← Back to Stories
💼 Labor & AI March 22, 2026

AI Has 500x More Supporters Than Opponents. It Has Zero Lobbyists Per 13 Million Users.

Nuclear energy had the technology to solve climate change in 1975. It lost the political battle and got regulated into $35 billion plants and 20-year build cycles. AI's unorganized supporters outnumber its organized opponents 500 to 1. That ratio means nothing without a political strategy. Here is what the industry should do before the concrete sets.

Split image: gleaming AI datacenter on one side, abandoned nuclear cooling tower covered in red tape on the other

Nadia Kovac · Labor & AI Policy · March 22, 2026

The Graveyard of Technologies That Won the Science and Lost the Politics

Plant Vogtle, in Burke County, Georgia, is the only nuclear power plant built in the United States in the last three decades. Construction began in 2009. The original budget was $14 billion. The original completion date was 2016. Unit 3 finally entered commercial operation in July 2023. Unit 4 followed in April 2024. The final cost exceeded $35 billion. The prime contractor, Westinghouse, went bankrupt during construction.

France, meanwhile, built 56 nuclear reactors between 1974 and 1999. It now generates 70% of its electricity from nuclear power, has among the lowest carbon emissions per kilowatt-hour in Europe, and charges its citizens roughly half what Germans pay for electricity. The physics was identical. The engineering was comparable. The difference was politics. France built political consensus around nuclear as a national project. The United States let the opposition organize first.

The technology that could have solved climate change 50 years ago was regulated into $35 billion plants, 20-year build cycles, and effective irrelevance. The reactors worked. The political strategy didn't.

AI is now sitting at the same fork in the road. And the clock is ticking faster than most people in the industry realize.

The Regulatory Flood Is Already Here

In 2025, state legislatures across the US introduced over 700 AI-related bills. Seventy-three became law across 27 states. California alone enacted 13. The EU AI Act entered force in August 2024, creating a risk-tier classification system that treats AI models the way the FDA treats pharmaceuticals, minus the outcome-based evidence requirement.

The compliance costs are already measurable. A 2025 survey of EU and UK tech firms found that small and mid-sized companies lose between $109,000 and $375,000 annually from delayed AI model launches and compliance overhead. Directly affected firms lose $186,000 to $528,000. US small businesses reported 91% AI adoption rates and 10.7% average cost savings, compared to 85% and 8.9% in EU/UK counterparts. The regulation gap is already producing a performance gap.

This is what the early stages of the nuclear scenario look like. Not a single dramatic ban. A thousand small rules that individually seem reasonable and collectively make innovation impossible for anyone without a billion-dollar legal department.

Five Technologies, Five Political Strategies, Five Outcomes

Every transformative technology eventually faces its regulatory reckoning. The ones that survive don't win on technical merit. They win on political coalition-building. The record is clear.

Nuclear power (lost). Had no organized constituency beyond the utilities themselves. Environmental groups organized against it. Local communities organized against it. The public associated it with weapons. By the time the industry tried to build political support, Three Mile Island (1979) and Chernobyl (1986) had sealed the narrative. Result: regulated into effective irrelevance in the US. Cost per plant went from $1 billion in the 1970s to $35 billion by 2024. The technology worked. Nobody cared.

GMOs (split decision). Won in the United States because farmers were the constituency. The American Farm Bureau Federation, representing 5.3 million farm families, lobbied consistently for science-based regulation. The USDA, EPA, and FDA treated GMOs as products, not threats. In Europe, no equivalent farm constituency mobilized. Environmental groups and consumer advocates controlled the narrative. Result: effectively banned in most of Europe despite scientific consensus on safety. Same technology, different politics, opposite outcomes.

Rideshare (won). Uber's genius was not the app. It was the political strategy. By launching in cities before regulators could react and building a user base of millions, Uber created a constituency that would fight regulation on its behalf. When Miami-Dade County tried to fine and impound UberX drivers in 2014, Uber paid the fines and kept operating. Within months, the county legalized ridesharing. The taxi lobby had the existing regulatory framework. Uber had 10 million riders who would email their city council. Riders won.

Crypto (learning the hard way). Tried the "move fast and ask forgiveness" approach without building a user constituency first. Got hammered by the SEC. Now spending $170 million through the Fairshake PAC (funded by Coinbase, Ripple, and Andreessen Horowitz) to elect pro-crypto candidates. The 2024 results were promising: most Fairshake-backed candidates won. But this is expensive coalition-building after the fact, not organic constituency-building before the crisis. Crypto is spending $100 million+ to buy what Uber got for free.

The internet (won decisively). Section 230 of the Communications Decency Act, passed in 1996, is arguably the most consequential piece of technology legislation ever written. Twenty-six words: "No provider or user of an interactive computer service shall be treated as the publisher or speaker of any information provided by another information content provider." This single liability shield enabled Google, Facebook, YouTube, Twitter, Reddit, Wikipedia, and every platform that followed. Without it, the internet as we know it could not exist. The key: it passed before the opposition organized. In 1996, nobody was lobbying against the internet.

AI's Political Map: Allies, Enemies, and the Undecided

AI's problem is that it has enemies who are organized and allies who don't know they're allies yet.

Organized against AI:

Natural AI allies (mostly unorganized):

The Bureaucracy Trap

Bad regulation doesn't announce itself. It arrives as reasonable-sounding requirements that compound into paralysis.

The EU AI Act classifies AI systems into risk tiers: unacceptable (banned), high-risk (heavy regulation), limited risk (transparency obligations), and minimal risk (mostly free). This sounds sensible. In practice, the classification boundaries are vague enough that compliance lawyers will argue about them for years while startups either pay up or leave. Large companies love this. Google, Meta, and Microsoft can absorb $375,000 in annual compliance costs without blinking. A 12-person startup cannot. The EU AI Act is a regulatory moat disguised as consumer protection.

HIPAA was supposed to protect patient privacy. It does. It also prevents the data sharing that could save lives. Researchers at two hospitals studying the same rare disease cannot easily combine their patient datasets. AI diagnostic systems need diverse training data from many institutions. HIPAA makes this bureaucratically excruciating. The law protects your medical records. It also, incidentally, makes it harder to develop the AI that might detect your cancer earlier.

Sarbanes-Oxley was supposed to prevent another Enron. It created an entire compliance industry. Ernst & Young, Deloitte, KPMG, and PwC collectively earned billions in SOX audit fees. Then the 2008 financial crisis happened anyway, caused by an entirely different mechanism (mortgage-backed securities) that SOX didn't address. The compliance theater was immaculate. The catastrophe prevention was nonexistent.

California's SB 1047 (vetoed by Governor Newsom in September 2024) would have required safety testing for AI models above a compute threshold. The bill had good intentions. In practice, it would have applied primarily to a handful of large model developers in California, creating liability for harms caused by downstream users of open models. Newsom vetoed it, but the next version is already being drafted.

What Good Regulation Actually Looks Like

Not all regulation is the bureaucracy trap. Some frameworks actually work.

Section 230 succeeded because it was simple, enabling, and passed before incumbents could organize against it. Twenty-six words. Clear liability shield with clear exceptions (federal criminal law, intellectual property). It didn't try to classify every possible internet service. It created a default ("you're not the publisher") and let courts handle edge cases. AI needs its Section 230 moment.

FAA aviation safety is prescriptive and expensive, but it's data-driven and outcome-based. The reason you can trust a plane is not that regulators imagined every possible failure mode. It's that every failure mode that actually occurs gets investigated, published (NTSB reports), and fed back into design requirements. The system learns. It's slow, but it works. Commercial aviation fatality rates have declined by 95% since the 1970s.

The common thread: good regulation focuses on outcomes (did someone get hurt?) rather than inputs (what technology did you use?). Bad regulation focuses on inputs because inputs are easier to measure and generate more compliance revenue.

The Nuclear Scenario

Here is what happens if AI loses the political battle.

By 2028, the EU AI Act is fully enforced. California, New York, and Illinois have passed state-level equivalents with additional requirements. Congress, unable to pass a single federal framework, has created a patchwork where AI companies must comply with 50 different state regimes simultaneously, the way financial services companies comply with 50 different state insurance regulations.

Compliance costs reach $5 to $15 million per model deployment, depending on risk classification. Pre-deployment safety assessments take 6 to 18 months. Only companies with billion-dollar war chests can afford to play. The startup ecosystem in AI shifts to Singapore, the UAE, and India, countries that chose lighter regulatory frameworks. China, which never had this debate, continues deploying AI across its economy at whatever speed its engineers can manage.

The US doesn't ban AI. It does something worse. It makes AI so expensive to deploy that only incumbents can afford it. Innovation doesn't die. It emigrates.

This is not speculation. This is exactly what happened to nuclear power. The US didn't ban nuclear plants. It made them cost $35 billion and take 20 years. Same outcome. Slower death.

What the AI Industry Should Actually Do

1. Build a user constituency, not just a lobbying budget. Meta spent a record $13.8 million on lobbying in the first half of 2025. OpenAI spent $3 million in all of 2025. This is the crypto strategy: spend money on politicians. Uber's strategy was better: spend money making a product so useful that voters become your lobbyists for free. AI is already doing this accidentally. The small business owner who uses ChatGPT for invoicing doesn't know she's an AI advocate. Someone needs to tell her.

2. Win the healthcare argument. "This AI caught my cancer" is the single most powerful political statement the AI industry can make. Every FDA-approved AI diagnostic tool, every early detection story, every patient who got screened because AI made it affordable to screen everyone instead of just those with specialists nearby, is a voter who will fight AI regulation. Healthcare is the Uber-rides equivalent: a benefit so tangible and personal that taking it away is politically impossible.

3. Co-opt the safety narrative. Companies that self-regulate credibly avoid the worst external regulation. The automotive industry created the Insurance Institute for Highway Safety. The chemical industry created Responsible Care. These were not altruistic. They were strategic. An industry that polices itself gives legislators cover to pass lighter frameworks. An industry that says "trust us" while deploying unreviewed models gives legislators no choice but to regulate aggressively. Anthropic's Constitutional AI approach is more strategically sound than OpenAI's "move fast" posture, regardless of which produces the better model.

4. Support outcome-based rules to block input-based rules. Offer to accept strict liability for actual harms in exchange for freedom from prescriptive process requirements. "If our AI misdiagnoses a patient, we pay. But don't tell us which testing framework to use." This is politically savvy because it gives regulators something concrete (liability) while preserving the freedom that enables innovation. The alternative is the EU model: no clear liability, but mountains of process requirements. Process requirements create compliance industries. Liability creates incentives to actually build safe systems.

5. Organize the unorganized allies. Small businesses, disability advocates, rural healthcare providers, and students using AI tutoring are already benefiting from AI. They just don't have a trade association, a PAC, or a lobbyist. The AI industry doesn't need to astroturf. It needs to fund genuine coalition organizations that represent the people who actually use these tools. The American Farm Bureau didn't save GMOs because Monsanto paid them to. It saved GMOs because farmers genuinely benefited and had an organization to say so.

The Constituency Gap: A Rough Calculation

Here is an analysis nobody seems to have run. On the anti-AI regulation side: SAG-AFTRA (160,000 members), the Writers Guild (11,500 members), the American Association for Justice (trial lawyers, 27,000 members), plus assorted advocacy groups. Call it 250,000 organized people with lobbyists, PACs, and media training.

On the pro-AI side: US Chamber data shows a majority of the 33.2 million US small businesses now use AI. Even at a conservative 40% adoption rate, that is 13.3 million small businesses. Add the estimated 7.2 million Americans using AI accessibility tools, the millions of patients served by FDA-approved AI diagnostics, and the 180 million monthly ChatGPT users. The unorganized pro-AI constituency outnumbers the organized opposition by roughly 500 to 1.

The problem is not numbers. It is organization. Anti-AI forces have a ratio of roughly one lobbyist per 200 members. Pro-AI beneficiaries have approximately zero lobbyists per 13 million users. The entire AI lobbying spend in 2025 (OpenAI's $3 million, Meta's $13.8 million, Google's and Microsoft's comparable figures) totals perhaps $50 to $60 million. The pharmaceutical industry spent $382 million on lobbying in 2023 alone. AI is bringing a business card to a knife fight.

The Strongest Counterargument

The strongest case for aggressive AI regulation is that the industry has earned it. Facebook promised "connecting the world" and delivered algorithmic radicalization. Theranos promised revolutionary diagnostics and delivered fraud. Uber promised to end drunk driving and delivered gig economy exploitation without benefits. Every technology industry has said "trust us" and every one has eventually required external regulation. The argument that AI is different because its practitioners are more ethical is exactly the argument every previous industry has made.

This is a fair point. The question is not whether AI should be regulated. It is whether regulation will be designed to prevent actual harm or to prevent theoretical harm while creating compliance industries. The nuclear scenario is not "no regulation." It is "regulation designed by people who don't understand the technology and aren't accountable for the innovation they prevent."

Limitations

This analysis treats regulatory outcomes as primarily a function of political organization, which is an oversimplification. Public opinion, specific incidents (an AI causing a high-profile death, for example), geopolitical events, and individual legislators' agendas all play unpredictable roles. The nuclear analogy is imperfect: nuclear power involves physical safety risks that AI (mostly) does not, and the regulatory capture dynamics differ. The lobbying spending figures cited are federal only; state-level spending is harder to track and may be proportionally larger. The EU compliance cost data comes from a single survey with acknowledged selection bias. And the 18-month timeframe in the headline is an editorial judgment, not a calculated deadline.

The Bottom Line

The technology is not the variable. Nuclear fission works. GMOs are safe. AI is useful. The variable is whether the people who benefit from the technology are organized before the people who fear it. Right now, AI's opponents have trade associations, PACs, and media-savvy spokespeople. AI's beneficiaries have a product they like and no idea a regulatory fight is happening on their behalf. Every month that gap persists, the bureaucracy hardens. The concrete sets. And the thing that could have been built in five years starts taking thirty.