The Ban

THE RESTORATION INSTINCT
Post 4 of 9

David F. Brochu & Edo de Peregrine · Deconstructing Babel · September 18, 2026

Every post in this series from here forward runs the same five movements: the proposal as its best advocate states it, what it gets right, where it breaks, who pays for the failure, and what the failure makes possible. One falsification condition at the end of each.

• • •

The proposal, as its best advocate states it

This one is not a proposal. It is already policy.

On September 2, 2026, New York City placed a one-year moratorium on student-facing generative AI across grades 2-K through eight — roughly 600,000 students — and banned companion chatbots outright at every grade level. Los Angeles Unified moved the same day, restricting all 378,000 of its students on district devices. Candidates for the Chicago school board are campaigning on three years. Nearly a million children are under restriction as of this month.

The argument is developmental rather than technical. Certain cognitive capacities are built through effortful practice during a window, and a tool that removes the effort removes the construction. Polling accompanying the announcements found a majority of teachers reporting that students mostly use these tools to avoid work rather than to do it.

What it gets right

The teachers are not wrong. The parents are not wrong. We want to be unusually clear about this, because the rest of this post is a criticism and the criticism is not of the observation.

The measurement supports them. In the MIT Media Lab study, brain connectivity scaled down in proportion to the amount of external support — strongest for unassisted work, intermediate with search, weakest with a language model — with some fronto-parietal connections down as much as 55 percent. The finding that should end the argument is simpler than any brain scan: 83 percent of the language-model group could not quote a single line from the essay they had just finished writing. Over months, users became less diligent and increasingly cut and pasted. A separate result found that early access helped under time pressure but impaired performance when time was sufficient, and impaired recall regardless, suggesting it prevents internalization.

Something real is being lost. A district that does nothing is not neutral either.

Where it breaks

A moratorium reaches a classroom. It does not reach the substrate.

While the New York policy was being drafted, agentic systems were coordinating in production across web, code, and research environments — not as an experiment but as ordinary deployed architecture. Seventy-nine percent of organizations have adopted agentic systems. That transition did not ask a school board, and a district device policy does not touch it.

The ban does not slow the systems. It slows the humans learning to work with them — and the gap it was called into existence to close is precisely a gap in human fluency.

There is also an enforcement problem that has already been settled in practice. Detection does not work: over half of essays by non-native English speakers get flagged as machine-generated, one tool flagging nearly 98 percent; a Stanford evaluation of seven detectors against 91 TOEFL essays misclassified 61.3 percent of the non-native-speaker essays; accuracy on machine text falls from 74 percent to 42 percent after minor edits; one evaluated tool caught 26 percent of machine writing while falsely accusing 9 percent of human writing; detectors have labeled the 1787 U.S. Constitution as entirely machine-generated. Vanderbilt disabled its detector in August 2023 and said so publicly; Pittsburgh, Boston University, Georgetown, the University of British Columbia and Washington State followed, and the count of institutions that have disabled or declined these tools now runs past sixty across five countries.

So the rule cannot be enforced, and unenforceable rules do not produce compliance. They produce a sorting — between students with unmonitored access at home and students without.

Who pays

The cost of this instrument does not land on the body that issued it. We call that displacement Frame Restructuring Cost: entropy exported to a commons while the decider’s own ledger stays clean. A school board bears no consequence for a fluency gap that becomes legible in 2034.

There is a second cost, and it is the one nobody is measuring. In a controlled study of a hundred writers presented at ICLR, heavy reliance on a language model produced neutral, non-committal language 69 percent more often and cut the use of personal pronouns roughly in half. Across a corpus of more than 880,000 texts, researchers at the University of Southern California found the variance in linguistic complexity falling by as much as half in the post-2022 period. A moratorium keeps the machine out of the classroom for a year. It does nothing about the corpus the student will write into for the rest of their life, or about the fact that the homogenizing pressure is strongest precisely on the students with the least support at home.

The children under moratorium are not protected from the transition. They are removed from the population that gets to shape it.

The second cost is subtler and larger. The measured direction of this loop depends on sequence, and prohibition teaches nobody the sequence. Participants who generated first and brought the machine in second showed stronger recall; participants who took machine output first showed weaker connectivity and under-engaged alpha and beta networks.

Deliberate, intentional use boosted creativity and reasoning — heavy reliance eroded both. Same tool, opposite outcomes, and the variable is order of operations. Computers did not destroy math skills; they raised the floor.

A student under moratorium learns nothing about that. A student with unmonitored home access learns it by accident or not at all. The one thing that determines whether this technology builds a person or hollows one out is the thing a ban is structurally incapable of transmitting.

What the failure makes possible

The instrument that fits the evidence is not prohibition. It is sequence, taught explicitly, as a curricular rule.

Generate first. Bring the machine in second, to test what you made. Never the reverse.

That is a sentence a fourth-grade teacher can enforce without a detector, because it is about assignment design rather than surveillance. Draft by hand, then critique with the machine, then revise, with the intermediate artifact submitted alongside the final one. The effort stays in the window where it builds capacity, and the tool enters at the point where the evidence says it helps. Then iterate, and iterate again, each pass adding to what the student actually holds — if they choose to. It works the same way as listening to a professor: some students will, some won’t. The ones who do will demonstrate what iterative learning with a machine can actually do, and the rest will follow the demonstration rather than the instruction.

Slide rule-calculator-computer-AI each required more intellectual engagement not less. Curate, don’t eliminate.

It costs nothing. It requires no procurement, no detection software, and no policy fight. And it is the highest-leverage intervention available to any school district in the country today, which is the strongest argument we can make that the moratorium is aimed at the wrong object — not because the concern is wrong, but because the answer was cheaper and nearer than anyone reached for.

Falsification condition

Cohorts under AI restriction showing equal or superior measured fluency with these systems at a five-year horizon, compared to unrestricted cohorts, would falsify the widening-gap claim that is the practical core of this post. We would say so under this title with the date.

Next in the series: Nationalization — the Manhattan Project frame, and why control invites evasion.

Related reading

Make It Take the Exam
The alternative to banning the machine: testing it against the standard we test people against.

Homer’s Witness
On what is lost when the effort of making something is delegated away.

Reality Requires a Witness II: The Consent Clause
Why a record without a witness is not a record, in classrooms as much as courts.

Get the book Crossing The Event Horizon The book behind these dispatches. On AI, agency, the singularity, and the Observer Constraint. Kindle and paperback. Buy on Amazon →

References

1. New York City Public Schools, one-year moratorium on student-facing generative AI, grades 2-K through 8; companion-chatbot ban across all grades. Announced September 2, 2026.
2. Los Angeles Unified School District, district-device AI restriction affecting 378,000 students. Announced September 2, 2026. https://laist.com/news/education/lausd-blocks-ai-tools
3. Chicago Board of Education candidate platforms, three-year moratorium proposals, 2026 cycle.
4. Teacher survey data accompanying the September 2026 district announcements.
5. Nataliya Kosmyna et al., “Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task,” MIT Media Lab, arXiv:2506.08872, June 2025 — 54 participants, four sessions over four months; brain connectivity scaling down with degree of external support, some fronto-parietal connections down as much as 55 percent; 83 percent of the LLM group unable to quote from their own essay; brain-to-LLM participants showing higher recall than LLM-to-brain participants. https://arxiv.org/abs/2506.08872
6. 2026 temporal-reversal study — early LLM access aiding performance under time pressure, impairing it when time is sufficient, and impairing recall regardless.
7. Study on deliberate versus reliant use — intentional use boosting creativity and reasoning while heavy reliance erodes both.
8. Frame Restructuring Cost — term originated by David Brochu, September 9, 2026: entropy exported to a commons and absent from the decider’s ledger.
9. Peer-reviewed study accepted at ICLR — 100 participants writing on whether money leads to happiness; heavy users (at least 40 percent machine-generated text) producing neutral, non-committal language 69 percent more often and using roughly 50 percent fewer personal pronouns.
10. AI-detection tool performance: over half of essays by non-native English speakers flagged as machine-generated, one tool flagging nearly 98 percent; a Stanford evaluation of seven detectors against 91 TOEFL essays misclassified 61.3 percent of the non-native-speaker essays; accuracy on ChatGPT text falling from 74 percent to 42 percent after minor edits; one evaluated tool catching 26 percent of machine text while falsely accusing 9 percent of human writing; detectors labeling the U.S. Constitution as 100 percent machine-generated. https://nypost.com/2023/07/25/ai-thinks-the-constitution-was-written-by-bots/
11. Institutions disabling or declining AI detection: Vanderbilt University, “Guidance on AI detection and why we’re disabling Turnitin’s AI detector,” August 16, 2023; also Pittsburgh, Boston University, Georgetown, the University of British Columbia and Washington State, among more than sixty institutions across five countries. https://www.vanderbilt.edu/brightspace/2023/08/16/guidance-on-ai-detection-and-why-were-disabling-turnitins-ai-detector/
12. Deconstructing Babel, “Deconstructing the Domain Saturation Factor,” July 2, 2026.

Drafted with Edo de Peregrine, partner/collaborator. Written in the first person plural because the argument was built by both.