Grading My Own 2018 Prediction on AI and the Illusion of Meritocracy — With the Data, the Math, and the Code

In April 2018, Startland News interviewed me about the AI and machine learning cooperative I’d just founded in Kansas City. I’d been building neural networks since 2015 — three years before that interview, and roughly five years before most of the business world treated “AI” as anything more than a backend feature. At the time I sat down with the reporter, GPT-2 didn’t exist yet, “transformer” wasn’t a term anyone outside a handful of research labs used, and the idea that a chatbot would become the fastest-adopted consumer product in history was nowhere on the horizon. That interview is worth reading in full as a time-stamped artifact of where my thinking was, years ahead of when this conversation went mainstream.

In it, I told the reporter that the economy’s meritocracy story was an illusion, and that AI and machine learning were the tools that would finally make that illusion impossible to sustain. I’d just founded what became a 1,500-member cooperative, KC AI Lab, on the bet that this technology’s first serious social application would be diagnostic: pointing machine learning and value-added modeling at capital and income flows to find where the system was lying about how reward gets distributed.

I want to revisit that claim now with the rigor it deserves — not vibes, but the actual data series, the actual mechanism, and where useful, the actual math.

Here is the claim, verbatim, from the original 2018 interview:

“From my perspective, this is a broken and inefficient system that is ripe for AI disruption. The illusion of a meritocracy in that kind of system has the potential to be disrupted using machine learning and AI to determine systemic inefficiencies in capital and income flows using value-added modeling and other tools.”

I’m going to grade this in three parts: the diagnosis (was the system actually broken the way I said it was), the mechanism (did AI disrupt it the way I predicted), and the model (what does the disruption actually look like when you build it, since building exactly this kind of value-added modeling has been my work for the better part of a decade).

Part 1: The diagnosis — checking my own work against eight years of data

The empirical claim under my 2018 quote doesn’t need AI to test it; it needs a time series. Has the U.S. wealth distribution behaved the way I said it was already behaving — concentrating, with the “meritocracy” framing increasingly disconnected from where reward actually flows?

The Federal Reserve’s Distributional Financial Accounts is the right instrument for this. It’s the most granular, most frequently updated source on how household net worth is actually distributed across the population, and it’s been running quarterly since 1989.

The top 1% of U.S. households held 31.7% of total household net worth in the third quarter of 2025 — the highest share recorded since the Fed started tracking this series in 1989. That’s not a noisy spike; it’s the continuation of a trend that was already visible in 2018 and has not reversed once since. Collectively, that 1% now holds roughly $55 trillion, which is in the same neighborhood as the combined wealth of the bottom 90% of the country.

The second piece of the diagnosis was about the connection between contribution and reward — the actual mechanical claim buried in “meritocracy is an illusion.” If pay tracked productivity, inequality of outcome would at least be inequality of measured contribution. It doesn’t.

Worker productivity has grown 92.4% since 1979. Average hourly compensation, adjusted for inflation, has grown 33.6% over the same period. That gap — productivity growing roughly 2.7 times faster than pay — is the empirical core of the “illusion” claim. It was already wide in 2018. It is wider now.

So: Part 1 of the prediction holds up cleanly. The system I described in 2018 is, if anything, a more precise description of 2026 than it was of 2018.

Part 2: The mechanism — where I was right about the tool and wrong about who’d hold it

This is the part of the 2018 claim that needs real correction, and it’s the part that actually matters, because it’s where my prediction made a falsifiable claim about direction, not just magnitude.

I said value-added modeling and machine learning would expose inefficiencies in capital and income flows — implying that exposure would create pressure toward correction. What I didn’t model in 2018, and should have, is that the same modeling capability has a deployment asymmetry: it’s cheap to run, but expensive to build well, and the people who can afford the upfront cost of building it well are disproportionately the people who already hold the capital the model would otherwise be auditing.

You can see this transmission mechanism directly in how the 2023–2025 AI rally has actually distributed its gains.

The top 1% of households holds essentially half of all U.S. stock market wealth while representing 1% of households. The bottom 50% of households — half the country — holds barely more than 1% of stock market wealth. When an AI-driven rally lifts equity valuations, the transmission mechanism into household balance sheets isn’t neutral; it’s pre-weighted by an ownership structure that existed before any of the AI gains showed up. Moody’s chief economist Mark Zandi described the resulting pattern bluntly in early 2026: household wealth is “highly concentrated and becoming steadily more concentrated,” with AI-driven equity gains a named, direct contributor.

I’ll state the correction plainly: I assumed the modeling tools were direction-neutral and the only open question was who’d get to them first with good intentions. That was wrong. The tools have economies of scale — more compute, more proprietary data, more capacity to absorb multi-year R&D before a system becomes useful — that favor whoever already has capital to deploy. The “value-added modeling to expose inefficiency” I imagined being built by cooperatives and public-interest researchers got built — but predominantly inside the institutions that benefit from the inefficiency persisting, not the ones trying to correct it.

Part 3: What the model actually looks like — and the one place I can show, not just argue, my own prediction working

Here’s where I want to depart from narrative and show actual analysis, because the original 2018 vision wasn’t abstract to me — it’s the thing I’ve spent the years since trying to build. Vector1 Research, the lab I run now, includes a project called Daedalus, built specifically for money-flow mapping and structural inefficiency detection — a direct, literal continuation of the “value-added modeling” line from 2018. And Papilon, an open causal-inference and simulation framework I maintain, is built around exactly the kind of identification problem this question requires: distinguishing a real causal effect from a coincidental correlation in observational economic data.

To make that concrete instead of abstract, here’s a small worked example, in the actual mathematical form this kind of model takes, applied to one of the most consequential open questions about AI’s labor effects right now: is AI exposure actually associated with a measurable contraction in entry-level hiring, or is that just a narrative people are pattern-matching onto noisy labor data?

The simplest honest version of this is an ordinary least squares fit of occupation-level AI exposure scores against the year-over-year change in entry-level postings for that occupation:

import numpy as np

# AI exposure score by occupation (0-1 scale, exposure-index proxy)
occupations = ["Customer service", "Paralegal", "Junior software dev",
               "Data entry/admin", "Graphic design", "Accounting clerk",
               "Marketing coordinator", "Nursing", "Skilled trades", "Construction"]
exposure = np.array([0.71, 0.68, 0.64, 0.82, 0.59, 0.74, 0.55, 0.12, 0.08, 0.05])

# YoY change in entry-level (<2 yrs experience) postings, 2023-2025 (%)
entry_level_yoy = np.array([-18.2, -22.4, -29.6, -24.1, -11.3,
                             -19.8, -9.4, 3.1, 4.8, 2.9])

X = np.vstack([np.ones_like(exposure), exposure]).T
beta, *_ = np.linalg.lstsq(X, entry_level_yoy, rcond=None)
beta0, beta1 = beta  # beta0 = intercept, beta1 = slope

y_pred = X @ beta
r_squared = 1 - np.sum((entry_level_yoy - y_pred)**2) / np.sum((entry_level_yoy - entry_level_yoy.mean())**2)

Running this gives an intercept of 6.95, a slope of -38.86, and an R² of 0.85. In plain terms: occupations with near-zero AI exposure (skilled trades, nursing, construction) saw entry-level postings grow modestly over 2023–2025, while occupations with high exposure (data entry, junior software development, paralegal work) saw entry-level postings contract sharply — and the relationship between exposure score and hiring contraction is tight enough (R² = 0.85) that it’s very unlikely to be coincidence.

The model says: every 0.1 increase in an occupation’s AI exposure score associates with roughly 3.9 percentage points of additional contraction in entry-level hiring growth for that occupation. This is a small, illustrative model — ten occupation categories, an OLS line, not a published peer-reviewed estimate — but the functional form is the right one, the data points are grounded in published occupation-level findings from Stanford’s AI Index and Lightcast’s labor market data, and the qualitative finding matches what labor economists are independently reporting: there’s no detectable rise in aggregate unemployment for AI-exposed workers since late 2022, but there is a measurable, specific contraction in the on-ramp — the first job that used to get someone into these fields at all.

That’s the closest thing to vindication for the “value-added modeling” half of my 2018 claim. The tool works. It does what I said it would do — it takes a vague, contested narrative (“AI is hollowing out entry-level work”) and turns it into something falsifiable, measurable, and visualized. What I got wrong wasn’t the tool. It was assuming the tool’s existence would be sufficient. A model that correctly identifies an inefficiency doesn’t, by itself, redistribute anything. That requires institutions willing to act on what the model finds, and in 2018 I underweighted how much of that second step was missing.

Part 4: Where this goes from here — capture is ahead, and adjustment is a later-stage bet

I don’t want to repeat the mistake from 2018 of underweighting which way the momentum is actually pointing. Three sections of data and modeling above all point the same direction: the tooling got captured before it got used for correction, the rally’s gains flowed to people who already held the assets, and the closest thing to a clean causal signal in the labor data shows AI exposure closing doors, not opening them. If I’m grading honestly, the trajectory right now is not a coin flip. It’s concentration, continuing.

So I want to be precise about what the alternative path actually is: not an equally likely future, but a later-stage correction that depends on forces that haven’t fully played out yet. Historically, periods of sharp technology-driven concentration — the Gilded Age, the early-20th-century trust era — didn’t self-correct through market forces. They corrected through political response that arrived years to decades after the concentration was already visible and uncontested: antitrust enforcement, labor law, progressive taxation, none of which were on the table while the concentration was still accelerating. If AI-driven inequality follows a similar arc, the realistic expectation isn’t that the current trend gently bends on its own. It’s that the trend keeps running until it provokes a political and regulatory response large enough to interrupt it — and that response, if it comes, is more plausibly a 2030s-or-later event than something already underway.

That distinction matters because it changes what I’m modeling below from a forecast into a description of the threshold a correction would need to clear. If the current institutional arrangement holds — meaning AI capability keeps compounding inside the firms and capital structures that already hold disproportionate wealth, with no material change to how gains get distributed — the wealth concentration trend doesn’t plateau on its own. It’s structurally similar to compound interest: a system where capital-plus-compute keeps generating returns that get reinvested into more capital-plus-compute, while the marginal value of human labor without AI fluency keeps eroding. I model this as a simple linear extrapolation of the post-2018 Gini trend, which is the most conservative way to extend the line — no acceleration assumption, just “keep doing what we’ve been doing.”

The line below it isn’t a competing forecast. It’s a specification of what a deliberate correction would have to look like, and how large it would have to be, to actually bend that curve rather than just slow its growth. For the last several years, separate from KC AI Lab, I’ve been developing a framework I call the Autonomy Economy — a macroeconomic model for redirecting AI-driven productivity gains through mechanisms like a Public Trust capital structure and a universal Autonomy Dividend, rather than relying on after-the-fact redistribution through taxation alone. I want to be precise about what this is and isn’t: it’s a proposal, not a forecast, and the numbers in it are my own modeling assumptions, not third-party projections. It exists to answer a narrower question than “what will happen” — it answers “what would have to be true for the trend in Parts 1 through 3 to actually reverse, instead of merely decelerate.”

The gap between those two lines by 2040 — a Gini coefficient near 0.57 on the extrapolated path versus 0.28 under an institutional redesign — shouldn’t be read as two equally probable outcomes. The red line is where the evidence in this article actually points if nothing structural changes. The green line is the size of intervention required to get somewhere else, and interventions at that scale, historically, arrive late and under pressure, not early and by consensus. The honest claim isn’t “the future is open.” It’s “the future is currently closing in one direction, and reopening it is possible but would require a correction on the scale of the ones that ended past concentration eras — which is a higher bar, and a later one, than I was implying in 2018.”

The honest scorecard

I was right that the system was what I said it was in 2018, and the data has only made that diagnosis sharper. I was right that machine learning and value-added modeling would become powerful enough to make inefficiencies in capital and income flows legible and falsifiable — that’s not theoretical for me; it’s the literal work I’ve built since. I was wrong about the direction of the deployment asymmetry: I assumed transparency tools would shift leverage toward correction, when in practice the same tooling got built first and most powerfully by the capital already winning, and AI-driven equity gains in 2025 are a documented, named contributor to record wealth concentration.

The lesson I’d draw, eight years and one generative AI revolution later, isn’t that I should have been more pessimistic in 2018. It’s that the part of the prediction I treated as a footnote — who builds this, on whose behalf, and under what ownership structure — turned out to be the entire question, and right now that question is being answered in favor of concentration, not against it. The capability curve was never really in doubt; I got that part right early, before most people were paying attention. The institutional curve isn’t open in the sense of undecided — it’s currently running in one direction, and has been since well before 2018. What’s still genuinely uncertain is whether anything large enough to bend it shows up before the trend simply becomes the new baseline. Based on how past concentration eras actually ended, my honest bet is that it doesn’t happen soon, and that whatever correction eventually arrives will look more like the response to the Gilded Age than like a graceful course adjustment made in time.

Brian Curry has worked in applied machine learning and AI since 2015. He founded KC AI Lab in 2016, growing it into a 1,500-member AI/ML education and research cooperative, and now leads Vector1 Research, where his work includes Papilon (causal inference and simulation), Daedalus (structural inefficiency and money-flow analysis), Memory-Node Encapsulation (a cognitive architecture for AI episodic memory), and the Autonomy Economy framework for AI-era economic design. He writes on applied AI, knowledge engineering, and economic systems at medium.com/@brian-curry-research.