# Guided AI Use Gains Evidence as Schools Set Guardrails *By AI in EdTech Weekly • March 30, 2026* Research and implementation are drawing the same line: supervised, purpose-built AI can help learning, while open-ended answer tools often undermine it. This brief covers new tutoring evidence, NYC's traffic-light policy, narrow deployments in reading and counseling, and the global turn toward stronger edtech evaluation. ## Structured AI is pulling ahead of generic chat The clearest signal this week is not that AI tutoring works in the abstract. It is that **constrained, teacher-mediated AI** looks very different from open-ended answer machines. That distinction matters even more because model releases are now moving faster than traditional efficacy studies: panelists pointed to a flood of new GenAI studies, capability jumps every 5-7 months, and the practical problem that a model can change before a long RCT is even finished [^1]. In one of the strongest classroom findings this week, ED reported a randomized trial comparing human tutoring with a **supervised** AI tutor on its platform. Across more than 3,200 conversations, students tutored by the human-in-the-loop AI did better on the next math question than students tutored only by humans. The AI exchanges were longer and more Socratic, with more questions that surfaced student thinking and misconceptions [^1]. [![AI & Efficacy What Works, What Doesn’t, and What We Need to Know Webinar](https://img.youtube.com/vi/JD_cFNAPN-8/hqdefault.jpg)](https://youtube.com/watch?v=JD_cFNAPN-8&t=1335) *AI & Efficacy What Works, What Doesn’t, and What We Need to Know Webinar (22:15)* A very different result showed up in the Wharton math study shared by Ethan Mollick: students given ChatGPT during practice solved more problems, but the basic ChatGPT group later scored **17% worse** on a no-AI exam than the no-tech group. Researchers found many students were simply asking for the answer and later believed that had not hurt their learning [^2]. At the same time, Mollick pointed to a separate RCT showing that **well-prompted AI tutors** can boost learning, reinforcing the idea that prompt design and use constraints are not minor details; they are the difference between scaffolding and shortcutting [^3]. > “Learning needs to feel like a struggle. If you're struggling, you're learning. If it feels easy, you're not learning.” [^1] That principle also surfaced in practitioner commentary: panelists argued that narrow, teacher-mediated AI can help with scaffolding, reading, writing, math, and teacher time-saving, while wide, unscaffolded student-facing AI use can undermine cognitive and social-emotional development and make cheating easier [^1]. ## Policy is moving from abstract debate to usable rules New York City's Education Department moved the conversation from general concern to an actual framework. Its preliminary guidance uses a **traffic-light system**: green-light uses include brainstorming lesson plans and drafting non-critical communications; yellow-light uses include finding trends in student data, translation, and adapting materials for students with disabilities with trained human review; red-light uses ban AI from grading, special education and 504 planning, discipline, counseling and crisis intervention, and academic placement decisions [^4]. The city also drew a hard line on privacy. Personal student information cannot be entered into AI tools, approved products must go through a formal vetting process, and final guidance is due in June after public feedback. One unresolved issue: free tools do not go through the same contract review process [^4]. School-level practice in New York is already converging around the same supervised-use logic. East Side Community School prohibits the **unsupervised** use of generative AI for schoolwork and assessments, while Brooklyn Collaborative asks teachers to label each assignment with green, yellow, and red AI permissions. Many English and social studies teachers have also moved back toward in-class handwritten writing to reduce AI-assisted cheating, despite the time costs [^5]. The broader U.S. policy picture remains fragmented. Tina Austin, who advises on California education AI policy, described a landscape of framework fever, uneven district access to enterprise tools, and widespread confusion about using consumer AI with student data under FERPA and COPPA. Her practical advice is to start with local problems and school-approved tools rather than chase generic frameworks [^6]. ## The most credible deployments are narrow, grounded, and workflow-specific A useful counterweight to the hype came from EdSurge's conversations with 17 teachers: most are **not** reorganizing their classrooms around generative AI. They are using it first for productivity — lesson planning, newsletters, and administrative drafting — while testing instructional use cases more cautiously [^7]. Where classroom use does look promising, it is usually tied to a specific learning job. **Google Read Along** is a good example. The tool's AI tutor, Diya, supports phonemic awareness, phonics, fluency, vocabulary, and comprehension through leveled and decodable texts, read-aloud/silent/listen modes, real-time feedback, and comprehension checks [^8]. Inside Google Classroom, teachers can see accuracy, phonics gaps, fluency, comprehension patterns, and progress over time. Gemini can also help teachers create custom stories, re-level text, generate quizzes, and add their own content in multiple languages [^8]. [![Edu on Air: Turbocharge Reading Skills with Read Along & Gemini in Google Classroom](https://img.youtube.com/vi/AQ5DWP9Hdfo/hqdefault.jpg)](https://youtube.com/watch?v=AQ5DWP9Hdfo&t=1004) *Edu on Air: Turbocharge Reading Skills with Read Along & Gemini in Google Classroom (16:44)* Just as important are the limits. Read Along is framed as a **supplement**, not a replacement for teachers, and its strongest value comes from targeted practice and feedback rather than open-ended conversation [^8]. Google says the product has already supported hundreds of millions of stories read by tens of millions of learners, and highlighted pilots in India and the Philippines that found significant reading improvement, along with differentiated deployments in Pakistan, Malaysia, and Australia [^8]. Outside direct instruction, North Kitsap School District in rural Washington is using AI to strengthen **multi-tiered systems of support**. Staff use AI across well-being, academic, attendance, and behavior data to spot patterns, synthesize long plans, identify outlier interventions, and generate action steps. The district paired that with tiered professional development, including 27 lighthouse teachers who support classroom adoption [^9]. That kind of data work still needs human interpretation. Another Tech & Learning analysis warned that AI can surface patterns in underused school data, but schools should stay **data-informed rather than data-driven** because the same pattern can mean very different things depending on context, and leaders still need direct observation and professional judgment to validate what AI finds [^10]. Student support is also becoming a serious AI use case. High schools in New York are piloting **CounselorGPT** and **EVA** to answer procedural college-going questions, surface labor-market information, link students to resources, and give counselors better visibility into what students are asking. The goal is explicitly to free humans for fit, trust, and encouragement — not to automate the relationship itself [^11]. ## Global edtech is getting more evidence-conscious A second important shift this week is around **how** systems decide what to fund and scale. At UNESCO's Global Education Coalition meeting, participants argued for pedagogy-driven edtech transformation grounded in the science of learning and backed by evidence-informed investment, especially as AI-driven labor-market changes increase pressure on education systems to respond [^12]. That logic is starting to show up in financing and evaluation. ICEI launched an **EdTech Financing Advisory Facility** to help governments assess cost-effectiveness, learning outcomes, equity, ethics, and environmental considerations when making edtech decisions [^12]. UNICEF's **Blue Unicorn** portfolio is an even clearer signal. Its first cohort of edtech tools will be deployed across Egypt, Ghana, Malaysia, Rwanda, Uzbekistan, and Zimbabwe, with an explicit focus on foundational literacy, numeracy, teacher effectiveness, and inclusion. ICEI is running a quasi-experimental evaluation with about 600 lower-primary learners per intervention, using EGRA/EGMA-style measures, teacher surveys, and implementation data such as dosage, fidelity, and engagement [^13]. For education leaders and edtech investors, the question is moving from 'Does this tool have AI?' to 'Can this tool show learning gains, equitable access, and realistic implementation conditions?' [^12][^13]. ## What This Means - **For schools:** Treat AI as a design choice, not a category. The best results this week came from systems that kept AI narrow, supervised, and tied to a clear instructional or support role [^1][^8][^11]. - **For policy and procurement:** Low-risk drafting, medium-risk support, and high-risk student decisions need different rules. NYC's traffic-light model is one concrete template, but privacy and vendor review still need as much attention as pedagogy [^4]. - **For teachers and learning designers:** The current sweet spot is selective use — revision support, guided reading practice, data synthesis, and procedural advising — while keeping humans responsible for interpretation, motivation, and relationship-building [^7][^8][^10][^11]. - **For higher ed and workforce learning:** Institutions should expect private AI use and redesign around explanation, coaching, and real judgment. Mollick argues students already use AI quietly, universities are still figuring out how to teach in that world, and the traditional apprenticeship model is starting to fray when interns route first-draft work through AI [^14]. - **For researchers and investors:** Evidence cycles will have to speed up. If model capabilities change every few months, long trials alone will not be enough; faster research sprints and implementation-aware evaluations are becoming more important [^1][^13]. ## Watch This Space - **Faster evaluation models.** Investors and researchers are actively experimenting with research sprints and quasi-experimental designs because traditional RCT timelines no longer match model update cycles [^1][^13]. - **AI-native mastery platforms.** Khan Academy says its reimagined product is rolling out with clearer learning paths, a more central Khanmigo, proactive teacher assistance, and early pilot signals of higher skill growth when teachers assign yearlong units. Summit Public Schools argues AI only improves systems when schools are already clear about outcomes, adult roles, and the whole model [^15][^16]. - **Hands-on AI literacy.** BBC Bitesize built an AI guide around young people stress-testing AI in real time, against a backdrop where 47% of surveyed students already use AI for homework or revision and 24% say they do not know where to find trusted information about it. Teachers in EdSurge's study are also using AI literacy lessons to teach prompting, fact-checking, and bias rather than treat AI as authoritative [^17][^7]. - **Higher-ed and workforce pipelines.** Mollick's warning is bigger than cheating: if novices stop doing the early work that builds judgment, companies and universities may need new ways to develop deep knowledge, wide knowledge, taste, and agency [^18][^14]. --- ### Sources [^1]: [AI & Efficacy What Works, What Doesn’t, and What We Need to Know Webinar](https://www.youtube.com/watch?v=JD_cFNAPN-8) [^2]: [𝕏 post by @asanwal](https://x.com/asanwal/status/2037943638144205021) [^3]: [𝕏 post by @emollick](https://x.com/emollick/status/2037997630500856069) [^4]: [NYC releases guidelines for AI in schools. Some say it raises more questions than it answers.](https://www.chalkbeat.org/newyork/2026/03/24/preliminary-ai-policy-nyc-schools) [^5]: [NYC’s Education Dept. is releasing a long-awaited AI policy. Some schools already have their own.](https://www.chalkbeat.org/newyork/2026/03/23/schools-develop-ai-policies-awaiting-city-guidance) [^6]: [UnBlooms: Tina Austin on thinking well with AI and rethinking Bloom’s](https://www.youtube.com/watch?v=6iuhlMAmceQ) [^7]: [The AI Use Case Question Teachers Are Still Asking](https://www.edsurge.com/news/2026-03-27-the-ai-use-case-question-teachers-are-still-asking) [^8]: [Edu on Air: Turbocharge Reading Skills with Read Along & Gemini in Google Classroom](https://www.youtube.com/watch?v=AQ5DWP9Hdfo) [^9]: [Supporting MTSS With AI](https://www.techlearning.com/technology/ai/supporting-mtss-with-ai) [^10]: [Human Judgment Still Matters in an Era of AI-Powered Data Analysis](https://www.techlearning.com/technology/ai/human-judgment-still-matters-in-an-era-of-ai-powered-data-analysis) [^11]: [On-demand college counseling, courtesy of AI](https://www.chalkbeat.org/newyork/2026/03/27/high-schools-embrace-ai-tool-for-college-counseling-help) [^12]: [Investing in What Works: Rethinking EdTech Financing for Better Learning Outcomes](https://edtechpartnerships.substack.com/p/investing-in-what-works-rethinking) [^13]: [Evaluating EdTech at Scale](https://edtechpartnerships.substack.com/p/evaluating-edtech-at-scale) [^14]: [DD26 - From Disruption to Opportunity: Embracing the AI Revolution](https://www.youtube.com/watch?v=-10xtePtjuM) [^15]: [Fireside Chat with Sal Khan and Dr. Amy Alzina](https://www.youtube.com/watch?v=f88rKzt4IMw) [^16]: [Behind the Reinvention of Summit Public Schools with AI](https://michaelbhorn.substack.com/p/behind-the-reinvention-of-summit) [^17]: [#321 Navigating the Future of AI Education with BBC Bitesize](https://www.youtube.com/watch?v=2SfcWXueO24) [^18]: [Preparing for an AI-Driven Future in Business – Wharton in Focus: New York](https://www.youtube.com/watch?v=vmNLxEtfG8s)