AI Tutoring Tools Flood Schools as New Trials Show Mixed Results

A school year built around AI

Before the 2026-27 school year started, the biggest edtech companies released a wave of AI features for classrooms. Under the hood, the products have a lot in common. The common thread is that they no longer behave like a general chatbot. Each one sits inside the class, the assignment and the curriculum. In a roundup of ten back-to-school updates, Forbes contributor Dan Fitzpatrick described the move from basic chatbots to workflow systems that know the context they're working in. His examples included Gemini in Google Classroom, Copilot's learning coach, OpenAI's K12 Educator plugin and a ChatGPT version for teenagers.2

School policy pulled in two directions at once. New York City and Los Angeles both restricted student use of generative AI in early September, even as states and vendors pushed adoption harder.11 The research is no more settled than the rules. A large randomized trial published this month found that an AI tutor lowered grades. Other studies report solid gains.3023 Read together, the coverage points to one conclusion: what matters is how a tool is designed and whether students actually use it, much more than whether AI is in the classroom at all.

What the vendors shipped

Google made the broadest push. Gemini now uses a class's own materials, including the assignment title, the teacher's instructions and the curriculum, as context. Students can turn sources their teacher picked into flashcards, quizzes, study guides and audio overviews.2 Google says school administrators decide whether Gemini is available in Classroom, and students of every age can be given access.4 The company's study notebooks start with a diagnostic quiz and then build a personal plan with shorter lessons. Students can also copy notebooks their teachers created and edit them.2 Google's own announcement stressed tools that let teachers shape how students use AI and follow their progress as it happens.1

Microsoft folded several products into one education system. Study and Learn is a Copilot coach for students aged 13 and older. Unit Plans drafts full units from the grade, subject and standards a teacher enters. Assignments now include Student AI Guidelines.2 With those guidelines, teachers can mark each task as allowing no AI, limited help or full Copilot use.4 Microsoft says the Study and Learn agent is built to keep the "productive struggle" learning depends on, rather than handing over answers. It now comes up as the default option when K-12 students with Copilot Chat access open the tool.8

OpenAI went after teachers and teenagers at the same time. Its K12 Educator plugin builds materials around local standards, but only for ChatGPT Edu and ChatGPT for Teachers district deployments. It also launched ChatGPT for Teens for users aged 13 to 17, which includes Study Mode and reminders about responsible homework use.2 On September 1 it expanded ChatGPT for Teachers to 55 more U.S. school systems, bringing the total to more than 340,000 educators in 30 states.6 Anthropic had already made Claude for Teachers free for verified U.S. K-12 educators through at least June 2027.6

Smaller companies are fighting for the same teacher workflows. Canva's Learn Grid pairs more than 50,000 curriculum-mapped resources with AI-generated activities. MagicSchool redesigned its platform around an assistant that recommends which of its 80-plus tools to use. Kahoot now runs inside ChatGPT, and Diffit's Chrome extension turns any webpage into handouts and assessments.2 Nearpod added an AI text leveler and lesson and activity generators, though only with school and district licenses.5

The tutoring evidence splits

The research is where the coverage disagrees most. Optimists usually point to a Harvard physics study. In coverage of that study, students using a custom-built AI tutor scored about 30% higher on a later test than students in an active-learning classroom, and they took less time.27 Those same reports note the limits: 194 Harvard undergraduates, one subject and a tutor written with carefully designed prompts.27 Some summaries go much further and claim that students in AI-enhanced settings score 54% higher.24 That number seems to come from an edtech vendor's internal impact study, not independent research, and it deserves skepticism.21

The newest large trial points the other way. An EdWorkingPaper released in October tested a course-integrated AI tutor with 2,379 undergraduates and 30 instructors at a large U.S. public university. In matched sections, access to the tutor lowered final grades by 0.37 standard deviations and cut participation in the learning platform by 0.90 standard deviations.30 First-generation students saw grade losses more than twice as large as their peers.26 About 74% of student requests asked for information, explanations or solutions. Fewer than 1% asked for feedback on their own work.26 The authors are clear that the data cannot show that lower engagement caused the lower grades. But the pattern suggests students used the tutor as a shortcut, not as a study partner.26

Results from K-12 fall in between. Stanford's National Student Support Accelerator reviewed two randomized trials. In one, Google's LearnLM matched human tutors at helping students fix mistakes, and its students did better on harder topics that came later. In the other, tutors using Tutor CoPilot were four percentage points more likely to get students to topic mastery, with the biggest gains among less-experienced tutors.23 A review of K-12 math pilots found effects of about 0.2 to 0.4 standard deviations, mostly in structured settings, and much weaker evidence in literacy.22 Other coverage of tutoring trials found that gains cluster in subjects like algebra and that AI tutors tend to tie or trail small-group human tutoring when both are well staffed.25

The most practical finding came from Stanford's SCALE Initiative. Its AI literacy tutor barely got used. Students in one setting averaged only a few minutes a week, many never logged in, and reading scores showed no meaningful difference.28 The lead researcher said access to an AI tutor is not the same as using it. Students who did log in were more likely to already be high performers.28 A Wharton trial of an AI lesson-planning assistant raised a different warning. In classrooms where marks had been lower before the study, achievement and student confidence both fell over the 10 weeks.29

The key point is that these findings fit together better than they first seem to. Purpose-built tutors that guide students through problems in structured subjects tend to help. Open-ended assistants used for answers tend not to, and they may hurt the students with the least support. The new university paper itself argues that purpose-built and general-purpose AI tools should be evaluated separately.26 Vendors have adopted the language of guided learning, but none of the new products has public outcome data of its own yet.

Policy fragments

Policy has not kept up with how fast the tools are spreading. New York Mayor Zohran Mamdani announced a moratorium on student-facing generative AI through eighth grade and banned companion chatbots in every grade. About 5% of high schoolers will join AI pilots, and all high schoolers will take AI literacy classes.11 Los Angeles Unified said it would ban most student AI use on district devices this year. Boston gives students supervised access to approved tools starting in third grade, and Denver blocks ChatGPT but permits MagicSchool.11

Other states are moving faster toward adoption. Utah agreed with Google to deploy Gemini for Education in every public school, reaching about 680,000 students at no cost.6 Washington, D.C.'s first AI rulebook for teachers uses a stoplight system and bars AI from grading, discipline, IEP decisions and teacher evaluations.6 One count puts the number of states with published school AI guidance at 35, plus Puerto Rico.9

Academic integrity is the gap

State guidance leaves cheating mostly unaddressed. A Chalkbeat review found that about three in four states have published AI guidance for education. Those documents spend more space on AI literacy and teacher efficiency than on concrete ways to prevent cheating.14 More than a third of states warn schools not to rely only on AI detectors, which can wrongly flag genuine student writing. That leaves teachers with little beyond acceptable-use policies.14 Half of the ten largest U.S. districts explicitly discourage detection tools.17

The risks are now showing up in court. A Palo Alto family filed a federal civil rights lawsuit after their son was flagged by detection software and his semester grade was cut. They want the court to stop the district from treating detector scores as decisive without review by an educator.19 Meanwhile, a Pew survey reports that 59% of U.S. teens say AI cheating happens at least somewhat often at their school.12 Teachers are responding with more graded work done in class, often on paper.14 Others are moving toward oral defenses and portfolios.25

The reading

This season's product launches agree on one thing: AI should be grounded in the curriculum and controlled by teachers. That direction fits the strongest research, which favors guided, structured tutoring over open-ended answer engines.1126 It does not fix the problem of students not using the tools, the risk that AI widens gaps, or the missing integrity rules, all of which the research and policy coverage keep coming back to.2814 Teachers will spend this year sorting out those issues themselves, district by district, while the companies keep releasing features.