AI tutoring apps take on the summer slide as research results split

A summer pitch with year-round consequences

In early August, WTOC aired a segment about an AI tutoring app meant to help students hold on to math and other subjects over summer break32. The segment is one example of a wider push. Companies are selling generative-AI tutors to parents as a cure for summer learning loss, at the same moment schools are trying to decide whether these tools belong in classrooms at all.

The products look alike. Lumos Learning's tutor, called Luma, recommends about 15 minutes of daily math and English practice to prevent the summer slide33. It has a free tier and a paid premium plan, plus a dashboard where parents can track progress33. Its marketing leans on parent testimonials and a before-and-after table, not independent evaluation33. UnlockGenius.io offers a free plan with one hour of AI tutoring a month and a $29.99 monthly K-12 plan, and it pitches tutoring in August as a way to get ready for school35. Thinkster Math pairs weekly one-on-one human sessions with a 24/7 AI math coach, starting at $22.50 per session34.

The marketing promises are clear. The evidence is less so. When you put the latest research next to the policy fights now under way in American schools, the honest answer is that AI tutoring can help, but only under certain design conditions that many consumer apps don't clearly meet.

The summer slide is real, but mostly a math problem

The problem these apps target does exist, though it's narrower than the ads suggest. An NWEA analysis reported in May found significant summer loss in math and much smaller loss in reading. Procedural skills, vocabulary, isolated facts and multi-step tasks are the areas most likely to fade without practice. Megan Kuhfeld, an NWEA researcher, said researchers still can't explain why some students do well over the summer while others fall far behind. She called the cause of that variation "the million dollar question".

This matters because summer-slide marketing often treats loss as uniform and predictable. Even one AI-tutoring vendor's blog acknowledges that NWEA's review shows a more nuanced picture than the old summer-slide story35. Prodigy, which sells game-based math practice, cites estimates that 70 to 78 percent of elementary students lose some math skill over the summer, with the steepest drop around the move from 5th to 6th grade. Taken together, the data supports the most modest version of the pitch: short, regular math practice addresses a real, measurable gap. The broad claims of across-the-board gains go further than the evidence.

What the research actually shows

The research on AI tutoring now splits into two camps, and this fall that split became hard to ignore.

On the positive side, the Stanford-affiliated National Student Support Accelerator described two randomized trials in which AI built into live, chat-based math tutoring improved results21. In one, supervising tutors approved 76.4 percent of a Google LearnLM model's responses with little or no editing. Students who worked with the model then succeeded on harder topics 66 percent of the time, compared with 61 percent for students tutored only by humans21. In the other, a tool called Tutor CoPilot made students four percentage points more likely to reach mastery, and up to nine points more likely when their human tutors were less experienced21. A widely cited Harvard physics trial found that a custom-built AI tutor helped students learn significantly more in less time than an in-class active-learning lesson. Students in the AI group also said they felt more engaged25.

The other camp got a major data point at the end of September. A University of Maryland team ran a randomized trial with 2,379 undergraduates and 30 instructors. Access to a course-integrated AI tutor lowered final grades by 0.37 standard deviations and cut participation in instructor-designed online activities by 0.90 standard deviations30. First-generation students lost more than twice as much as their peers24. The authors conclude that the tutor may have pulled students away from existing learning activities without improving performance24.

The same paper's literature review explains why the two sets of results differ. It cites work showing that open access to general-purpose AI hurt later independent performance in high school math, while tutoring guardrails reduced that harm24. It also points to evidence that generative-AI adoption in high school improved homework scores but lowered exam results24. The OECD's 2026 Digital Education Outlook reached a similar conclusion. Students using general-purpose chatbots produced better work, but that advantage disappeared or reversed on exams without AI. Tools designed with a clear teaching purpose showed lasting gains23.

Khanmigo is the most prominent tool that says it follows the guardrail approach. Khan Academy markets it as a tutor that guides students toward answers instead of giving them away, unlike ChatGPT8. Even so, the Maryland review notes that Oreopoulos and Low found only modest gains from Khan Academy combined with Khanmigo, and limited sustained use of the tutor24.

In short, design is the main factor in whether these tools help. A Socratic tutor that's built into a structured program and paired with human attention can raise achievement. A tool that hands out answers, or that replaces existing practice, can lower it. Some industry roundups describe AI tutors as delivering Bloom's "2-sigma" effect at almost no cost27. The 2026 trial evidence doesn't support that kind of claim.

Why that matters for a summer app

Summer is about the weakest setting for the conditions the research rewards. There's no teacher reviewing the AI's feedback and no structured course around the practice. Everything depends on whether a child keeps logging in. One review of 2026 trials found that gains concentrate in structured subjects like algebra and grammar, and that eight-week programs do better than two-week pilots that end before habits form28. That favors math-focused, standards-aligned practice, which is how Lumos describes itself33. It also means a free tier that families try in June and drop by July is unlikely to show up in fall test scores.

There's also an equity problem. Low-income students are the ones summer-learning programs most want to reach. Yet the strongest recent negative finding, from Maryland, fell hardest on first-generation college students24. If unsupervised AI widens gaps instead of closing them, a summer product with a free entry tier and paid premium features could make the divide worse.

Schools are moving in the opposite direction

While companies market AI tutors directly to families, some of the country's largest school systems are pulling back. New York City Mayor Zohran Mamdani announced a moratorium on student-facing generative AI. It bans companion chatbots in every grade, removes other student AI tools through eighth grade, and limits high school AI use to pilots reaching about 5 percent of students11. Los Angeles Unified said it would ban most student AI use on district devices this year11. Denver has banned ChatGPT but allows MagicSchool, an education-specific platform11. Boston gives students from third grade up access to approved tools under teacher supervision11.

The Denver approach, which bans general chatbots but allows a purpose-built tool, lines up with what the research suggests. Matthew Agnew, quoted in USA TODAY's coverage, makes the same distinction. He describes general chatbots as answer engines that can take away students' "productive struggle," while narrower Socratic tools keep students doing the problem-solving11.

The academic integrity gap

The AI-integrity debate usually centers on take-home essays, but it reaches tutoring apps too. Many consumer tools advertise that a student can photograph a problem and get the solution. CK-12's Flexi says it will solve a snapped math problem instantly and for free4. Mathos AI promotes instant step-by-step solutions9. In a summer setting with no grades, that's mostly a question of whether students actually learn anything. In September, the same tools come back into homework.

Policy hasn't caught up. A Chalkbeat review found that about three in four states have issued AI guidance for schools, but those documents mostly avoid the subject of cheating. They focus on AI literacy and efficiency for teachers rather than concrete steps to deter misuse14. More than a third of states warn schools not to rely on AI detectors alone14. In Manassas City, Virginia, a board member's proposal to add AI misuse to the student code of conduct failed after a colleague worried it would burden teachers who have no tools to identify AI-generated work17. Meanwhile, Pew found that 59 percent of U.S. teens say students at their school use AI to cheat at least somewhat often18. Teachers have responded by moving more graded work into class and onto paper1419.

The bottom line

AI tutoring for the summer slide addresses a real problem, especially in math. But it reaches families just as the evidence turns more cautious and schools tighten their rules. The research supports buying from the guardrail camp: tools that teach step by step and withhold answers, used consistently over weeks, ideally with an adult checking in. It doesn't support apps whose main feature is solving the problem for the child. Parents, like districts, should choose based on whether a product teaches or just answers, not on testimonials.