🛝 What Happens When Everything Schools Rely On Stops Working

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

This week's reality check: A NeurIPS award-winning study found that 70+ AI models from different companies produce nearly identical essays, undermining every current detection approach. Meanwhile, AP scores doubled in a single year because the tests got easier, and a hack of an anonymous tip app may have exposed the most sensitive data schools hold on 30,000 campuses.

Data Gem

NWEA analyzed MAP Growth data from more than 3 million first-time kindergartners and found that about 5% were redshirted between 2017 and 2025. Those children showed short-term academic advantages, but the gains faded entirely by third grade.

70 AI Models Were Asked to Write an Essay. All Wrote the Same One.

Bruce Maxwell, a computer science professor at Northeastern University, was grading exams for his online master's course when he noticed something off. 

His students' essays used the same phrases, the same commas, the same word choices.

His former student, Liwei Jiang, now a Ph.D. student at the University of Washington, decided to test the hunch scientifically. Working with researchers at UW, the Allen Institute for AI, Stanford, and Carnegie Mellon, Jiang analyzed the output of more than 70 large language models, including ChatGPT, Claude and Gemini.

The team posed 100 open-ended questions to all 70 models, with each model answering 50 times. 

The result: answers were frequently indistinguishable across different models by different companies with different architectures and different training data. 

Same metaphors, same imagery, same sentence structures.

When asked to come up with a metaphor for time, the overwhelming answer from every model was the same: a river. 

When asked to write a story about a colorful toad, AI kept naming the toad Ziggy or Pip, and a hungry hawk and mushrooms kept appearing. 

Chinese-developed models produced similar answers to American ones.

The explanation is in how chatbots are built. The "alignment" step, designed to ensure responses are reasonable and appropriate, penalizes unconventional answers and favors safe, consensus-based ones. 

Originality gets stripped away.

For education, the finding reshapes how schools should think about AI in student work. 

The problem isn't just that one student used a chatbot. It's that an entire class submitting AI-assisted work will produce essays with the same metaphors, the same structure, and the same safe, consensus-driven ideas, because the alignment process that makes chatbots helpful also strips originality. 

For innovators the opportunity is in assessment formats that reward what AI structurally cannot produce: genuine creativity, personal voice, and demonstrated reasoning. Process-capture tools, oral assessment platforms, and portfolio systems that track how student thinking develops over time address a gap that detection tools alone cannot fill.

AP Top Scores Doubled in One Year. The Tests Got Easier.

Massachusetts politicians are celebrating the highest AP scores any state has ever received.

There's a problem. 

The scores almost certainly reflect easier tests, not more learned students.

On the AP U.S. Government exam, the share of students earning top marks (scores of 4 or 5) jumped from 24.1% in 2023 to 49% in 2024. 

The College Board admits its questions are easier and passing scores have been lowered. 

Their justification: AP is adapting to a "less demanding curriculum in high school and lowered expectations by colleges and universities."

Frederick Hess, director of education policy studies at the American Enterprise Institute, put it bluntly: "Students and families are happier because they get college credit. Schools are happier because they look good. Governors and state agencies are happier because they get to brag about it."

The trend extends beyond AP. National high school GPAs climbed over half a letter grade, from B to B+, between 1985 and 2020, according to NCES data.

A three-university economics team examined what that inflation actually costs. 

Studying teacher grading practices in Los Angeles and Maryland over a combined two decades, they found that students taught by teachers who inflate grades by one level are less likely to finish high school, less likely to enroll in college, and earn less as adults. 

The estimated annual societal cost: $213,872 per grade-inflating teacher, accounting for all students affected over an average career.

The researchers found that weaker teachers are more likely to inflate grades than effective ones, and that early-career teachers inflate more than experienced ones. Grade inflation may be used to ease student disappointment or to mask how little has been taught.

For education innovators, the data creates a case for assessment integrity. Standards-aligned grading platforms that calibrate teacher scoring against external benchmarks address the root cause. External validation systems, competency-based assessments, and employer-facing credential verification tools become more valuable as traditional grades and AP scores lose meaning.

School App Was Hacked. It Held Data on Self-Harm, Abuse, and Violence Threats.

Navigate360 markets its P3 Global Intel tip line to schools as an anonymous reporting channel. 

Students use it to report self-harm, abuse, substance use, and threats of violence. More than 30,000 schools rely on it.

A hacker claimed to have accessed Navigate360's systems, and early reports suggest the claims are legitimate. 

The full extent of the breach is still unclear.

David Riedman, founder of the K-12 School Shooting Database and a professor of security and risk management at Idaho State University, described the stakes: "This is an app that is sold to identify students who are thinking about self harm, being abused, abusing substances, or making threats of violence. That is the most sensitive information possibly available about a child."

The damage goes both ways. 

Students who submitted tips they believed were anonymous could now be identified, potentially making them targets.

Kenneth Trump, a school security expert, warned of the broader consequence: "School administrators work so hard to create that trust to get kids to come forward, and kids are not going to trust anonymous reporting if the system is actually not anonymous."

The incident follows a pattern. The PowerSchool breach exposed millions of student records and triggered dozens of lawsuits. In 2023, Raptor Technologies, another school safety vendor, leaked evacuation plans, lockdown procedures, and flagged student information through unsecured databases.

Doug Levin, national director of the K12 Security Information Exchange, recommended districts suspend use of the platform during the investigation.

For education innovators, the breach signals a market shift from "buy the cheapest safety tool" to "verify the vendor." 

Districts need continuous vendor security monitoring, procurement risk assessment tools, and privacy-by-design alternatives to current anonymous reporting systems. The fact that a tool designed to protect the most vulnerable students may have exposed them creates both urgency and a clear product specification: security, anonymity, and auditability as non-negotiable features.

⚡️More Quick Hits

This week in education:

• Science of reading bill wins unanimous committee approval in Congress — The Science of Reading Act (H.R. 7890) would require states receiving federal literacy grants to prove alignment with phonics-based instruction, with at least 35 literacy bills already enacted in 25 states

• Education Department plans to ease college merger pathways — Under Secretary Nicholas Kent said "not all" of the nation's approximately 6,000 institutions "are going to make it out of the next decade," announcing revised rules for nonprofit and for-profit mergers

• Conservative activists push alternative math standards across states — The National Association of Scholars' "Archimedes Standards" are influencing South Dakota's math rewrite, shrinking standards from full-length to just 36 pages, with Louisiana also reviewing

• Child care subsidy waitlists expanding in 14 states as pandemic funding expires — About 8 million children are eligible for subsidies but only 1.8 million receive them, creating a structural barrier for working families despite $39 billion in prior federal relief

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