Reform of University Physics. Vol 1. American Association of Physics Teachers; 2007:1-42. doi:10.1119/RevPERv1.1.4
  11. Bastani H, Bastani O, Sungu A, Ge H, Kabakcı Ö, Mariman R. Generative AI without guardrails can harm learning: evidence from high school mathematics. Proc Natl Acad Sci U S A. 2025;122(26):e2422633122. doi:10.1073/pnas.2422633122
  12. Kestin G, Miller K, Klales A, Milbourne T, Ponti G. AI tutoring outperforms in-class active learning: an RCT introducing a novel research-based design in an authentic educational setting. Sci Rep. 2025;15(1):17458. doi:10.1038/s41598-025-97652-6
  13. Fan Y, Tang L, Le H, et al. Beware of metacognitive laziness: effects of generative artificial intelligence on learning motivation, processes, and performance. Br J Educ Technol. 2025;56(2):489-530. doi:10.1111/bjet.13544
  14. Stadler M, Bannert M, Sailer M. Cognitive ease at a cost: LLMs reduce mental effort but compromise depth in student scientific inquiry. Comput Human Behav. 2024;160:108386. doi:10.1016/j.chb.2024.108386
  15. Roediger HL III, Karpicke JD. Test-enhanced learning: taking memory tests improves long-term retention. Psychol Sci. 2006;17(3):249-255. doi:10.1111/j.1467-9280.2006.01693.x
  16. Freeman S, Eddy SL, McDonough M, et al. Active learning increases student performance in science, engineering, and mathematics. Proc Natl Acad Sci U S A. 2014;111(23):8410-8415. doi:10.1073/pnas.1319030111

About this page

Written by Jesse Rogers, Learning Lab Coordinator, Palm Beach State College. Last updated October 15, 2026.

AI disclosure. Model used for research, drafting, editing, and verification: claude-sonnet-5-5 (Claude, by Anthropic). The author is responsible for every claim and citation. The first draft of this page was written with claude-haiku-5-5. It was then rebuilt for this site, and its citations were re-checked and corrected, using claude-sonnet-5-5.

How it was checked. Bibliographic details were checked against DOI records. Numbers and findings were checked against abstracts or open full texts. Where that was not possible, the page says so. See the method page.

What I could not confirm.

  • Three placeholder citations in the first draft (the SCALE-UP chapter's pages, and the volume and article numbers for the Bastani and Kestin studies) were replaced with checked details. The SCALE-UP page range (1-42) comes from the OpenAlex record for that chapter.
  • The first draft cited a 1998 book by Johnson, Johnson, and Smith and a Phi Delta Kappan article by Black and Wiliam for the claims about cooperative learning and feedback. This page cites a 2009 review of cooperative learning research by Johnson and Johnson and the 1998 review of formative assessment by Black and Wiliam in Assessment in Education instead, because those are the sources whose details could be checked.
  • For the protégé-effect study (Chase 2009), the record I could reach had no abstract, so it is cited for effort toward learning on the strength of its title, and described as supporting evidence only.

License. CC BY 4.0. Share and adapt it with credit. This is an independent project and is not an official resource of Palm Beach State College.

Corrections. If you find an error, message me on LinkedIn and I will check it and fix it.

Cite this page (AMA). Rogers J. Explain It Yourself: AI-assisted group projects with individual accountability. Learn With Jesse. Published October 15, 2026. Updated October 15, 2026. Accessed [date]. https://learnwithjesse.com/techniques/explain-it-yourself/