THE REVIEW
“When nobody knows how anything actually works, everything starts to look like magic—or fraud”
The Invisible Machines Are Broken
This week: What happens when no one can tell if the instructions actually work
We live, dear reader, in an age of marvels. Your phone can summon a car. An algorithm can write your code. A chatbot can compose your child’s birthday dinner menu. Thomas Edison once predicted motion pictures would make textbooks obsolete; we are still waiting, but this time it surely feels different.
And yet something peculiar is happening across seemingly unrelated domains: developers who feel 20 percent faster are measuring 19 percent slower. Recipe bloggers are watching AI systems confidently instruct home cooks to wash poultry with soap. Farmers who own their tractors cannot repair them. Teachers are spending 12 minutes daily on phonemic awareness exercises that research suggests may not improve reading. In each case, the same pattern emerges: the systems that once verified whether instructions actually worked have been quietly replaced by systems optimized for something else entirely.
This edition of The Review examines what we’re calling the verification collapse—the gradual disappearance of the invisible infrastructure that once connected instruction to outcome. We’ve assembled five dispatches from the front lines: from code editors where productivity metrics have become dissociated from productivity, from kitchens where AI-generated “Frankenstein recipes” are ruining Thanksgiving dinners, from farms where software now enforces what lawyers could not, from classrooms where curriculum fads outpace research, and from the revisionist historians who are reconstructing what the forgetting machines have erased.
The connecting thread is uncomfortable: we have become very good at making things look verified—complete with citations, step numbers, and professional formatting—while severing the connection to whether any of it actually works in the physical world. The 40-point gap between perception and reality in one study may be the defining number of our moment.
We do not offer solutions, mostly because solutions require admitting the problem exists. We offer instead the news, such as it is, with the hope that naming the pattern is the first step toward recognizing it in the wild.
❧ ❧ ❧
Developers Feel Fast, Work Slow, Keep Coding Anyway
A landmark study finds programmers are 19% slower with AI tools—but won’t stop using them
The number that should haunt Silicon Valley is 40—as in the 40-point gap between how fast software developers feel they’re working and how fast they actually are.
A randomized controlled trial published by the AI research organization METR in late 2025 delivered a verdict that should have stopped the industry cold: experienced open-source developers estimated their AI coding assistants made them 20 percent faster. When researchers actually measured their task completion times, the developers were 19 percent slower.
That’s not a rounding error. That’s a perception-reality gap wide enough to drive a business model through—and indeed, several companies already have.
“I kept feeling like the AI was really dumb, but maybe I could trick it into being smart if I found the right magic incantation.” — Mike Judge, software developer at Substantial
The implications ripple outward. If roughly 41 percent of code in production workflows is now AI-generated, according to 2025 industry surveys, and only 3 percent of developers report they “highly trust” AI output while 46 percent say they don’t fully trust it, we have built an industrial process on the digital equivalent of a coin flip.
Source: METR randomized controlled trial, 2025
The METR study adds a twist that should concern anyone with equity in AI companies: after being told they were slower, 69 percent of participants continued using the tools anyway. The feeling of productivity persists even when contradicted by data. This is not technology adoption; it is something closer to habit formation.
The skeptics have data now. A September 2025 report from Bain & Company described real-world productivity gains from AI coding tools as “unremarkable.” GitClear, which analyzes millions of pull requests, found developers producing roughly 10 percent more “durable code”—code that isn’t deleted or rewritten within weeks—since 2022. Ten percent is helpful. It is not transformational.
The boosters have data too. A Jellyfish analysis of companies that went from zero to full AI tool adoption showed median cycle times dropping 24 percent. GitHub reports 81 percent of Copilot users say it helps them complete tasks faster.
The reconciliation is uncomfortable: AI coding tools feel like magic and are useful for specific tasks—boilerplate, documentation, debugging explanations—while potentially degrading performance on complex, novel, or safety-critical work. The platforms cannot tell the difference. Neither, it appears, can many of the developers using them.
The Stack Overflow 2025 Developer Survey captures the mood shift: positive sentiment toward AI tools dropped from above 70 percent in 2023-2024 to 60 percent in 2025. The honeymoon is ending. The addiction, however, continues.
For Further Reading: Perspectives
PRO
“Generative Coding: 10 Breakthrough Technologies 2026” — MIT Technology Review makes the case that AI coding assistants represent genuine transformation, despite limitations.
technologyreview.com (January 2026)
CON
“The Productivity Paradox of AI Coding Assistants” — Cerbos analyzes why dopamine rewards activity in the editor, not working code in production.
cerbos.dev (September 2025)
❧ ❧ ❧
Google’s AI Told Someone to Wash Their Chicken With Soap
The holiday recipe crisis reveals what happens when instructions are severed from kitchens
Somewhere in America this past Thanksgiving, a home cook followed Google’s AI-generated advice and washed raw poultry with dish soap. We know this because recipe bloggers have been cataloging the damage.
The incident is not anomalous. It is the new normal. Google’s AI Overviews, powered by Gemini, now serve synthesized cooking instructions to roughly 1.5 billion users monthly. The system works by combining ingredients from one recipe blog with directions from another, creating what food creators have come to call “Frankenstein recipes.”
The results read like a culinary uncanny valley. One AI summary instructed bakers to cook a six-inch Christmas cake for three to four hours. “You’d end up with charcoal,” observed Eb Gargano, whose legitimate recipe traffic has cratered by 40 percent year over year.
“For websites that depend on the advertising model, I think this is an extinction event in many ways.” — Matt Rodbard, founder and editor-in-chief of Taste
The arithmetic is grim. Twenty-two independent food creators interviewed by Fortune reported traffic declines ranging from 30 to 80 percent from Google search.
Source: Fortune interviews with food creators, November 2025
The mechanism is instructive. Recipe development is not just writing; it is testing. A professional recipe typically requires multiple iterations—adjusting oven temperatures, ingredient ratios, resting times—until the dish reliably works. This labor is invisible in the final product, which looks identical whether or not anyone has actually cooked it.
AI systems, trained on the surface features of recipes, have no way to distinguish between tested and untested instructions. They can tell that a recipe looks like a recipe. They cannot tell whether the recipe works.
Google has responded that AI Overviews are “a helpful starting point” and that users still click through to original sources. The bloggers observe the opposite. When a complete recipe appears in search results, there is no reason to visit the source—and no way for the source to maintain the testing pipeline that made the recipe trustworthy in the first place.
The food blogging business model—free recipes supported by ads—now feeds the machine that will replace it. But the person following untested instructions is not the blogger. It is the home cook serving Christmas dinner.
For Further Reading: Perspectives
PRO
“AI slop recipes are taking over the internet” — Fortune provides comprehensive reporting on how 22 food creators are experiencing the transformation.
fortune.com (November 2025)
CON
“Google AI Summaries Are Ruining the Livelihoods of Recipe Writers” — Slashdot commenters debate whether bloggers’ own SEO practices contributed to the current crisis.
slashdot.org (December 2025)
❧ ❧ ❧
Your Tractor Knows You Fixed It, and It’s Not Happy
How John Deere’s software turned repair manuals into compliance theater
The repair manual exists. You can download it. You can read every word. You can follow every instruction perfectly. And when you’re done, your $500,000 combine will sit in your field, refusing to harvest, because the software doesn’t recognize your repair.
This is the state of agricultural repair in 2026, and a federal judge has had enough movie references to describe it.
“Sequels so rarely beat their originals that even the acclaimed Steve Martin couldn’t do it on three tries,” U.S. District Judge Iain D. Johnston wrote in June 2025, denying John Deere’s motion to dismiss the Federal Trade Commission’s right-to-repair lawsuit.
“In technologizing its equipment, Deere makes farmers reliant on Deere’s own ADVISOR software. And, in only licensing that software to its authorized Dealers, Deere forces farmers to visit those shops instead of using closer, cheaper options.” — U.S. District Judge Iain D. Johnston
The FTC’s lawsuit, filed in January 2025, alleges Deere monopolized the repair services market, violated antitrust laws, and burdened farmers with billions in excess costs. A 2023 U.S. PIRG report estimated the annual toll: 1.2 billion in inflated repair bills.
Source: U.S. PIRG 2023 report, court filings
Modern John Deere tractors contain engine control units (ECUs) that manage everything from fuel injection to GPS navigation. When a component fails, the ECU may enter “limp mode”—reduced functionality until a dealer clears the error code. Farmers can physically install a replacement part, but without ADVISOR software to authenticate it, the tractor may refuse to recognize the repair.
Deere’s documentation, meanwhile, has achieved a peculiar compliance-without-function. U.S. PIRG reported that farmer-level software displays redacted troubleshooting links, partial guides, and blocked payload files. The manual exists. It tells you how to repair the tractor. It just doesn’t tell you how to make the tractor accept the repair.
This is what malicious compliance looks like when written in code.
The repair manual, it turns out, was never the problem. The problem is who gets to decide whether repair counts as repair.
For Further Reading: Perspectives
PRO
“Right to Repair Gains Traction as John Deere Faces Trial” — Barn Raiser examines the FTC case and its implications for agricultural monopoly.
barnraisingmedia.com (July 2025)
CON
“Opening the Door to Right-to-Repair: Deere & Co. and Its Implications” — University of Chicago Business Law Review analyzes the legal theory and precedent concerns.
businesslawreview.uchicago.edu (2025)
❧ ❧ ❧
The Reading Wars Found a New Battlefield: The Ears
A curriculum that dominated 70% of U.S. schools may not improve reading—but the trains have left the station
If you teach first grade in America, you probably spend 12 minutes each day asking children to identify sounds in words without showing them any letters. This is called phonemic awareness instruction, and it may be one of the most successful educational fads of the decade.
The market leader, Heggerty, claims its programs are used in 70 percent of U.S. school districts. In Ohio, Heggerty commands 50 percent market share—a penetration rate that would be the envy of any consumer brand. Teachers love it. Children clap along. The lessons are scripted and easy to deliver.
There is just one problem: a growing body of research suggests oral-only phonemic awareness instruction doesn’t improve reading outcomes.
A November 2025 study from the University of Connecticut tested Heggerty in 13 elementary schools with 782 first-graders. The program improved students’ phonemic awareness skills—their ability to identify and manipulate sounds—with a moderate-to-large effect size. But students who received Heggerty didn’t do significantly better than control groups on measures of word-reading and oral reading fluency.
They got better at the thing the program teaches. They didn’t get better at reading.
“If you teach phonemic awareness, students will learn phonemic awareness, which isn’t the goal. If you teach blending and segmenting using letters, students are learning to read and spell.” — Tiffany Peltier, literacy consultant at NWEA
Source: University of Connecticut study, November 2025; Evidence for ESSA
The finding is consistent with four meta-analyses over the past 25 years: combining auditory instruction with visible letters works better than auditory-only exercises.
Heggerty has quietly revised its approach. The company acknowledged that “the science of reading has evolved,” which is why it revised its program in 2022 to incorporate letters. But many schools cannot afford new curricula, and the company still recommends 8–12 minutes daily through the end of first grade—more than many researchers suggest.
The momentum problem is real. The pattern is familiar: a plausible instructional idea achieves rapid scale on teacher enthusiasm and vendor marketing, outruns the research, and becomes institutionally embedded before evidence can catch up. By the time the studies are in, the trains have left the station.
For Further Reading: Perspectives
PRO
“A Popular Method for Teaching Phonemic Awareness Doesn’t Boost Reading” — Education Week reports the University of Connecticut findings and industry response.
edweek.org (November 2025)
CON
“Educators Were Sold a Story About Phonemic Awareness” — Curriculum Insight Project argues reviewers failed to warn teachers about oral-only limitations.
curriculuminsightproject.substack.com (February 2026)
❧ ❧ ❧
The Recipes They Refused to Forget
How one culinary historian is rebuilding the instructions that erasure tried to delete
“Food is an archiver, a keeper of secrets.” So writes Michael W. Twitty in The Cooking Gene, his James Beard Award-winning exploration of African American culinary history. What he doesn’t say—but what his work demonstrates—is that archives require archivists, and secrets require people who refuse to forget.
In October 2025, Twitty published Recipes from the American South, a 432-page cookbook that is also something more: a counter-archive, a reconstruction of culinary knowledge that erasure tried to delete. Each recipe arrives with context—genealogy, oral history, agricultural history, technique—that transforms a list of ingredients into what Twitty calls a recipe with “memory.”
The timing feels deliberate. As AI systems vacuum up the world’s recipes to produce Frankenstein mashups, Twitty offers the opposite: instructions so deeply embedded in human experience that they resist abstraction. You cannot remix a recipe that is also a story about your great-grandmother’s survival.
“I want to restore cooking to its rightful heritage, whatever it is.” — Michael W. Twitty
Two philosophies of preserving culinary knowledge
Twitty’s method is laborious. He traces family lines back to enslaved ancestors in specific plantation kitchens. He recreates historical conditions—picking cotton, chopping wood, cooking over open fires—to understand the constraints that shaped technique. He consults genetic tests, historical documents, and seed catalogs. He maintains heirloom seeds from the African American Heritage Collection he helped compile for the D. Landreth Seed Company.
The result is not just recipes but a methodology for culinary justice—what Twitty defines as the project of honoring the food past while providing for the food future.
This is verification of a different kind. Where AI systems collapse context to extract replicable patterns, Twitty expands context to reveal unreplicable specificity. The recipe for hoecakes is not just cornmeal, water, and fat; it is the economic reality of what enslaved people could grow, the thermodynamics of cooking on a hoe over coals, the cultural memory of West African grain preparation.
In an age when instructions are increasingly severed from the people who created them, Twitty’s work offers a model: verification through memory, context as resistance, recipes that cannot be scraped because they were never merely text.
For Further Reading: Perspectives
PRO
“Michael Twitty at Dayton Food and Culture Event” — Edible Ohio Valley profiles Twitty’s approach to culinary justice and identity.
edibleohiovalley.com (January 2026)
CON
“Recipes from the American South” — NPR coverage of Twitty’s new cookbook and debates about Southern food ownership.
npr.org (October 2025)
❧ ❧ ❧
EDITORIAL
The 40-Point Gap Is Everywhere Now
We have buried the lede across five stories, so let us state it plainly: the 40-point gap between perception and reality in the METR coding study is not an anomaly. It is the signature of our moment.
Developers feel 20 percent faster; they are 19 percent slower. Recipe bloggers spend years testing; AI serves untested instructions to 1.5 billion users. Farmers buy repair manuals that omit the one piece of information required to make repair possible. Teachers spend 10 percent of their literacy block on exercises that improve a proxy metric but not reading. In each case, the feeling of verification persists while the substance of verification has been hollowed out.
This is not a technology problem. It is not a policy problem. It is an epistemology problem—a collapse in our collective ability to distinguish between things that look verified and things that are verified.
The pattern across this edition’s five stories
The mechanisms differ: AI confidence scores are not the same as corporate repair restrictions, which are not the same as curriculum marketing, which are not the same as systematic historical erasure. But the pattern is consistent: systems that once connected instruction to outcome have been replaced by systems optimized for something else—engagement, compliance, profit, institutional momentum—without anyone quite noticing the substitution.
The telling detail from the METR study is that 69 percent of developers kept using AI tools even after being told the tools made them slower. The feeling was stronger than the evidence. We suspect this is not unique to developers.
What would it mean to rebuild verification? We don’t know, exactly, but the examples in this issue suggest directions. Twitty’s methodology—recipes embedded in genealogy, technique rooted in physical constraints, context that resists abstraction—offers one model. Right-to-repair advocates demanding that diagnostic software be opened offer another. Literacy researchers calling for letters to be shown alongside sounds offer a third.
Each involves making the invisible visible: the testing that precedes a trustworthy recipe, the software that gatekeeps a physical repair, the difference between learning a proxy skill and learning to read.
Thomas Edison predicted in 1913 that motion pictures would replace textbooks. Radio was supposed to revolutionize education in the 1930s. Television in the 1950s. Computer-assisted instruction in the 1960s. The Internet in the 1990s. MOOCs in the 2010s. AI now.
The pattern is not that these technologies failed. The pattern is that we forgot, each time, that technology does not verify itself.
Verification is infrastructure. Like plumbing, it is invisible when it works and catastrophic when it fails. Like plumbing, someone has to maintain it. Like plumbing, when the maintenance stops, the failures compound silently until something obvious breaks.
The 40-point gap is not sustainable. But sustainability has never been the metric.
For Further Reading: Perspectives
PRO
“Why Measuring Impact Effectively Is So Important in EdTech” — World Economic Forum argues for universal monitoring standards across educational technology.
weforum.org (April 2024)
CON
“Ed-Tech in 2026: Only for Profit and Misuse, or There is Hope?” — Gaurav Tiwari examines the structural incentives pushing EdTech away from evidence.
gauravtiwari.org (December 2025)
Production Note
This edition of The Review was produced through collaboration between human editorial judgment and AI research assistance. The facts reported here were verified against primary sources where possible; the interpretations are the editor’s own. We have attempted to represent multiple perspectives fairly, including perspectives critical of our framing. Your skepticism remains appropriate and encouraged. If you find errors, please let us know.
Coming Next
The Apprenticeship Gap—examining why video tutorials feel like learning but may not produce competence. Also: the Content Authenticity Initiative and whether cryptographic provenance can rebuild trust.
Subscription Notice
This publication is not available to the general public. Distributed by subscription only to readers vouched for by existing members.
If you know, you know.
© 2026 The Review. All rights reserved.
Editor: Daniel Markham | Submissions: submissions@thereview.pub
Wednesday, February 5, 2026