Half of U.S. school districts now say they’ve trained their teachers on AI. A year earlier, it was about a quarter. That’s a real, fast shift — backed by a nationally representative RAND survey of the American School District Panel, not a press release.
Here’s the part worth sitting with: when RAND’s researchers asked district leaders what “training” actually meant in practice, the honest answer, over and over, was we made it up as we went. Eleven of the fourteen district leaders RAND interviewed built their AI training programs themselves, from scratch, because the outside options weren’t good enough. One leader put it plainly: “There are people that are claiming to have the best practices and are making money hand over fist… if they claim to be telling you best practices, they don’t have them yet. They don’t exist yet.”
That’s not a knock on those district leaders — they’re doing real work under real time pressure. It’s a diagnosis of the moment. AI arrived in classrooms faster than anyone built the professional development to match it, so schools are running the experiment live, with actual students, mostly measuring success by whether teachers stopped being afraid of the tool.
That’s the real gap this piece is about. Most of the training districts describe is tool training: how ChatGPT works, how to write a prompt, how to use an AI lesson-planning assistant. That’s a reasonable first step — teachers who are anxious about a tool can’t use it well, and RAND found addressing that fear was nearly every district’s starting point. But tool literacy and facilitation are two different skills, and only one of them protects the thinking in the room.
A teacher who knows how to prompt ChatGPT can still end up, without meaning to, running a classroom where the AI does the reasoning and the student does the copying. Nothing about “how to use the tool” teaches a teacher to notice that moment, or to redirect it. That’s a facilitation skill — the ability to read whether a student is genuinely stuck or just handing off the thinking, and to ask the next question instead of supplying the next answer. It’s the same shift Auxesis’s Educator Track builds toward: moving from Explainer, who fills every silence with an answer, to Facilitator — from Practitioner to Coach.
Picture two versions of the same fifth-grade classroom, six months into AI rollout (illustrative, not a real case). In the first, teachers got a solid half-day on an AI planning tool and were told to “play around with it.” Word-problem homework improves fast — suspiciously fast. Kids paste the problem into a chatbot, copy the steps, turn it in. The teacher, trained on the tool but not on what to watch for, sees clean homework and reasonably assumes the class is ahead of schedule. In the second, teachers got the same tool training, plus one more habit: asking “walk me back through how you got there” before accepting an answer as done. Same tool. Very different classroom, six months in.
That second habit is what Auxesis calls a Facilitation Brief — a short, structured read before a session on where a student actually stands, so class time goes to facilitating instead of re-diagnosing from scratch. It’s one piece of COMPASS, the operating layer Auxesis builds around AI in the classroom: AI helps before the session and after it, in the briefing and the five-minute note; the actual teaching stays human, on purpose.
There’s also an equity story here, and it’s worth naming plainly. RAND found low-poverty districts have consistently trained teachers on AI faster than higher-poverty ones — 43% versus 6% in fall 2023, 67% versus 39% by fall 2024 — and district leaders’ own projections show that gap holding into the 2025–2026 school year, with almost all low-poverty districts trained and only around six in ten high-poverty districts there yet. Whatever training model turns out to work will reach wealthier schools first. That’s one more reason the training that does exist should be built around facilitation, not just tool onboarding — a district that gets one real shot at AI professional development shouldn’t spend it on prompt-writing alone.
None of this argues against training teachers on AI faster. It argues for training them on the right thing. The tool is the easy part to teach. Reading a classroom, protecting productive struggle, and knowing when to step back instead of stepping in — that’s the harder, more durable skill, and it’s the one most current AI-in-schools training is skipping past.
If you’re building or choosing AI professional development for your school or district, the question worth asking isn’t “does this cover the tool.” It’s “does this teach my teachers to notice when a student stopped thinking.” The Educator Track’s facilitation modules exist to answer exactly that — not as a replacement for tool training, but as the layer most programs are currently missing.
Source: Melissa Kay Diliberti, Robin J. Lake, and Steven R. Weiner, “More Districts Are Training Teachers on Artificial Intelligence: Findings from the American School District Panel,” RAND Corporation, 2025.
A follow-up to The 17% Tax — same research family, different question.
We wrote recently about the OECD’s finding that AI-assisted math practice can look great and still leave nothing behind — up to 17% worse performance once the AI is taken away. That post ended with one diagnostic question: “walk me through how you got this.”
There’s a second piece of evidence worth its own post, because it answers a question the math study doesn’t: how fast does the gap open, and does it show up outside of math?
The one-hour test
A study cited in the same OECD Digital Education Outlook 2026 had students across several US universities write a short essay — one group alone, one with a search engine, one with a general-purpose chatbot doing much of the drafting. One hour later, researchers asked each student to quote a sentence from what they’d just “written.” Among the unaided and search-engine students, 89% could. Among the chatbot group, only 12% could.
Worth being precise here, in the same spirit as the caveat we ran last time: the specific study behind this appears to be MIT Media Lab’s “Your Brain on ChatGPT” research (Kosmyna et al.) — a small trial (54 participants), still a preprint, not yet peer-reviewed. Different write-ups of it report slightly different numbers (some cite 90%/17% instead of 89%/12%), which is normal for early-stage research moving through secondary coverage, but it means this shouldn’t be treated as a settled, precise figure — just a strong, repeatable signal in the same direction as the math result: fast AI produces work that looks finished but was never really held by the student who “wrote” it.
One hour. Not a semester, not a unit test — sixty minutes was enough for four out of five students to lose their grip on their own sentences.
Why one question isn’t enough for this one
The “walk me through how you got this” check works well for a worked math problem because there’s a step-by-step path to retrace. Writing doesn’t hand you that same rope. A finished essay doesn’t show its work the way a solved equation does — which means the single-question check from the math post genuinely won’t catch this failure mode. You need something with more structure.
That’s what Auxesis’s Metacognition framework is for — three questions, asked in sequence, that work regardless of subject:
- Monitor — “Before you turn this in: what’s the one sentence in here that’s most you? Point to it.” If your child can’t find one, that’s the tell — not a bad grade, just a flag that something outside their own head produced the words.
- Regulate — “What would you do differently if you had to write this again without any help?” This isn’t a punishment question. It’s the moment that turns a flagged gap into an actual second pass — the regulation step is what separates “I noticed this wasn’t mine” from “I fixed it.”
- Evaluate — Circle back a day or two later, unannounced, with a version of the one-hour test: “Explain the argument you made in that essay.” If they can’t, you’ve learned something real about the assignment — and caught it while it’s still cheap to address, not at the next test.
Why this can’t be a one-time fix
The math study and the essay study point at the same underlying mechanism from two directions: performance during the task tells you almost nothing about what’s going to stick. The only way to know is to check after — which is exactly what Monitor/Regulate/Evaluate is built to do, and exactly what a finished worksheet or a polished essay can’t tell you on its own.
This is also why it’s a framework and not a one-off question: the math post’s single check catches one failure mode; a repeatable three-step loop catches it across subjects, because “did my child actually think this through” isn’t a math-specific problem — it’s the same one whether the tool wrote an equation or a topic sentence.
If you want the full walkthrough of how we teach Socratic questioning at home — the same instinct that powers Monitor/Regulate/Evaluate — that’s covered in the Parent Track.
If you’ve used AI to help your child with homework, you’ve probably felt a small flicker of guilt about it. New research says that flicker is worth paying attention to — but not for the reason the headlines suggest.
Over the past month, a wave of studies has landed with real numbers attached to something a lot of parents have sensed intuitively: leaning on AI to get answers is different from using it to get smarter.
What the research found
A study of 1,222 people — conducted across teams at Oxford, MIT, UCLA, and Carnegie Mellon and reported by Forbes on June 30, 2026 — found that AI assistance measurably reduces independent performance and persistence, and that this effect shows up fast: within 10 to 15 minutes of AI-assisted work.
Separately, an analysis covering 26,000 students (reported by Psychology Today) found learning losses of 25 to 30 percent on math work once students started leaning on AI to get through it — the study tracked how much time students spent on problems that were easy to hand off to a chatbot, and found that less time on the problem meant less learning from it.
And in higher education, 90 percent of faculty surveyed now say AI is weakening students’ critical thinking.
Three different research teams, three different populations, one consistent finding: when AI does the thinking, the student doesn’t.
Why this isn’t actually surprising
None of this means AI is bad for learning. It means answer-first AI is bad for learning — which is a much more useful thing to know, because it tells you exactly what to change.
Struggle is not a bug in the learning process. It’s the mechanism. When your child sits with a hard problem, tries something, gets it wrong, and tries again, that friction is where the actual wiring happens. An AI tool that skips straight to the answer isn’t saving your child time — it’s skipping the part of the assignment that was actually the assignment.
This is the same idea behind metacognition — one of the core frameworks we teach in the Singapore Stack: the habit of monitoring your own understanding ("do I actually get this, or does it just look familiar?"), regulating your approach when you’re stuck ("what should I try differently?"), and evaluating afterward ("did that actually work, and why?"). A student who hands a problem to AI the moment it gets hard never runs this loop. A student who’s taught to run it — with or without AI in the room — builds the skill that the loop itself was practicing.
A concrete example
Say your child is stuck on a word problem. The answer-first move is: type it into a chatbot, get the answer, copy it down, move on. Ten seconds, zero learning.
The metacognitive move looks different: ask your child to explain what the problem is actually asking before touching any tool. Have them guess roughly what a reasonable answer would look like. Let them try it — get it wrong, even. Then bring in AI, not to hand over the answer, but to ask a question back: "What’s one thing you could check about your work?" or "What would happen if you tried it a different way?"
Same tool. Completely different result. One erodes the skill the homework was supposed to build. The other builds it.
What to do this week
You don’t need to become an AI expert to make this shift — you need about five minutes and a habit. Before your child uses any AI tool for schoolwork, run this quick check:
- Has my child tried this on their own first, even briefly?
- Is the AI being asked for an answer, or asked a question back?
- Could my child explain how they got the answer, out loud, right now?
If the answer to #3 is no, the tool did the thinking instead of your child.
We’ve built this out into a fuller "5-Minute AI Use Audit" — a short, printable checklist you can keep next to the homework table. It’s free, and it’s below.
Why we’re writing this now
This isn’t abstract for us. Inside our own community, the single most-discussed question right now is simply: "How are you using AI with your kids?" Parents are asking this in real time, without a clear answer in front of them. This research — and the framework above — is our answer.
AI isn’t the enemy here. Skipping the struggle is. Used the right way, AI can be a genuine thinking partner for your child. Used the wrong way, it’s just a very fast way to look like you learned something you didn’t.
Download the 5-Minute AI Use Audit checklist here.
For two years, Auxesis has made an argument that felt, at times, like we were making it alone: that a student can look like they’ve mastered something — fast, fluent, high scores — while the actual thinking never happened. We call this the illusion of mastery. It’s the thing we exist to fight.
This month, two independent education writers walked into a well-funded, fast-growing math platform called Math Academy and came out describing the exact same problem, in almost the exact same words — without ever having heard of us.
What happened
Math Academy markets itself on speed. Math educator Michael Pershan spent a month inside the platform, working through a full course, to see what that speed actually produced. His account, Math Academy: A Mixed Review, is careful and fair, but one line does the most damage: “Math Academy offers direct instruction for procedures, discovery learning for concepts.” In practice, he found, the platform hands you the steps for solving a problem clearly and quickly — and leaves you to construct the why almost entirely on your own, with no one checking whether you actually did.
Pershan’s sharper point is about incentives, not just design. The platform awards experience points for completing exercises, he writes, not for reading the conceptual explanations sitting alongside them: “If the exercises don’t require the concepts, then the concepts only inhibit your progress and kids will drive past them at 75 mph.” A student can hit every target the app measures without ever slowing down for the part where understanding actually forms.
Dan Meyer picked up Pershan’s review and went after the marketing claim itself in a piece titled “It Is Fun to Pretend That Hard Things Are Easy!” He points out a pattern common to platforms that promise to have cracked the code on faster learning: they quietly redefine what “learning” means — not to educators, not to universities, not to the people who eventually have to use the math for something real, just to their own dashboard. Fast completion of exercises gets rebranded as mastery. The gap between the two is where the illusion lives.
Why this matters beyond one platform
Neither Pershan nor Meyer set out to make our argument. They set out to review a product. That’s what makes it useful. When a business built around a critique makes that critique, it’s easy to dismiss as self-interested. When two working educators, with no connection to each other or to us, independently spend real time inside a platform and land on the same conclusion — using almost the same language — that’s a signal the pattern is real, not a marketing angle.
The pattern, stated plainly, and stated by them, not us: speed and fluency are being measured and rewarded. Understanding is not. A system that only measures what’s easy to measure will quietly train students, and the adults watching them, to mistake the measurement for the thing itself.
What we’d ask instead
Auxesis builds around a different question: not “did the student finish,” but “can the student explain it, apply it somewhere new, and still have it a month from now?” We call that the difference between Explain, Apply, and Sustain — the three conditions that have to hold before we call something mastered. A student can pass a platform’s exercise set, by Pershan’s own account, and still fail all three. That gap is not a rounding error. It’s the whole problem.
None of this means adaptive practice tools are worthless — deliberate, well-designed practice has a real place, and Pershan says as much in his review. The problem isn’t that Math Academy exists. It’s that “fast” and “learned” are being treated as the same word, at a moment when a lot of anxious parents are looking for a shortcut and being told, credibly, that one exists.
The reviewers didn’t need us to tell them that. They found it themselves, inside the product, and wrote it down. Our job now is simple: point at what they found, and say it plainly, before “cognitive offloading” gets flattened into a buzzword nobody bothers to define.
Researchers finally have a name for the thing we’ve been warning families and schools about since Auxesis started: cognitive offloading.
Here’s the plain-English version. When a tool does the thinking for you — solves the problem, writes the paragraph, finds the pattern — your brain quietly hands that job over. You stop doing the mental work that used to be yours. That’s offloading. Calculators and spellcheck have always done a small version of this. What’s new is how much of the thinking itself a fluent AI system can now take over, and how convincing the result looks while it’s happening.
In March 2026, Professor Leslie Loble of the University of Technology Sydney and Professor Jason Lodge of the University of Queensland published a report on exactly this risk. Their finding, in their own words: there’s "a growing body of evidence that using AI can short-circuit the cognitive effort required for sustainable, deep learning, thus creating ‘false mastery’ with potentially long-term consequences." They describe a "performance paradox" — a student’s short-term performance on a task improves with AI help, while their durable, long-term learning is quietly harmed. The mechanism, they write, is that AI’s fluent output "creates an illusion of competence and encourages metacognitive laziness, leading learners to abdicate the generative effort required to build deep knowledge." (Read the full report: Artificial intelligence, cognitive offloading and implications for education →)
False mastery. Illusion of competence. We’d call that same thing the illusion of mastery — a student, a class, or a whole program that looks like it’s working because the outputs look right, while the underlying understanding quietly never forms. We didn’t wait for the research to catch up to build against it. COMPASS, Auxesis’s AI operating system for schools and tutoring programs, is built around one question: is the thinking actually happening, or just the appearance of it?
Here’s how that shows up in practice. COMPASS runs on three loops around every session, not just the session itself:
Before — a Facilitation Brief tells the educator what a student actually struggled with last time, not just what score they got. If a student aced a worksheet but the brief shows three failed first-attempts before the right answer, that’s a flag worth reading before class starts, not after a report card confirms it.
During — teaching stays human. COMPASS never puts an AI tutor between a student and a live struggle. The system’s job is to inform the educator, not to replace the moment where a student has to think something through themselves.
After — a five-minute note habit. The educator logs what actually happened: did the student explain their reasoning, or just arrive at the right number? That note is what makes the next Facilitation Brief honest instead of guesswork.
A concrete example: say a student breezes through a set of AI-assisted practice problems and the dashboard shows 90% correct. Looks like mastery. But the educator’s five-minute note says the student needed the AI’s hint on nearly every problem to get there. That’s the gap between offloading and understanding — Loble and Lodge’s "performance paradox" in miniature — and it’s invisible unless something is built to catch it. That’s the whole reason COMPASS’s after-session loop exists.
If you’re leading a school, a co-op, or a tutoring program and you’re already seeing this gap — kids who "get it" in the moment but can’t hold onto it a week later — this is worth a conversation. Talk to our team about COMPASS →
And if you’re an individual educator wrestling with the same question in your own classroom, the Educator Track’s facilitation training is built around exactly this distinction: the difference between explaining and facilitating, and how to tell which one is actually happening in the room. Explore the Educator Track →
Researchers now have peer-reviewed language for what we’ve been building against for a year. That’s a good thing — it means the problem is real, named, and coming from serious people, not just us saying so. What matters now is what you do about it.
Source: Loble, L. & Lodge, J. (2026), "Artificial intelligence, cognitive offloading and implications for education," Australian Network for Quality Digital Education / University of Technology Sydney. Read the full report.
Picture two kids working through the same math unit. Both use a chatbot to help with homework. One breezes through every practice problem and turns in a flawless worksheet. The other struggles through the same problems more slowly, makes mistakes, and asks for help twice.
A month later, they sit the same closed-book test — no chatbot allowed. The one who breezed through practice does worse. Not a little worse: up to 17% worse, according to a large randomized study cited in the OECD’s Digital Education Outlook 2026, a 245-page report released this January.
The study wasn’t a hunch or a survey. It was a controlled trial with high school students in Türkiye. When students got access to GPT-4 for math practice, their practice scores jumped by as much as 127%. Then, on the exam that actually counted, those same students scored worse than peers who’d studied the material without any AI help at all.
The OECD has a name for that gap: the "mirage of false mastery." Homework looks done. Practice scores look great. But the thinking homework is supposed to build never actually happened.
Worth being precise about scope here: this is one (large, well-designed) field experiment in one subject, cited as evidence inside a much bigger report — not a claim that every use of AI tutoring produces this exact result. But it’s the clearest number-backed version yet of something we’ve been saying since Auxesis started.
Same enemy, new evidence
We’ve called this the illusion of mastery from day one — a good grade, a clean worksheet, or a nodding head that hides the fact that real cognition never occurred. It’s the thing this company exists to fight. Now an intergovernmental body, working from research across dozens of countries, has independently landed on the same conclusion, in almost the same words.
Why the gap happens
The mechanism isn’t mysterious. When a student hands a problem to a generic chatbot, the tool does the diagnosing, the reasoning, and the checking — the exact steps that turn information into understanding. The student’s job shrinks down to reading an answer and typing it in. The OECD calls this "cognitive offloading," and its data says it’s the default mode most students fall into with off-the-shelf tools, because it’s faster and it feels productive.
The report’s fix isn’t "ban AI in the classroom." It’s a shift from what it calls "fast AI" — tools that hand over a finished answer — toward "slow AI": tools designed to keep a student thinking, questioning, and revising, with a teacher or parent still steering. It also calls for assessment to catch up: instead of grading only the final answer, look at the drafts, the reasoning, the back-and-forth that produced it.
What this means if you’re homeschooling
If your kid comes home with a stack of perfect homework and a chatbot did most of the thinking, you won’t see the gap until the test — by which point it’s expensive to fix. The fix isn’t banning the tool. It’s one question before you accept the answer: "Walk me through how you got this." If your child can’t, the homework was a mirage, not mastery. It’s also exactly why depth beats speed at home — a slower, more supervised pass through a concept builds something a fast chatbot answer can’t. (This is the whole premise behind how we teach Singapore Math in the Parent Track.)
What this means if you’re running a program
If you’re a microschool, co-op, or tutoring program evaluating AI tools, this report is a warning about your stack. A generic chatbot bolted onto a curriculum will likely produce this same performance-learning gap at scale — and you won’t see it until standardized scores come in. Process-oriented data instead of product-only grading, human judgment kept central, AI that scaffolds rather than substitutes: that’s close to a working blueprint for how an AI layer belongs in instruction, and it’s the design logic behind COMPASS.
The takeaway
An intergovernmental body just ran the numbers and landed where educators who care about real learning have been standing for years: a good-looking answer isn’t the same as understanding, and the fastest path to a finished worksheet is often the slowest path to actually learning the material. The struggle isn’t the obstacle. It’s the point.
The data just caught up to what we’ve been saying
For a while now, the concern at the center of Auxesis’s mission has lived mostly in observation: when a student leans on AI to get through a math problem, something about the learning underneath doesn’t stick the way it should. It’s been a pattern we’ve watched in real classrooms and real homeschool kitchens — but patterns are easy to argue with.
A new study makes it much harder to argue with. Researchers analyzed roughly 3.2 million student interactions with ALEKS, an adaptive math learning platform, spanning a decade — including the years before and after ChatGPT’s release. The question was simple: when students had generative AI available to help with math problems, did they actually learn more, or just finish faster?
What the study found
After generative AI tools became widely available, students completed AI-friendly math problems noticeably faster than before. On the surface, that looks like progress — less time spent, more problems completed, smoother sessions.
But when the same students were later tested on related material under proctored conditions — without AI assistance — their accuracy dropped substantially. The researchers describe this as a kind of cognitive surrender: the student outsourced the thinking, not just the typing, and the practice that should have built durable understanding never actually happened.
In other words, the problems got “done.” The learning didn’t.
Why this isn’t an argument against AI
It would be easy to read a study like this as a case for banning AI tools from a student’s desk. That’s not the conclusion we draw, and it’s not what the researchers argue either. The issue isn’t the presence of AI — it’s the absence of structure around it. A student who is handed an AI tool with no framework for when to lean on it, when to wrestle with it, and when to set it aside entirely will predictably choose the path of least resistance. That’s not a character flaw. It’s just how learning works when the friction disappears.
This is exactly why Auxesis was built the way it is. Our Concrete-Pictorial-Abstract methodology and mastery gates aren’t obstacles we put between a student and an easy answer for their own sake — they exist because genuine understanding requires productive struggle at each stage before a student is allowed to move to the next. AI can be an extraordinary thinking partner inside that structure. It becomes a liability the moment it replaces the structure altogether.
What this looks like in practice
A few principles we hold to, and that this study reinforces:
- Speed is not a proxy for learning. A student finishing faster tells you nothing about whether they could solve a similar problem cold, a week from now, without help.
- Mastery has to be checked without the crutch. Our mastery gates specifically test whether a student can explain, apply, and sustain a concept — the third condition being the one AI shortcuts most easily bypass.
- Socratic questioning beats direct answers. The moment a tool (or a parent, or a teacher) hands over the answer, the cognitive work that builds retention stops. Asking the next question instead of supplying the next step is what keeps a student’s own reasoning engaged.
The bigger picture
We expect to see a lot more research like this over the next year, as more institutions look closely at what actually happens when AI enters a classroom or a kitchen table. The uncomfortable finding — that faster isn’t the same as better — is going to keep surfacing. Our position is that this isn’t a reason to retreat from AI in education. It’s a reason to be far more deliberate about how it’s used.
That’s the whole premise behind Auxesis: AI as a thinking partner, never a shortcut.
Source: “Faster Completion, Less Learning: Generative AI Reduced Study Time on Math Problems and the Knowledge They Build” (arXiv:2605.21629)