Show Notes
Homework rose 18% while exams fell 20%. The Generative AI Learning Penalty Evidence from Chinese Secondary Education tracked 26,811 students. In this episode: A study of 26,811 students in China revealed an "AI learning penalty": an 18% rise in homework scores coincided with a 20% fall in closed-book exam performance. Students completing AI-assisted homework in under 50 minutes showed strong homework scores but extremely weak examination results, indicating a significant AI homework impact on AI student performance. High-attaining and younger students were more susceptible to the generative AI education learning penalty, with estimated Zhongkao and Gaokao results 5-7% lower overall for AI adopters. The research by David Strömberg, Victor Lei, and Yanhui Wu highlights that improved homework performance could actively conceal declining understanding and poor AI study habits. Effective generative AI education strategies must focus on tasks that reward critical thinking, like challenging AI responses, rather than simply fast completion, to avoid the learning penalty. Chapters: 00:00 — Cold open & welcome 00:20 — Introducing the AI learning penalty research 00:50 — Homework productivity vs. actual student learning 01:25 — Strength of the study and contextual limitations 02:00 — Impact of homework time on AI student performance 02:50 — Rethinking AI use: from output to student interaction 03:30 — Measurement challenges for teachers and leaders 04:10 — Differential impact on subjects, attainment levels, and age groups 05:00 — Long-term consequences on Zhongkao and Gaokao results 05:40 — Designing for productive AI study habits What is the AI learning penalty in education? The AI learning penalty, observed in a study of Chinese secondary students, refers to the phenomenon where students using AI for homework see improved assignment scores but experience a decline in their closed-book examination performance due to outsourcing intellectual effort. How does AI homework impact student performance on exams? AI homework can negatively impact student performance on exams if students use AI to complete tasks quickly without engaging in the intellectual work, leading to strong homework scores but weak results on assessments like Zhongkao and Gaokao where AI is not permitted. What are the long-term effects of generative AI on student learning? The long-term effects of generative AI on student learning, according to one study, include significantly lower scores on major national examinations like Zhongkao and Gaokao, with losses estimated at 5-7% overall for AI adopters and reaching 24% and 18% respectively for students who used AI consistently for two years. Featuring: Dan Fitzpatrick, The Generative AI Learning Penalty: Evidence from Chinese Secondary Education, David Strömberg, Victor Lei, Yanhui Wu, Zhongkao, Gaokao. Follow AI in Education with Dan Fitzpatrick for more on AI in education.
More from AI for Educators
Are screen time bans protecting children, or taking away tools some need to learn? Discover the accessibility blind spot in the screen time debate and what it means for parents navigating AI in education. Paddy McGrath, author of The AI Equity Paradox, joins Dan Fitzpatrick to explore how technology can help children communicate, build confidence and express ideas that might otherwise stay hidden. The conversation begins with a girl who couldn’t speak aloud. An iPad helped her deliver a speech about owls and win a prize. The technology hadn’t made her more intelligent. It had given her a way to show what she already knew. That story raises a difficult question: when we remove screens from classrooms, whose opportunities do we remove with them? As parents, we want our children to spend less time staring at screens. We also want them to grow up capable, creative and prepared for a world shaped by artificial intelligence. Paddy and Dan explore how to navigate those decisions, with practical examples from schools and their own families. IN THIS EPISODE ■ The screen time blind spot: Why blanket restrictions can overlook children who rely on assistive technology to learn and communicate. ■ “Necessary for some, useful for all”: How captions, dictation and other accessibility tools can benefit a much wider range of learners. ■ AI and homework: How to recognize when AI is supporting your child’s thinking and when it is doing the thinking for them. ■ The smartphone dilemma: Navigating friendships, online risks, parental controls and boundaries at home. ■ Children’s creativity: How AI can help bring a child’s original ideas to life, including Dan’s experience turning his son’s drawing into an animated creature. ■ The AI equity paradox: Why two schools with the same access to AI can offer their students very different opportunities. ■ Beyond test scores: Paddy’s argument for paying closer attention to confidence, independence and the barriers a tool helps remove. ■ When your school bans AI: How parents can model responsible use at home and share useful examples with teachers. If you’re a parent, teacher or school leader trying to make sense of screen time, AI literacy and inclusive education, this conversation offers questions you can take into your next discussion at home or school. What has technology helped your child do that they previously struggled with? Share your experience in the comments. Find us at https://theaieducator.io
OpenAI's always-on AI agents can now work around the clock. Here's what school leaders need to decide before they arrive in your district. This week's AI in education news: ChatGPT's new "dots" agents, the OpenAI model pulled over safety concerns, Meta's approval-first agent for small businesses, the America.gov AI assistant, the executive order directing federal agencies to say "Super Intelligence", Anthropic's leaked IPO prospectus, and Pope Leo XIV on AI risk. Each story comes with a practical action for principals, teachers and district leaders. Three actions for this week: build an agent register with an answer, draft, act rule, decide what student data never goes into shared AI workspaces, and put your AI vendors through a short risk and contract check. More at theaieducator.io
The UN launched a child-protection AI coalition this week, while Stanford's review shows schools have almost no US classroom evidence to go on. This week's episode covers the UN coalition and the White House access initiative, what NPR and Stanford found about the evidence base for AI in K-12, MIT's "illusion of learning" warning, England's Public First report on young people and AI, and what Massachusetts districts and the Microsoft/AFT privacy standard mean for your school's policies and contracts.
What should schools expect from the companies bringing AI into their classrooms? Microsoft President Brad Smith joins Dan Fitzpatrick to discuss the new AI Safety and Privacy Standard agreed with the American Federation of Teachers and its New York City affiliate, the UFT. Their conversation covers enforceable protections, school AI bans, teacher control, independent schools and whether the agreement could extend beyond the United States. Brad also addresses whether stronger safeguards could slow innovation. Dan then examines the latest PISA findings, the evidence behind Utah’s reported critical thinking gains, AI in special education advocacy and OpenAI’s Learning Lab. The episode closes with three practical actions to help schools check both the protections around their tools and what students are actually learning. Join The AI Educator Lab waitlist: https://lab.theaieducator.io Help shape an ongoing professional development journey and community, with no commitment required. Follow the podcast for your weekly roundup of AI news and research in education.
The two biggest school districts in America pulled students back from AI this week. In the same 72 hours, the AI labs handed teachers everything for free. Dan walks through the five stories that matter and what each one means for your school on Monday. NYC's one-year moratorium through eighth grade and the carve-outs nobody reported. LAUSD's district-wide restriction. The Royal Society finding that only 14% of teachers are confident with AI, with principals scoring lower than the teachers they lead. IBM's survey showing 77% of parents want input while 20% understand what's happening. OpenAI's 16-state privacy agreement, and why that matters more than the free access. And the Stanford review and Tennessee trial on what AI tutoring actually does. Plus three things to do this week: separate your staff and student policies, teach the thing before you police it, and redesign one task. The Educators' AI Guide 2027: theeducatorsaiguide.com Back to School AI Summit, 8-9 September: summit.aieducator.tools You can also watch on YouTube at https://www.youtube.com/@IamTheAIEducator