• Playground Post
  • Posts
  • πŸ› Schools Can See Every Student AI Conversation. It's Not Pretty.

πŸ› Schools Can See Every Student AI Conversation. It's Not Pretty.

What this means for educators + more

Welcome to Playground Post, a bi-weekly newsletter that keeps education innovators ahead of what's next.

This week's reality check: While an analysis of 1.2 million student AI interactions reveals one in five cross district boundaries, two university-led studies prove virtual tutoring can deliver months of reading gains that persist into second grade. Meanwhile, child care costs have crossed $28,000 a year and 61% of providers say families can't afford to pay.

Data Gem

About a third of U.S. schools have returned to pre-pandemic achievement levels in reading or math. But when you look at both subjects: just 1 in 7 schools have recovered in both math and reading. Schools serving higher-poverty and historically marginalized students saw the largest gains but remain the least likely to have fully recovered.

Schools Can Now See Every Student AI Conversation

Securly, a company offering internet filtering and safety services, analyzed nearly 1.2 million student interactions with generative AI across more than 1,300 districts between December 2025 and February 2026.

The finding: roughly 1 in 5 interactions involved cheating, self-harm, bullying, or other problematic behavior. 

About 1 in 50 were red flags for violence, cyberbullying, or self-harm.

Unlike most research on student AI use, which relies on surveys and self-reporting, this data captures what students actually type. Jeremy Roschelle, co-executive director of learning science research for Digital Promise, a nonprofit focused on equity and technology issues in schools, called the data set unique.

"What are students really doing when they're writing text into generative AI?" Roschelle said. "That's why it's fascinating."

In November, Securly began allowing district officials to set parameters around students' AI use, similar to website filtering. Large language models will "deflect" queries that fall outside district policy. Nearly all of the deflected queries, 95%, were from students trying to get AI to complete their schoolwork.

Securly CEO Tammy Wincup wasn't surprised. She expects students to "experiment with understanding the guardrails" and try to get around them.

The remaining flagged interactions painted a more troubling picture. 

About 2% of interactions were identified as potentially unsafe, compared to just 0.4% for traditional internet searches. That 2% represents more than 24,000 queries. 

In one case, a student directed a large language model to help draft an email to their mother explaining they had suicidal thoughts. In another, a student searched questions about nerves and arteries before switching to an AI tool and asking how to commit suicide. In both cases, student identities were unmasked and district officials were notified.

ChatGPT dominated student usage, accounting for 42% of all interactions. Google's Gemini made up 21%. Education-specific tools like MagicSchool, SchoolAI, and BriskTeaching comprised just 9%.

For education innovators, the data reveals three distinct product needs. Districts want AI governance dashboards that give administrators visibility into how students interact with large language models on school networks and devices. There's growing demand for classroom-safe LLM wrappers that deflect inappropriate queries while keeping learning productive. And the biggest gap is tools that help educators determine whether "appropriate" AI interactions are actually advancing learning or just avoiding harm.

Virtual Tutoring Adds Five Months of Reading Gains

Struggling readers in Missouri's Kansas City Public Schools showed statistically significant gains in literacy when virtual high-impact tutoring was embedded in their school day.

The intervention: 30 minutes of 1:1 virtual tutoring, at least three times a week, over 20 weeks.

The result: participating students saw 10.84 percentage points of additional annual typical growth and 5.24 percentage points of additional stretch growth on the i-Ready reading assessment, according to a study from Stanford University's National Student Support Accelerator

First graders gained nearly an additional month of learning. Fourth graders gained about two months.

A second study, released in January by the Center for Research and Reform on Education at Johns Hopkins University, found similar results in a different context. Researchers analyzed a 1:1 virtual tutoring literacy intervention based in the science of reading across 13 Massachusetts public school districts.

The 1,872 participating first graders made progress equivalent to more than five months of additional instruction. 

On average, students received 33 hours of tutoring over 36 weeks.

85% of students who completed the program at or above grade level by the end of first grade maintained those gains into second grade without additional tutoring.

These results arrive at a moment when most schools still haven't recovered from pandemic learning losses. The evidence is strong, but the question is scale.

Scalable virtual tutoring platforms that deliver consistent 1:1 sessions need tutor supply networks and scheduling infrastructure to match. Dosage-tracking and outcomes-measurement tools can help districts prove ROI and sustain funding beyond the pandemic recovery window. 

The persistence finding is the most investable signal: if interventions that get students to grade level by end of first grade don't need to continue into second grade, the cost equation for district-wide adoption fundamentally changes.

65% of Child Care Programs Raised Tuition. Now Families Are Leaving.

In Philadelphia, Mary Graham has run the early learning program Children's Village since it opened in 1976. 

This year, for the first time in more than three decades, the program is running a deficit.

The deficit: $200,000.

Liability insurance jumped from $45,000 in 2024 to $62,000 this year. Costs for food, health insurance, and worker's compensation all climbed. Graham had to cut back on substitutes and scale back planned salary increases for her 76 full-time staff.

"I almost had a heart attack," Graham said of the insurance bill.

She isn't alone. A new report by the National Association for the Education of Young Children, based on a survey of more than 7,000 early childhood educators, found that 65% of center-based providers and 31% of home-based providers raised tuition in the past year. 

The cost for food and supplies increased the most, followed by facilities maintenance and liability insurance.

More than half of the programs that raised prices reported families leaving as a result. 

61% of respondents said their programs are underenrolled because families simply cannot afford to pay.

A January 2026 LendingTree study found the average annual cost of child care for an infant and a 4-year-old now exceeds $28,000. 

A family with two children would need to earn more than $400,000 for child care to account for 7% or less of household income, the federal benchmark for affordability.

The timing makes it worse - the federal pandemic-era child care funding has expired. Some states are making up for budget shortfalls by slashing state funding for child care, adding pressure to an industry already running on thin margins.

Graham's response captures the field: do more with less. Despite an increase in children identified with disabilities in her program, she cannot add the extra classroom teachers they need.

"Kids need it, but we can't," Graham said.

For education entrepreneurs, the structural crisis is the product opportunity. The underenrollment problem points to flexible pricing tools, employer-subsidized benefit platforms, and family financial assistance navigation that can help programs fill seats without pricing out the families they serve. 

And the insurance cost spiral suggests demand for group purchasing cooperatives or aggregation platforms that could prevent a $45,000 policy from becoming $62,000 in two years.

⚑️More Quick Hits

This week in education:

β€’ California fines edtech company $1.1 million for violating student data privacy β€” PlayOn, which contracted with roughly 1,400 schools, forced students to consent to data-sharing through its GoFan ticketing platform just to access event tickets

β€’ Indiana governor signs law cutting college degrees where graduates don't out-earn high school grads β€” Public colleges must end programs below $24,000 to $35,000 earnings thresholds; over 400 degrees already eliminated under a prior law

β€’ Kids Internet and Digital Safety Act advances to full House vote β€” The bill would require chatbots to disclose they are not human, show suicide hotline resources, and encourage breaks after three hours of use

β€’ Schedule dissatisfaction hits teacher morale 3.6 times harder than salary β€” EdWeek's survey of 5,000 teachers across all 50 states finds the Teacher Morale Index dropped from +18 to +13, with 64% reporting worsened student behavior

To stay up-to-date on all things education innovation, visit us at playgroundpost.com.

What did you think of today’s edition?

Login or Subscribe to participate in polls.