Latest News and Comment from Education

Show All

Saturday, August 22, 2026

THE HOUSE THAT AI BUILT: GRADING ITS OWN HOMEWORK

 

THE HOUSE THAT AI BUILT: GRADING ITS OWN HOMEWORK

On August 20, the U.S. Department of Education sent school districts a six-page love letter to the algorithm. Assistant Secretary Kirsten Baesler's Dear Colleague Letter draws a bright line between "recreational" screen time — the TikTok-scrolling, dopamine-farming kind everyone from state legislatures to parent Facebook groups has spent 2026 trying to ban — and "instructional" screen time, the kind delivered by an adaptive AI tutor with a chipper avatar and a Series B round behind it. One is poison. The other, we're told, is medicine. Same delivery device. Different lawyer.

It's an elegant distinction, and like most elegant distinctions in education policy, it happens to be extremely convenient for somebody.

Same Playbook, New Chapter

Here's the part nobody in the press release wants to dwell on: we have done this exact rodeo before, with the exact same animal. Two weeks before Baesler's letter dropped, an eight-person jury in Oakland began hearing 29 states argue that Meta spent a decade engineering infinite scroll, push notifications, and recommendation loops specifically to hijack the developing adolescent brain — and that Meta knew it, had the internal research to prove it, and buried it anyway. The state's opening line, roughly: hook them, hold them, harvest them, then lie about it to Congress.

Now hold that thought next to the second half of the Baesler letter, the part about classroom AI. It cites glowing effect sizes from randomized trials on "domain-specific AI tutoring systems." Great news. Who ran those trials? Follow the money and you land on a tidy little ouroboros: federal science agencies, yes, but also the Gates Foundation, the Chan Zuckerberg Initiative, Schmidt Sciences, and — sitting right there in the acknowledgments next to Khan Academy and OpenAI — the AI labs themselves, funding "independent" efficacy research on the products they sell. It's not illegal. It's not even unusual. It's just the exact incestuous funding structure that let the ed-tech industry mark its own homework for fifteen years running, except now the homework is graded by a chatbot instead of a screen-time app, and the grader has a nine-figure compute budget.

Nobody is accusing Kirsten Baesler of moonlighting for Anthropic. The point is structural, not personal: when the entities with the deepest pockets to fund "rigorous, independent" studies are also the entities whose valuations depend on the studies coming back positive, "independent" starts doing a lot of load-bearing work for one word.

The Social Media Precedent, Rebranded

The Oakland trial matters here for a reason beyond schadenfreude. It establishes, in a courtroom under oath, that a platform's own internal research showed harm to kids years before the public — or the regulators — found out. That's the pattern the Department of Education's new guidance should be most worried about repeating, and instead it reads like a permission slip to speed-run it again with a shinier acronym.

Because the honest version of the "AI vs. social media" comparison isn't flattering to either side. Both arrived in classrooms and bedrooms faster than the research could keep up. Both were sold on projected outcomes rather than demonstrated ones. Both had enthusiastic executive-branch cheerleading — recall the President's AI-in-education executive order, and Melania Trump's State of the Union guest slot for a student from an AI-powered micro-school network. And both, when the receipts eventually surfaced, revealed that the people building the technology had a much clearer picture of its downsides than the parents signing the permission forms.

The DOE letter's own citations gesture at this without quite saying it out loud: unstructured generative AI use produces measurable "cognitive offloading," where kids stop doing the thinking and just take the answer. That's not a hypothetical risk cooked up by anxious school boards — it's in the research the guidance itself leans on. The letter's solution is to wave that risk through with a "human-in-the-loop" clause and move on to the procurement rubric, the same way a decade of "just add parental controls" clauses waved through the harms Oakland is currently litigating.

Outcomes-Based Everything, Except Honesty

To be fair to the guidance — and a fair accounting requires being fair — the Outcomes-Based Contracting model it champions is a genuine improvement over the old regime of flat annual licenses and vibes-based renewal. Tying 40% of a vendor's payment to verified student growth, building in exit clauses for tools that flatline, demanding ESSA Tier 1/2 evidence instead of a glossy case study PDF — districts that adopt this seriously will waste less money on dead software. Texas and Louisiana's early rate-card experiments are a real, and reasonably sober, template.

The trouble is that "evidence-based" is only as good as who's allowed to define the evidence, and right now the same small cluster of foundations and labs sits on both sides of that ledger — funding the research, building the products, and, increasingly, writing the policy memos that decide what counts as proof. That's not a conspiracy. It's just an incentive structure, and incentive structures don't need anyone to be lying for them to produce a predictable result.

So when the letter tells districts to stop counting minutes and start counting outcomes, that's the right instinct pointed at the wrong finish line. The actual question isn't "does this AI tutor move a test score." It's "who paid to find out, who profits from the answer, and will we still trust that answer in five years the way we used to trust the research that told us social media was basically fine." Oakland's jury is currently finding out how that story ends the last time we asked it. The Department of Education might want to read the transcript before it prints the next permission slip.

— Big Education Ape




Sources — "The House That AI Built"

The DOE Dear Colleague Letter (Aug. 20, 2026)

The Meta Trial (Oakland, N.D. Cal., opened Aug. 18, 2026)

Trump Administration AI-in-Education Push

Note on the funding/research material

The claims about specific funders (NSF, IES, Gates Foundation, Spencer Foundation, Schmidt Sciences, CZI, OpenAI/Anthropic, Khan Academy), the Outcomes-Based Contracting rollouts in Texas (Ector County ISD) and Louisiana (Act 771), and the cited effect-size/"cognitive offloading" research findings were drawn from the briefing document you provided rather than independently re-verified against primary sources in this pass. If you want those specific figures and grant numbers checked and re-sourced to primary pages (NSF award pages, Gates Foundation grant database, the original Scientific Reports study, etc.) before publication, I can run that pass separately.


Posted by at