may 2026 · field guide for ai-era builders

The technology
trap, applied.

Carl Benedikt Frey's history of automation has six lessons for AI-era founders. The trap is structural — it springs from the cluster's own success.

based on The Technology Trap (Frey, 2019) · talks at google · social europe podcast · LSE review

01 / the thesis

Tech is destiny. Politics is not.

Frey's central claim: whether a new technology lifts a society or breaks it depends on the distribution of political power, not on the technology itself.

technology steam · electricity computers · AI political power who has the vote who has capital who writes the rules who can resist enabling outcome wages rise middle class grows replacing outcome wages stagnate middle class hollows

Same machine, two outcomes. A steam engine in 1810 Britain: replacing. The same engine 80 years later in welfare-capitalist America: enabling. Frey: "whether workers lose their job to robots ultimately depends on the societal distribution of political power."

why this matters Most "tech is destiny" debates assume the technology determines the outcome. Frey's history shows the opposite: the same technology produces opposite social outcomes depending on who has political power. The leverage isn't in the chip; it's in the coalition.

02 / the distinction

Labor-enabling. Labor-replacing.

Every wave of technology falls into one of two molds. The same physical machine can be either, depending on how it's deployed.

labor-enabling "makes the worker more valuable" worker + machine → ↑ wages ↑ jobs examples: → unit-drive electric motors (1920s) → word processors (1990s) → Cursor / dev copilots (2024–) → AI for radiology read-aloud labor-replacing "makes the worker disappear" worker × machine → ↓ wages ↓ jobs examples: → mechanized factory (1810s) → industrial robots (1980s+) → self-checkout (2010s) → end-to-end agent for X workflow

Frey's claim is that history alternates. The First Industrial Revolution (~1780–1860) was largely replacing: artisan wages collapsed, the labor share of income fell, riots followed. The Second Industrial Revolution (~1880–1970) was largely enabling: electrification + welfare capitalism + unions created the broad middle class. The Computer Revolution (1980–) returned to replacing: middle-skill clerical work hollowed out. AI continues that mode by default — but doesn't have to.

the founder version Pure-replacement framings ("we replace the engineering team") trigger the same political coalition that delayed industrial automation in the 1810s. Enabling framings ("a 10× engineer") preserve the political license to operate. The product can be identical; the framing isn't.

03 / engels' pause

Productivity goes up. Wages don't.

During Britain's First Industrial Revolution, output per worker grew 46%. Real wages grew 12%. The 60-year gap is what economists call "Engels' Pause" — and it's repeating now.

+50% +40% +30% +20% +10% 0 1780 1800 1820 1840 1860 +46% productivity per worker +12% real wages engels' pause ~60 years of decoupling first industrial revolution · britain

The same shape exists right now in the US. Since 1979, labor productivity has grown ~8× faster than hourly compensation. The labor share of income has fallen for 40 years. We are five-plus years into AI's contribution to that gap. History says it can persist for two generations if not corrected — Frey's central warning.

if you're under 35 The "rising tide lifts all boats" assumption fails Engels' Pause. The median British worker in 1820 was no better off than the median in 1780 — the pie grew, but their slice didn't move. The historical base rate disagrees with the optimist case. Plan for capture, not for trickle-down.

04 / why the luddites lost

An army larger than Wellington's.

The Luddites weren't anti-progress reactionaries. They were rationally opposing labor-replacing machinery. They lost because they didn't have the vote.

1810s britain no enfranchisement property requirement to vote → army deployed larger than Wellington's at Peninsular War 1900s us universal suffrage + unions + welfare capitalism → progressive era → new deal → middle class today vote · weak unions no riots — populism instead → trump · brexit → EU AI Act → AI bills, 12 states → ~90 years → → ~120 years →

Frey: "if horses could have voted, the tractor wouldn't have spread." The Luddites' political voicelessness is what let mechanization steamroll them. Today's displaced workers have the vote. They don't break machines — they elect populists who break policy.

the modern luddite Doesn't smash a machine. Votes for someone who passes a state copyright law, an AI disclosure rule, a class-action authorization for AI training data. Same political function — different mechanism. Faster, less violent, equally constraining.

05 / power

Capital concentration is more extreme now.

Frey is bullish on the present in one specific way: workers can vote. He's bearish on capital concentration. Foundation models compress the Carnegie/Rockefeller era into a handful of labs.

industrial age (~1900) capital concentrated in steel, oil, rail ~50 firms held most industrial capital AI age (2026) capital concentrated in foundation labs openai google anthropic meta xai ~7 labs hold most frontier capability

Industrial-age antitrust took ~30 years to develop (Sherman Act 1890 → Standard Oil 1911 → progressive era). Foundation-model concentration is more extreme — with network effects, data moats, and compute scarcity that didn't exist mechanically. The political coalition for AI antitrust is forming faster than it formed for steel and oil. Expect compressed timelines.

06 / the trap itself

The cluster traps itself.

The technology trap isn't tech vs progress. It's the feedback loop where the cluster's own success creates the political conditions that constrain the technology.

01 cluster wealth 02 visible wealth gap 03 populist coalition forms 04 constraints passed 05 tech adoption slows 06 cluster captures less the trap cluster's success → political coalition → cluster's constraint

The trap is structural, not tactical. You don't escape it by being personally virtuous about your tech. The mechanics are: concentrated capital + visible inequality + universal suffrage. If your industry has all three, the loop runs whether you want it to or not. Historical exit (1900s–1940s): unions, progressive taxation, the New Deal. The question for AI: what plays that role this time, and who builds it?

your cluster For most readers of this page: SF Bay Area + a strip of Manhattan. The political conditions shaping AI policy in 2028 are being set by people reacting to the wealth and visibility of your cluster right now. You're not outside the loop. You're inside it.

07 / patterns

What's the same. What's new.

Most of the dynamics rhyme with the First Industrial Revolution. A few are genuinely unprecedented.

same as 1810s britain

  • Productivity ↑↑, wages ↑. Engels' Pause is repeating. Labor share of income is at a 50-year low.
  • Middle-income job hollowing. IR killed artisans; AI is hollowing knowledge work that was supposed to be safe.
  • Geographic concentration. Manchester then; SF, NYC, Beijing, Shenzhen now.
  • Resistance from displaced workers. Luddites then; populism + lawsuits + state bills now.
  • Innovation needs a capital + power coalition. Same political prerequisite, different instruments.

genuinely new

  • Universal suffrage. Workers vote — no bayonets at protests. Backlash takes the form of policy.
  • Compressed timeline. IR took ~80 years (1780–1860). AI's analog may compress to 10–20.
  • Cognitive work is the target. Knowledge workers were "safe" in every prior wave. Not this one.
  • Capital concentration is more extreme. ~7 labs control foundation models. Carnegie's empire was less concentrated.
  • Network + data moats. Winner-take-all dynamics that didn't exist mechanically.
  • Welfare state already exists. Cushion + political constituency for keeping it intact.
  • Recursive risk. AI can plausibly substitute for AI engineers. First time tech can replace its own designers.

08 / the gap that kills you

Your buyer ≠ your rule-writer.

The people who use and pay for your product live in one place. The people who decide what your product is allowed to do live somewhere else. Optimizing for one without the other is how startups die.

rule-writers never see your demo · decide your fate your buyer enterprise team in the cluster reads your changelog → EU parliament (AI Act) → NY Times v. OpenAI judges → state attorneys general → labor boards → Brussels regulators → California legislature → Texas, Tennessee, Utah → federal Congress → city councils (Airbnb) → DMVs (Cruise) → Treasury, OCC (Stripe) → FAA (delivery drones) your TAM is the inner circle. your survival depends on the outer.

Concrete examples. Cruise sold robotaxi rides to consumers in SF — the California DMV killed the business. Airbnb sold to travelers — city councils decided one-by-one whether the business could exist. OpenAI sells API access to enterprises — the EU AI Act + NY copyright cases set what they can ship. Optimizing for the buyer doesn't insulate you from the rule-writer.

the implication Most early-stage founders treat policy and public communications as overhead. Frey's framing flips it: legitimacy work outside the buyer relationship is the actual cost of operating in a politically-exposed industry. AI is now politically exposed. Your TAM analysis is incomplete without a rule-writer analysis.

09 / six plays

What to actually do.

Six operating moves for AI-era founders, derived from Frey's history. Tactical, not philosophical.

  1. enabling Bet on labor-enabling, not labor-replacing. Pure-replacement framings ("we replaced the team") trigger the same coalition that delayed industrial automation in the 1810s. The product can be identical; the framing isn't. Cursor wins; "AI engineer in a box" loses.
  2. power-read Read the political-economy stack around your product. Who has model power, regulatory leverage, cultural legitimacy? AI labs hold model power. App layer is at commoditization risk if labs verticalize (Cowork, Stitch from Google, etc. already show this).
  3. engels now Engels' Pause is happening — plan for capture, not trickle. AI productivity gains are real. Wages aren't moving. History says this gap can persist for 60+ years if not corrected. Don't anchor on "the pie grows" — build for distribution.
  4. cluster-aware Geographic concentration is the moat — and the trap. Be in the cluster (SF, NYC, Shenzhen). But do legitimacy work outside the cluster: deployments in healthcare in the South, manufacturing in Michigan, government contracts spread across districts. Political durability isn't a market-expansion play; it's an existence play.
  5. past reskill "Just reskill them" is a policy myth. Frey is explicit: retraining alone is not adequate. Displaced workers historically can't pivot in their lifetimes. Real responses: relocation vouchers, place-based investment, antitrust on capital concentration. For founders, this means: don't pretend "AI for everyone" is the social contract. It isn't.
  6. TAM ≠ rule Map your rule-writers before you map your TAM. 50 state AGs, EU parliament, federal Congress, courts, labor boards. None of them buy your product; all of them can kill it. Public communications, policy engagement, and narrative work are the actual cost of operating in a politically-exposed industry — not overhead.