From the K–12 edition of September 18, 2026
CEPR/NBER evidence: GenAI raises homework scores but cuts closed-book exam learning, concentrated in homework outsourcing
CEPR Discussion Paper / NBER conference paper (Strömberg, Lei, Wu) · K-12 academic economics research (secondary education) · China (grades 7–12); transferability limits for U.S. districts
- Publisher
- The Generative AI Learning Penalty: Evidence from Chinese Secondary Education
- Original publication
- June 2, 2026 (CEPR DP21577)
- Source retrieved
- 2026-09-19
What happened
Using 30 months of panel data on 26,811 Chinese students in grades 7–12, staggered GenAI adoption raises homework scores ~18% and cuts completion time ~30%, but lowers monthly closed-book exam scores ~20% within six months; high-stakes entrance exams fall ~18–24% with the full penalty emerging after about two years. Losses concentrate among roughly 80% of AI users whose behavior matches homework outsourcing (very short completion time with high homework scores); users who keep similar time-on-task see small losses. Why now: U.S. districts debating bans on AI-completed assignments (e.g., APS draft; Montgomery PTA concerns about removing mental struggle) need measured evidence that assisted homework productivity is not retained learning.
Why it matters
Gives SLED buyers a quantitative caution against treating homework speed/score gains as proof of learning, and supports instructional designs that preserve effortful practice and off-tool assessment.
Evidence and measured results
Opened the NBER conference PDF (full paper) and the CEPR DP21577 abstract page. Verified abstract statistics (n=26,811; 18%/30%/20%; entrance-exam 18 and 24%; ~80% outsourcing concentration) and ethics approval note. CEPR hosted PDF URL returned an HTML paywall/login page, so the conference PDF is the verification object.
Limitations and uncertainty
China secondary context; DiD/staggered adoption not a U.S. RCT; college entrance exams are out of pure K-12 scope except as long-run learning signal; transferability to U.S. curricula and tool ecosystems is limited and must be stated.
Put this evidence to work
Lighthouse Advisory interpretation, grounded in this source. Enriched 2026-09-19; this does not change the original publication date. Each role follows the same checklist so takeaways compare across sources.
Sales
Role takeaway
- Customer problem
- Districts may misread AI homework score/time gains as learning success while unaided performance declines under outsourcing.
- Stakeholders
- Curriculum chiefs
- Principals
- Teachers
- Assessment leads
- Boards
- Discovery questions
- What unaided assessments will judge your AI homework pilot?
- Can the tool prevent full-answer outsourcing on graded work?
- How will you detect ultra-fast high-score homework patterns?
- Are consumer chatbots blocked for assignment completion?
- Value hypothesis
- Effort-preserving AI homework design with unaided endpoints reduces the risk of productivity-without-learning that this study documents.
- Potential engagement
- 5-week homework-AI redesign: analytics review, outsourcing heuristics, and an unaided assessment pilot protocol.
- Unsupported claims
- That U.S. districts will see the same 20% exam penalty
- Guaranteed learning gains from any AI tutor mode
- That banning AI alone fixes outsourcing on consumer devices
Pre-sales engineering
Role takeaway
- Fit
- Fit for homework/practice products that can enforce scaffolding and unaided checks; poor fit for answer-complete chatbots on graded work.
- Architecture and integration
- Homework AI behind LMS with process logs; delayed unaided quizzes; time-on-task signals; optional AI-off assignments.
- Prerequisites
- Baseline unaided assessments
- Teacher workflow for process review
- Filtering of consumer GenAI where policy requires
- Constraints
- Evidence from China grades 7–12
- Observational DiD, not U.S. RCT
- Students may bypass on personal devices
- Security
- Keep assignment assistance on approved enterprise tools under DPA; log enough process data for integrity review without over-collecting.
- Proof of value
- Pre/post unaided assessment in pilot sample
- Share of sessions flagged as outsourcing-like
- Homework time distribution vs non-AI peers
Delivery
Role takeaway
- Work
- Set unaided endpoints, configure anti-outsourcing modes, train teachers on process evidence, and report dual metrics (assisted vs unaided).
- Dependencies
- Assessment team
- LMS admin
- Academic integrity policy
- Ownership
- Chief academic officer owns learning endpoints; principals own classroom enforcement.
- Skills and adoption
- Teachers need PD to distinguish tutoring from outsourcing and to design AI-off checks.
- Governance checkpoints
- Pilot pre-registration of unaided outcomes
- Midpoint outsourcing-pattern review
- Go/no-go on expanding homework AI
- Acceptance criteria
- Primary outcome is unaided/closed-book measure
- Documented response when outsourcing patterns spike
- Consumer GenAI assignment path addressed in AUP
- Risks
- Over-generalizing China effect sizes
- Teachers equated homework scores with mastery
- Equity issues if only some students have stealth AI access
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?
If product analytics celebrate homework score/time gains, add off-tool or closed-book checks and time-on-task monitors; discourage answer-complete modes for graded homework.
Governance
Who approves, reviews and stays accountable for outcomes?
Academic integrity and AI-use policies should distinguish tutoring-with-effort from outsourcing; evaluate pilots on unaided assessments, not homework throughput alone.
Security and privacy
What data, permissions and controls need testing?
Indirect: outsourcing patterns may push students to consumer tools outside district DPAs—reinforce approved-tool boundaries when homework AI is allowed.
Accessibility and workforce
Who is affected, and what skills or accommodations follow?
Protecting productive struggle may conflict with legitimate assistive uses for some disabilities—pair integrity rules with IEP-aligned accommodations rather than blanket bans.
Procurement
What should contracts, pricing and exit terms secure?
Score vendors on unaided-learning evaluation plans and anti-outsourcing defaults, not homework throughput dashboards alone.
Operating model
Which teams own the service once it runs?
Shift AI homework pilots into the same operating cadence as assessment: pre-register endpoints, monitor process metrics, and stop when unaided learning falls.
Publication history
- 2026-09-18K–12 · Issue 134 resources
Stable resource ID: cepr-genai-learning-penalty-chinese-secondary-2026