Faster homework, poor exam results: What AI is doing to students’ learning

A large study of ~27,000 Chinese students found genAI raised homework scores by 18% while reducing time per assignment by ~30% (64 to 45 minutes). However, when the same students took monthly closed-book exams without AI, scores fell by 20% within six months, suggesting weakened learning when tools are removed. Survey evidence also indicates many students use AI for complete solutions (31%) rather than independent learning support, reinforcing a cautious view of genAI’s educational impact.

Analysis

The investable takeaway is not “AI hurts education” so much as “unstructured AI usage monetizes time, not learning.” That is a negative for consumer homework-help businesses because the value proposition gets commoditized by generic chatbots, while the more durable winners are workflow products that sit inside teacher/admin systems and can enforce guardrails, provenance, and assessment. In other words, the market should separate AI-assisted productivity from credential risk: anything judged on final scores rather than time saved is exposed to pushback.

Over the next 1-3 months, the likely catalyst is policy tightening rather than a demand shock. Schools and ministries will respond to falling exam outcomes by restricting open-ended use and pushing vetted, institution-level tools; that favors incumbents with distribution into districts and LMS integrations, not standalone chat interfaces. Over 6-18 months, the bigger effect is budget reallocation toward proctoring, adaptive testing, and “AI literacy” curriculum content, while generic tutoring and answer-generation features become easier to substitute and harder to price.

The contrarian point is that the headline risk is probably overstated for broad tech: students are not abandoning AI, they are just using it in a way that boosts throughput. That means the revenue upside for consumer AI education tools may be real, but the retention risk rises if outcomes disappoint and schools clamp down. The cleanest short is still weak because many education names are already deeply discounted, so the better setup is to fade any rally tied to AI-fueled engagement until there is evidence that exam performance, not just homework completion, is improving.

Key Decisions for Investors

• Avoid chasing any long in CHGG on AI-usage growth; if sentiment lifts the stock on engagement data, fade into strength over the next 2-6 weeks because the core monetization is vulnerable to generic chatbot substitution.
• Watch DUOL as a relative winner versus homework-help names: if AI is pushing students toward self-directed micro-learning, DUOL’s subscription model is better aligned than answer-generation businesses; consider a long DUOL / short CHGG pair only if DUOL user retention remains stable in the next earnings print.
• Do not force a broad AI-edtech long basket; instead wait for evidence of school-level procurement for proctored/adaptive tools. The cleaner beneficiaries are enterprise workflow names with education exposure, not consumer tutoring apps.
• Set a policy-alert trigger on district/state guidance that restricts open AI use in schools; that would be a short-term tailwind for compliance/proctoring vendors and a near-term headwind for any consumer AI study app.
• If CHGG trades back to levels implying AI-driven user growth without exam-linked conversion improvement, use a 1-3 month put spread to express the thesis that time-saved is not the same as value-created.