From the OpenAI edition of October 29, 2026
RCT of 1,053 Bocconi undergraduates: ChatGPT Edu raised rubric scores; causal-reasoning training broadened ideas but did not
Bocconi University; OpenAI Economic Research · Higher education · Milan, Italy
- Publisher
- OpenAI (working paper PDF)
- Original publication
- August 2026
- Source retrieved
- 2026-10-30
What happened
In a 2x2 randomized trial of 1,053 first-year students, ChatGPT Edu (GPT-4o) access raised expert rubric scores on a business case, while causal-reasoning training increased mechanism identification (about 0.5 SD) and idea diversity without raising rubric scores. The combined condition showed more coherent and falsification logic.
Why it matters
Gives universities experimental evidence that AI access improves graded output on well-defined tasks but that critical-thinking instruction adds distinct benefits the rubric did not capture, informing how ChatGPT Edu rollouts pair with curriculum.
Evidence and measured results
RCT across 13 classes with expert graders, inter-rater reliability reported in the appendix, and OLS with clustered errors; some covariates failed balance checks and were controlled.
Limitations and uncertainty
Co-authored with OpenAI employees and contractors; single Italian university and task; GPT-4o rather than current models; publication month only (August 2026); not shown as peer-reviewed.
Put this evidence to work
Lighthouse Advisory interpretation, grounded in this source. Enriched 2026-10-30; this does not change the original publication date. Each role follows the same checklist so takeaways compare across sources.
Sales
Role takeaway
- Customer problem
- Universities rolling out AI need evidence on whether access helps or harms student work quality and thinking.
- Stakeholders
- Provost
- Teaching and learning center
- Faculty
- Institutional research
- Discovery questions
- Which courses would pilot AI access with assessment changes?
- Do your rubrics reward originality as well as quality?
- Can you run a controlled comparison across sections?
- Value hypothesis
- AI access paired with reasoning instruction may improve work quality while preserving idea diversity.
- Potential engagement
- A bounded course-level pilot design with blinded grading and originality measures.
- Unsupported claims
- Long-term learning gains
- Generalization beyond well-defined tasks
- Independence from vendor influence
Pre-sales engineering
Role takeaway
- Fit
- Fits institutions with ChatGPT Edu or equivalent access able to run course pilots.
- Architecture and integration
- Standard ChatGPT Edu workspace access; no integration.
- Prerequisites
- Provisioned Edu workspace
- IRB or research approval for pilots
- Grading rubric design
- Constraints
- Model version differences
- Single-task evidence
- Vendor co-authorship
- Security
- Pilot data handling should follow institutional research policies.
- Proof of value
- Section-level randomized pilot
- Blinded expert grading
- Measure idea diversity alongside rubric scores
Delivery
Role takeaway
- Work
- Design and run course pilots that pair AI access with reasoning training and revised rubrics.
- Dependencies
- Faculty participation
- Research approval
- Grading capacity
- Ownership
- Teaching and learning center with institutional research.
- Skills and adoption
- Faculty development on assessment design and AI-aware pedagogy.
- Governance checkpoints
- Research approval
- Mid-term review
- Results review before scaling
- Acceptance criteria
- Pilot completed with blinded grading (proposed)
- Originality measure reported (proposed)
- Risks
- Homogenized student work
- Overreliance on AI
- Misreading vendor-coauthored results
Implementation considerations
Lighthouse Advisory interpretation across the operating dimensions a public-sector buyer must settle before this evidence becomes a design. Each note answers the question under its heading for this specific source.
Architecture and integration
What must connect, and where does the AI sit in the workflow?
Limited: used standard ChatGPT Edu access; no integration.
Governance
Who approves, reviews and stays accountable for outcomes?
Assessment design matters: rubrics rewarding closeness to expert answers may favor AI-assisted work and penalize original reasoning.
Security and privacy
What data, permissions and controls need testing?
Limited relevance; study data handling not a focus.
Accessibility and workforce
Who is affected, and what skills or accommodations follow?
Targets novice workers and students; suggests training in reasoning remains valuable alongside AI access.
Procurement
What should contracts, pricing and exit terms secure?
Limited relevance to procurement.
Operating model
Which teams own the service once it runs?
Supports pairing tool access with curriculum and assessment redesign.
Publication history
- 2026-10-29OpenAI · Issue 016 resources
Stable resource ID: bocconi-openai-rct-chatgpt-edu-causal-reasoning-202608