--- title: "Babysitting The Slop" date: 2026-03-18 outlet: forbes originalUrl: "https://www.forbes.com/sites/christiancatalini/2026/03/18/babysitting-the-slop/" canonical: original rights: full tags: [ai-agi] deck: "Generating the output is free. Knowing when it’s lying is the moat. Why verification—not intelligence—is the binding constraint on the AI economy." image: "/images/writing/babysitting-the-slop.jpg" --- In 1842, the managers of Lowell, Massachusetts's textile mills had what seemed like a straightforward idea. Their weavers each operated two power looms. The mills had just acquired more machines. The math was simple: give each weaver a third loom, output goes up by fifty percent. It didn’t work. Something counterintuitive happened instead. With three looms running, even the most experienced weavers couldn't keep up. Not with the physical labor—the machines did the weaving. The problem was *monitoring*. A weaver's real job wasn't making cloth. It was watching cloth get made: scanning for broken threads, catching defects, making micro-adjustments before a flaw could propagate through yards of fabric. Add a third loom, and the verification load exceeded human bandwidth. The mills had to cut loom speeds by 15% just to keep quality from collapsing. They'd upgraded the machines, only to discover the machines weren't the bottleneck. It took a full year of retraining before weavers could run three looms at full speed. By 1902, a single American weaver managed 18 power looms and produced over 50 times the output of a weaver a century earlier —but only after mills tripled their training investment per worker, from $47 to $162. [Bessen](https://x.com/JamesBessen), who painstakingly reconstructed this [history](https://scholarship.law.bu.edu/cgi/viewcontent.cgi?params=/context/faculty_scholarship/article/4181/&path_info=More_Machines__Better_Machines...Or_Better_Workers.pdf), found that once the power loom itself was in place, 62% of the remaining productivity gains through the Northrop loom era came not from further mechanical inventions, but from better-skilled humans who could monitor more looms at once. "The weavers were not simply idle during this time," he wrote. "They were monitoring". The binding constraint was never the loom. It was never the thread, the cotton, or the power source. It was the weaver's ability to verify the loom's output. Today, the loom is an AI agent. The constraint hasn't changed. ## The Bicycle and the Rocket In a [recent paper](https://arxiv.org/pdf/2602.20946), my co-authors [Xiang Hui](https://olin.washu.edu/faculty/xiang-hui), [Jane Wu](https://www.anderson.ucla.edu/faculty-and-research/strategy/faculty/wu), and I argue that the binding constraint on AI-driven economic growth is not intelligence, not compute, not execution capacity. It is human *verification bandwidth*—the scarce ability to confirm that AI agents actually did what they were supposed to do, correctly, safely, and honestly. The core concept is what we call the [Measurability Gap](https://x.com/ccatalini/status/2026311825898504608?s=46), and understanding it is the key to understanding why AI progress on benchmarks keeps diverging from AI progress in the real economy. Two cost curves are moving in opposite directions. The cost to *generate*—to have an AI produce code, analysis, text, decisions—is falling exponentially with every new model release. The cost to *verify*—to confirm the output is correct, safe, non-hallucinated, and aligned with intent—remains stubbornly bottlenecked by human time, domain expertise, and feedback latency. One curve is a rocket. The other is a bicycle. The widening space between them is where economic value goes to die. The numbers make this vivid. SWE-bench coding accuracy jumped from 4.4% to 71.7% in a single year. Task horizons for autonomous agents are doubling on a sub-year cadence. AI can now write entire pull requests, draft regulatory filings, generate onboarding workflows. But Google's DORA reports show that greater AI adoption is associated with *lower* delivery stability. Daron Acemoglu projects AI's total factor productivity contribution at a modest 0.53–0.66% over a decade—not because AI isn't capable, but because verification bottlenecks cap the value that actually gets realized. And when verification fails, it fails spectacularly. Days ago, a [Chinese cybersecurity company](https://x.com/ccatalini/status/2033742180607791529?s=46) accidentally shipped its private cryptographic key inside a public release—a verification lapse so basic it's almost comic, except that it exposed the entire signing infrastructure that authenticated their software. Nobody checked. The AI economy is increasingly full of outputs that nobody checked. The automation boundary, we argue, is no longer [“routine versus non-routine”](https://arxiv.org/html/2602.20946v2#S7)—the old framework economists used to predict which jobs machines would take. The new boundary is **measurable versus non-measurable.** If a human can verify the output in seconds—thumbs up on a chatbot response, quick visual check on a generated image, run the code and see if it compiles—adoption is fast and value capture is real. If verification requires domain expertise, extended time horizons, or adversarial stress-testing, the Measurability Gap yawns open and adoption stalls, even when the AI is technically brilliant. This is why the first killer AI products were chat, image generation, and code autocomplete. Not because those were the hardest problems. Because they were the most *verifiable*. The hard frontier is everything else—and it's most of the economy. We are getting dramatically better at producing output and not meaningfully better at confirming it's right. ## Your Data Moat Is Probably the Wrong Data For any organization sitting on proprietary data, this framework forces a surprisingly uncomfortable question: which of your data assets actually matter in an AI world? We distinguish two types. The difference is subtle but has enormous strategic implications. **1. Execution-grade knowledge** is the stuff that teaches an AI *what to do*. Finished code. Completed contracts. Final reports. Clean datasets. Polished outputs. This is what most companies think of when they hear "proprietary data advantage." And it is structurally vulnerable. As foundation models improve—as they ingest more public examples of good code, good contracts, good reports—the marginal value of your private stash of finished work declines. The model already knows how to write a contract. Your 10,000 historical contracts help a little less with each generation of frontier model. **2. Verification-grade knowledge** is different. It teaches systems *what to reject*—and more importantly, *why*. This is the data generated at the point of failure: redlines on deals that almost closed but didn't, deployment blocks on code that passed tests but broke in production, false-positive logs from fraud detection, near-miss records from compliance reviews, edge-case adjudications where expert judgment overrode the default answer. It is the institutional memory of things going wrong, encoded with enough context to be actionable. Here's why verification-grade data is so much more valuable—and why its value is actually *increasing* with AI capability. In a world where AI can effectively convert data and compute into labor—where a sufficiently rich dataset can train, fine-tune, or prompt a model to replicate expertise—the quality, recency, and granularity of your data determines how much skilled labor you can synthesize. This is the **labor as software** thesis: data doesn't just inform decisions, it *becomes* the decision-maker. Execution-grade data gives you a model that can produce outputs. Verification-grade data gives you a model that can *audit* outputs. And in a world where production is cheap and verification is expensive, the second capability is worth dramatically more. Think of it concretely. A bank's accumulated database of KYC/AML failures, fraud false-positives, edge-case adjudications, and compliance near-misses is verification-grade ground truth. Each record encodes a moment where a human expert said "this looks right but isn't" or "this looks wrong but is actually fine" or "this is the specific reason we blocked this transaction." That judgment, captured at scale with timestamps and context, is the raw material for training verification systems that can eventually widen the verification bottleneck itself. And because these records are more recent, more fine-grained, and more domain-specific than anything in a foundation model's training corpus, they get *more* powerful as AI improves—not less. Better base models mean you can extract more signal from each verified data point. Your failure library becomes a force multiplier. A competitor can replicate the ability to *generate* an onboarding workflow. It cannot replicate the 15 years of institutional failure-memory required to *trust* that workflow at scale. The verification-grade data is the moat precisely because it accrues slowly, is hard to fake, and becomes more valuable as AI makes generation cheaper. The asymmetry cuts deeper than most people realize. AI lowers the cost of executing tasks far faster than it lowers the cost of verifying whether those tasks were done honestly. A scanned passport used to be identification—now it's raw material for a generative model. A video selfie used to be a liveness check—now it's a challenge to the best deepfake on the market. [You can automate the paperwork faster than you can believe the paperwork.](https://www.forbes.com/sites/christiancatalini/2026/03/10/software-eats-the-brokerage-ai-lowers-cost-crypto-makes-it-verifiable/) Every advance in generation is simultaneously an advance in the attack surface for verification. The dominant strategy follows: rent cognition, own trust. Use frontier models for reasoning. Privatize your domain context and your verification stack. Execution commoditizes. Verification is the moat. ## What Happens When the Network Effect Runs in Reverse There is a related problem worth thinking through carefully, because it threatens one of the core assumptions of platform strategy. The standard playbook for platforms goes like this: get big, get network effects, let the flywheel spin. More users attract more users. The platform gets better because it gets bigger. This is the story of every successful marketplace, social network, and exchange.Now consider what happens when AI agents join your platform. Because they imitate human behavior so convincingly, platforms can’t easily filter them out. But beneath that plausible veneer, they are all trained on the same data and suffer from the same architectural blind spots. We call this "agentic slop," and it has a fatal flaw: common-mode failure. When these agents hallucinate or guess wrong, they don't make random, independent human mistakes. They make the exact same mistake. A million agents flooding a platform don't give you a million independent signals. They give you a million perfectly correlated, human-passing mistakes, making it impossible to tell what is actually real. So network effects invert. Instead of more participants making the platform more valuable, more AI-generated participants make the platform less trustworthy. The quality humans—the ones whose verified behavior is actually valuable—notice the rising noise floor and leave. The trust flywheel doesn't just slow down. It runs in reverse. The metric that matters isn't raw network size anymore. It's what we call *verified network scale*—the authenticated, verified share of participants whose outputs can actually be trusted. Platforms that measure and defend that number will survive. Platforms optimizing for raw volume are building on quicksand. Scale without verification isn't a moat. It's debt that compounds. For regulated industries—banking, healthcare, insurance—this dynamic cuts both ways. These sectors resist full automation the longest, because verification requirements are written into law. That's expensive and frustrating today. But once a verification stack clears regulatory gates, it becomes nearly impossible to dislodge. Nobody voluntarily rips out a compliance infrastructure that a regulator has signed off on, especially not to replace it with something newer and unproven. The barrier that slows you down on the way in is the same barrier that keeps competitors out once you're through. ## Liability Is the Product Now The emerging organizational model is what we call the [AI Sandwich](https://arxiv.org/html/2602.20946v2#S7). Directors at the top, navigating uncertainty and orchestrating agent swarms. Verified agents in the middle, executing at scale. And Liability Underwriters at the bottom, serving as adversarial auditors who absorb and price risk. The bottleneck in this model is verification bandwidth, not headcount. Revenue models follow. The shift is from SaaS to **Liability-as-a-Service**: the product is not the agent but the *indemnified outcome*. In February 2026, ElevenLabs launched an AIUC-1 certified insured AI voice agent—insurance bundled as the product boundary. Value these firms the way you'd value insurers: by underwriting margin, loss experience, and reserve adequacy. And short the firms that are liquidating verification capacity—automating junior roles without replacing the training pipeline—because they are converting future oversight capability into current earnings. That trade has a name in finance. It's called eating your seed corn. ## The Weavers Knew Return to Bessen’s weavers. The 19th-century mills that thrived didn't just buy more looms. They tripled their training investments. The ones that simply added machines and skimped on human oversight failed. Bessen found that the elasticity of substitution between capital and labor in weaving was strikingly low—between 0.23 and 0.26. In plain English: you couldn’t just throw machines at the problem. The two factors were so tightly coupled that upgrading the looms without comparably upgrading the weavers left most of the potential gains on the table. Capital and skilled labor weren't substitutes. They were complements. The machine was only as good as the human checking its work. The modern parallel holds exactly. The companies and institutions that will capture AI's value aren't the ones deploying the most agents. They are the ones investing in the capacity to verify what those agents produce—building the failure libraries, training the next generation of expert auditors, underwriting the liability when the machine is wrong, and turning their verification-grade data into the training signal that makes the next round of verification systems better. The defining challenge of the agentic economy is not the race to deploy. It is the race to verify. The loom has never been faster. The question is whether anyone is watching the cloth.