Welcome to the Great Stasis

Dear reader, you hold in your hands—or more likely, scroll upon your screen—something peculiar: a newspaper about why everything has stopped moving.

Not stopped in the dramatic sense. No crashes, no explosions, no singular catastrophe you could pin to a calendar. Something stranger: a creeping paralysis across the institutions that promised to liberate us through technology. The online courses that were supposed to democratize education? Ninety percent abandon them. The job markets that were supposed to reward talent? Frozen, with workers trapped and employers refusing to hire. The AI systems that were supposed to get smarter? Eating their own outputs and degrading into confident nonsense.

These are not isolated failures. They share a common cause: the cost of generating anything has collapsed to near-zero, while the cost of verifying anything has become prohibitive.

When you can't tell if a resume was written by the applicant or their chatbot, you stop hiring. When you can't tell if a student learned the material or just prompted their way to an A, credentials lose meaning. When you can't tell if training data is human insight or machine regurgitation, your AI models degrade. The rational response to a verification vacuum is paralysis.

This edition examines five interconnected domains caught in the stasis. We close with an editorial asking the only question that matters: Is this paralysis temporary, or is it the new normal?

When AI Eats AI: The Photocopier Problem Nobody Can Fix

Recursive training degrades models into "confident nonsense," sparking a scramble for data as precious as pre-nuclear steel

The machines have started eating themselves, and the results are exactly as digestible as you'd expect.

Researchers call it "model collapse"—artificial intelligence trained on content generated by previous AI gradually degenerates, losing the ability to produce diverse or coherent outputs. A Nature study demonstrated the mechanism: each generation trained on its predecessor is like a photocopy of a photocopy, losing fidelity until the original is unrecognizable. The difference is that a photocopier doesn't claim its degraded copy is superior.

The contamination is real. By mid-2025, over 74 percent of newly created webpages contained AI-generated text. In Google's top results, AI-written pages climbed from 11 percent to nearly 20 percent in a single year. The machines are citing machines, recursively, like a hall of mirrors where nobody remembers which reflection was the original.

"The internet is just a wire in the ocean, and now it's full of ghosts talking to ghosts." — Anonymous data scientist on model collapse

This has inverted traditional data economics. In the pre-AI era, "fresh" data was prized. Now archival data is the premium asset. Old books. Pre-2022 forum discussions. Scanned physical documents. Anything demonstrably human.

The Harvard Journal of Law & Technology made the parallel explicit: this is the "low-background steel" of the AI age. After atomic testing contaminated all post-1945 steel with radioactive isotopes, scientists who needed uncontaminated metal had to scavenge from sunken battleships. There will never be more pre-nuclear steel. There will also never be more pre-2023 internet.

AI companies have responded with unprecedented licensing deals—hundreds of millions of dollars for access to uncontaminated human data. Reddit's IPO filing revealed content licensing agreements worth $203 million. We have moved from hunter-gatherer data acquisition to a feudal phase, where high-quality data is hoarded behind castle walls.

Eighty Guidelines, Zero Changes: The Theater of AI Ethics

Principles proliferate while algorithms keep optimizing for outrage—and nobody can explain why

Between 2023 and 2025, the world produced over 80 distinct sets of AI ethical guidelines. You'd be hard-pressed to identify a single concrete outcome from any of them.

This is not cynicism but observation. The Business Roundtable's 2019 redefinition of corporate purpose pledged major corporations to serve all stakeholders. They did not explain how. The aspiration was articulated. The operational translation never occurred. Five years later, the gap between stated values and operational reality has become the defining feature of AI governance: elaborate principles that change nothing.

Researchers call this the "operationalization gap." It is relatively easy to achieve consensus that an AI system should be "fair." It is excruciatingly difficult to translate "fairness" into a line of Python code. What does "fair" mean when different definitions—equality of outcome, equality of treatment, equality of opportunity—are mathematically incompatible?

"We've been excellent at writing principles. We've been terrible at writing code that implements them." — AI ethics researcher on the gap between aspiration and operation

The problem becomes especially dangerous when systems determine which ideas reach which audiences. A September 2025 systematic review of 78 empirical studies identified the pattern: algorithms systematically privilege content that maximizes engagement, and divisive content engages more reliably than nuanced content.

Governments attempted "Algorithmic Impact Assessments"—the theory being that organizations would assess risks before deployment. In practice, AIAs have become bureaucratic checkboxes. Checklist-based frameworks can be completed in approximately 1.5 hours. They are efficient, scalable, operationally friendly—and fundamentally shallow.

Free Harvard, 5% Graduation: Online Learning's Broken Promise

Completion rates haven't budged in a decade—the difference isn't content, it's accountability

The promise was irresistible: world-class education, available to anyone with an internet connection, at no cost. A decade later, the data tells a different story.

The median completion rate across massive open online courses sits at 12.6 percent. For participants who don't pay, that number drops to 5 percent. Six years of investment in course development, learning research, and platform improvements have failed to move the needle. The completion rate for MIT and Harvard MOOCs actually fell from 2013 to 2018.

Here's the puzzle that should trouble every education reformer: cohort-based programs using the same lectures and materials achieve 80-96 percent completion rates. The content is identical. The technology is identical. The difference is structure.

"The 6-year saga of MOOCs provides a cautionary tale for AI-based instruction that could repeat the same errors." — Educational technology researcher

Cohort-based programs impose accountability: regular deadlines, peer interaction, instructor presence. Self-paced courses offer freedom—and 95 percent of free participants exercise that freedom by walking away. The lesson is uncomfortable for technologists: the problem isn't the content or the platform. It's that learning requires sustained effort, and sustained effort requires external structure for most people.

Meanwhile, AI has created a credentialing crisis. UK universities caught 7,000 students using AI inappropriately in 2023-24—triple the previous year. Research suggests 94 percent of AI use goes undetected. Students using AI perform 17 percent worse on independent assessments, suggesting the tools enable credential acquisition without corresponding skill acquisition.

Institutions are reverting to handwritten exams, in-person proctoring, oral defenses—the infrastructure of trust that technology was supposed to make obsolete.

No Juniors Allowed: Who Trains Tomorrow's Senior Engineers?

Entry-level positions have vanished—creating a pipeline paradox companies haven't solved

In the language of corporate planning, the labor market has reached "equilibrium." Translation: nobody is hiring, and nobody is firing.

The numbers beneath this equilibrium are stark. Entry-level technical positions—the P1 and P2 rungs on engineering ladders—have declined 73 percent. At major technology companies, entry-level hiring is down more than 50 percent over three years. UK technology graduate roles fell 46 percent in 2024, with projections of another 53 percent decline by 2026.

This isn't a hiring freeze in the traditional sense. Companies are still adding headcount—selectively, at senior levels, for specialized skills. What's disappeared is the training rung: junior analyst, associate developer, entry-level consultant. The positions where new graduates learned how organizations actually work.

"If there are no junior jobs, where do senior workers come from? Nobody has answered this question." — Technology workforce analyst

The pipeline paradox is emerging. Companies want five years of experience for positions that no longer exist at the entry level. They want senior talent while eliminating the apprenticeships that create senior talent. The solution—assuming someone else will do the training—works only until it doesn't.

AI accelerated this dynamic but didn't create it. Companies discovered they could maintain or grow revenue without corresponding headcount growth through "outcome-based delivery" and automation. The efficiency gains are real. The question of who develops the next generation of workers is someone else's problem.

Fifty-five percent of companies that made AI-related layoffs now regret the decisions. Forrester predicts half of those positions will return—but offshore, at lower wages, with the training burden shifted to workers themselves.

31 New Subscribers a Month Just to Stand Still

The creator economy's brutal math—and why the "sovereign creator" is a myth

Every month, a Substack writer earning $50,000 annually faces the same math: 50 percent of their subscribers will churn within the year.

At $8 per month, earning $50,000 requires approximately 900 paying subscribers. Annual churn of 50 percent means 450 will disappear. Replacing them requires adding 31 new paying subscribers every month—just to maintain current income, before any growth.

At a typical 3 percent conversion rate from free to paid, those 31 paying subscribers require 1,000 new free subscribers per month. Every month. Forever. Not to grow. To stand still.

This is the treadmill that creator economy rhetoric obscures. The pitch is independence: escape corporate employment, own your audience, capture the full value of your work. The reality is revealed in surveys where creators describe their experience in a single word: "Fatigue." "Overwhelmed." "Impossible." "Exhausted." "Drained."

"The sovereign creator is a myth. The reality is a gig worker with a newsletter and no health insurance." — Former full-time newsletter writer

The distribution of outcomes follows a brutal power law. The top 20 percent of content captures 76 percent of views. The most-viewed piece gets 64 times the views of the median. Among Substack's 63,000+ newsletters, 0.3 percent of writers earn over $100,000. The median earns essentially nothing.

July 2025 brought what some called an "extinction event"—writers reported 70-90 percent decreases in new subscriber growth, attributed to algorithm changes and market saturation. The platforms that promised liberation have become landlords extracting rent from attention.

EDITORIAL

The Verification Vacuum Demands New Infrastructure, Not Nostalgia

The five domains examined in this edition—data integrity, ethics governance, education, labor markets, and the creator economy—appear distinct. Beneath the surface, they share a common architecture of failure.

Each exhibits the same pattern: an aspirational premise, an operational reality that diverges predictably, and a gap that is not accidental but traceable. The divergence follows from incentive structures, measurement systems, and accountability mechanisms that were poorly designed—or designed for something other than the stated goal.

The meta-problem is verification collapse. As the cost of generating content, credentials, code, and communication approached zero, the cost of verifying provenance, authenticity, and quality became prohibitive.

"The question is not whether we can build new verification infrastructure. It's whether we will."

The rational response to a verification vacuum is paralysis. But paralysis is not stability. It is decay in slow motion.

There are good reasons for pessimism. The contamination of the data substrate is essentially permanent. The educational credentialing crisis has no obvious solution. The labor market's structural changes appear durable. The creator economy was never designed to support a middle class.

And yet. The data also reveals something else: where operational changes have been made, outcomes improve. Cohort-based programs achieve 80-96% completion with the same content that fails in MOOCs. Model collapse can be mitigated by preserving original data. Ethical deliberation correlates with better outcomes when individual accountability is present.

The solutions hiding in plain sight are procedural, not rhetorical: Heritage Data Trusts treating pre-2023 archives as protected baselines. Design constraints rather than advisory boards. Completion-contingent revenue for educational platforms. Apprenticeship revivals. Transparent unit economics for creators.

The temptation is nostalgia—retreat to handwritten exams, physical gatekeepers, the world before generation was free. This is understandable but insufficient. You cannot solve a verification crisis by pretending the technology doesn't exist.

What's needed is new verification infrastructure: provenance-tracking systems, cryptographic attestation, watermarking protocols, curation standards. The technology that created the crisis could also help solve it—if the operational will to deploy it can be mustered.

Is this paralysis temporary—a transitional phase before new verification mechanisms emerge—or is it the new normal? The answer depends on choices not yet made.

The Great Stasis is not a pause before resumption of normal service. It is a signal that "normal" has ended.