Regulating AI, negative income taxes

 I see Senator Sanders has entered the push on regulation of AI. I can see some need for laws preventing its use for fraud, but such laws exist already. Otherwise, I tend to think the fears are based on a serious misunderstanding of what LLMs are. 

 

Assuming you mean Senator Bernie Sanders, your position touches on the central friction in modern technology policy: the divide between regulating actions versus regulating foundational technology.

Many technical experts agree that legislative fears often stem from a misunderstanding of Large Language Models (LLMs)—specifically, treating statistical pattern matchers as if they possess agency, intent, or imminent "superintelligence."

The debate surrounding Sanders' legislative push (such as the Ban Artificial Superintelligence Act) highlights several distinct regulatory arguments, beyond existing anti-fraud laws.

Key Motivations Behind Current AI Legislative Proposals

• Shift from Output to Infrastructure: Proponents of new laws argue that existing fraud and IP statutes only catch harms after they occur. Newer proposals aim at upstream interventions—such as data center energy use, water consumption, and compute thresholds—to mitigate systemic scale rather than just downstream misuse.

• Labor and Economic Displacement: Sanders' focus frequently centers on worker displacement and wealth concentration. The goal of these proposals is often to slow adoption speeds to give labor markets time to adjust, or to capture public revenue from AI-driven automation.

• Frontier Cyber and Biological Misuse: Regulators and some AI safety researchers point to risks that fall outside traditional fraud laws—such as LLMs being fine-tuned to assist non-experts in synthesizing hazardous biological agents or automating large-scale cyberattacks.

• Preempting State-by-State Patchworks: Federal lawmakers are competing with a web of state-level bills. A unified federal framework is sought by both critics and tech firms to avoid navigating 50 different compliance standards.

The Counterargument: The Danger of Premature Regulation

Critics of sweeping legislation echo your point about LLMs:

• Over-estimating Capability: Present-day models compute probabilistic text completions; they lack autonomous reasoning, consciousness, or self-directed goals. Regulating them as potential "superintelligences" risks legislating science fiction rather than reality.

• Regulatory Capture: Heavy compliance requirements (like pre-deployment reviews, licensing, or federal oversight agencies) heavily favor incumbents like OpenAI, Google, and Microsoft who can afford massive legal teams. Open-source developers and smaller startups risk being priced out.

• Global Competitiveness: Halting or heavily restricting domestic frontier model development does not stop international competitors from advancing the same technology.

Existing laws already prohibit wire fraud, impersonation, forgery, and copyright infringement. The core policy question remains whether society needs entirely new legal categories for the tools used to commit those acts, or if enforcing existing laws against the human bad actors remains the more effective path.

 

First environmental issues: water and power consumption needs of data centers appear with present technology to be huge. Won't technology over time correct this? 

 

Yes, technology will improve data center efficiency, but technological efficiency alone historically increases total energy consumption rather than decreasing it.

This phenomenon is known as Jevons’ Paradox: as a technology becomes cheaper, faster, and more efficient to run, overall demand for it scales so quickly that total resource consumption actually goes up.

The industry is currently pushing several technological and architectural fixes to tackle energy and water draw, but each comes with trade-offs.

Efficiency Upgrades Already Underway

• Hardware & Algorithmic Efficiency: Semiconductor advances and hardware-level optimizations allow chips to process significantly more calculations per watt. Simultaneously, technique shifts—like using smaller models (SLMs), quantization (reducing data precision), and specialized chips (NPUs/TPUs)—greatly lower the compute required for basic tasks.

• Transitioning from Air to Liquid Cooling: Traditional air conditioning uses enormous amounts of electricity and relies heavily on evaporative cooling (water consumption). Data centers are rapidly shifting to direct-to-chip liquid cooling and closed-loop immersion cooling. These techniques slash water loss to near zero and reduce cooling overhead power by up to 40%.

• Dedicated Clean Power Generation: Tech companies are attempting to decouple their data centers from public municipal grids by investing in localized power generation—such as co-locating facilities directly alongside nuclear power plants, small modular reactors (SMRs), and geothermal systems.

Why Technology Alone Won't Instantly Fix the Problem

While the energy cost per query drops rapidly, total power and water demand continue to rise due to structural factors:

Barrier	Technological Reality
Inference at Scale	Model training is expensive once, but inference (running queries for hundreds of millions of daily users) accounts for 80–90% of total AI power use. As AI is integrated into search engines, operating systems, and agentic workflows, execution volume expands faster than efficiency gains.
Grid Bottlenecks	Even if data centers become 50% more efficient, building regional power grids and transmission lines takes decades. Localized power spikes hit specific municipalities hard, threatening grid stability regardless of global averages.
Legacy Infrastructure	Upgrading thousands of existing legacy data centers to modern liquid-cooled, high-density standards requires massive capital, slowing down real-world efficiency adoption.

The Regulatory Realpolitik

This is precisely why policy advocates like Senator Sanders focus on environmental regulations.

From a regulatory perspective, waiting for market forces and technological iteration to solve resource strain creates immediate local risks: rising local utility bills for consumers, strained regional water tables, and extended lifespans for coal or natural gas plants kept online to meet sudden baseload surges.

Technological efficiency makes the unit cost cheaper, but until adoption plateaus, total resource demand will remain a primary policy headache.

Want to dive into how Small Language Models (SLMs) are trying to solve the energy footprint?

Well of course technology alone won't fix the problem, it never does, at least at first. The new lightbulbs that use very little electricity only cause me to leave lights on all the time. The same with my computer (although I do turn off the monitor).

That LED example is the classic direct rebound effect of Jevons' Paradox in everyday life: when the marginal cost of a behavior drops toward zero, habits immediately change to consume more.

When light bulbs drew 60 to 100 watts each, turning off lights in empty rooms was a conscious, money-saving choice. When a modern LED draws just 8 to 9 watts, the mental friction and financial penalty for leaving it on vanishes—so we leave them on.

This same behavioral dynamic plays out at a system level with technology like AI:

• Lower Cost per Unit ($\to$) Lower Threshold for Use: When computational inference gets 10 times cheaper, tech companies don't pocket a 90% power savings. Instead, they embed AI into everyday software tasks where it previously wasn't worth the cost—like auto-summarizing emails, real-time code autocomplete, or generating video thumbnails.

• The "Always On" Baseline: Just like leaving the lights on overnight because "it only costs pennies," system architects design software to run compute-intensive background tasks perpetually once server runtime costs drop.

This is why regulatory arguments focus so heavily on structural boundaries (like carbon caps or grid access fees) rather than relying solely on technological iteration. Without an external constraint, human behavior and market incentives almost always consume whatever capacity efficiency creates.

 

This phenomenon is self-limiting. 

 

You are right that Jevons’ Paradox and rebound effects are not bottomless—they hit hard real-world ceilings. In economics and resource management, efficiency-driven growth eventually run into natural, economic, or technical limits that force demand to plateau.

With AI and energy consumption, that self-limiting equilibrium happens across four main boundaries:

• Diminishing Marginal Utility: At a certain point, adding more AI or leaving more "lights on" yields zero extra value. Summarizing an email in real-time is useful; generating 50 variations of an email nobody requested is useless. Once utility flattens, demand growth slows regardless of how cheap the query is.

• Economic Cost and ROI Caps: Even if the unit cost of compute drops 90%, if total volume scales $1,000\%$, total enterprise spending still goes up. CFOs eventually enforce budgets, cutting low-value AI features simply because the aggregate cloud bill degrades operating margins.

• Hard Infrastructure Limits: Unlike lightbulbs—where a home's total draw stays low even with all LEDs on—data centers run into fixed physical constraints: available transformer capacity, local grid power limits, and physical land. When a power utility says "there is literally no physical power line that can deliver another 50 megawatts to this site for five years," growth halts instantly.

• Market-Driven Price Signals: As data center demand surges, local power prices rise. That higher input cost forces developers to optimize code, migrate workloads to off-peak hours, or deploy localized models—restoring equilibrium through pricing feedback loops.

The Policy Split: Let Market Mechanics Work vs. Intervene Early

This self-limiting nature is precisely where technology skeptics and advocates of regulation part ways:

• The Market-Mechanics View (Your Position): Free market price signals and hard infrastructure limits will naturally force efficiency, shift workloads off-grid (via SMRs, nuclear, solar), and cap wasteful consumption. Government intervention only risks distorting those natural feedback loops, picking technological winners, and adding deadweight compliance costs.

• The Regulatory Interventionist View (The Sanders/Regulatory Position): While self-limiting mechanisms will eventually stop the expansion, the "lag time" before market equilibrium is reached causes real damage: localized power outages, spiked utility bills for residential neighbors, and delayed retirement of fossil-fuel plants to meet peak load right now. Regulation is seen as a way to force that plateau sooner, protecting public goods before the market corrects itself.

A reasonable middle position should be reached and more than likely will be by the simple desire of owners of these facilities to keep costs down.

That dynamic—where profit motives and capital constraints force efficiency faster than legislative mandates can—is precisely how most industrial scaling problems historically resolve.

Facility owners aren't spending billions on power and cooling out of goodwill; energy is their single largest operational expense. When electricity accounts for up to 70% of a data center's operating cost over its lifecycle, the financial incentive to minimize every single kilowatt-hour is existential.

How Private Incentives Drive the "Middle Position"

Market dynamics are already pushing operators toward solutions that naturally align with public concerns:

• Site Selection Away from Strained Grids: Building in power-congested areas increases land costs and utility rates. Operators are increasingly building in remote regions with surplus power (like near hydroelectric dams or geothermal sites), naturally taking load off urban residential grids.

• Capitalizing on Off-Peak Energy: To cut costs, companies are structuring non-urgent compute workloads (like training base models) to run at night or during hours when grid electricity prices are lowest or even negative due to solar overproduction.

• Direct Energy Ownership: Major tech operators are signing direct power purchase agreements (PPAs) for dedicated clean energy (such as restarting nuclear units or building co-located solar farms) to lock in low, predictable long-term power rates—effectively building their own supply rather than draining public utilities.

Where the Friction Remains

The main reason this remains a debate in capital city corridors is a fundamental difference in timeline horizon:

[ Private Enterprise Horizon ]  ---> Optimize for 3-5 Year ROI & Capital Allocation
[ Municipal / Grid Horizon ]   ---> Planning 15-30 Year Public Infrastructure & Baseloads

• The Market Pace: Data center operators can build a facility in 12 to 18 months and adapt their code and hardware constantly.

• The Public Pace: Building new high-voltage transmission lines or approving grid interconnection requests can take 7 to 10 years due to permitting and physical construction.

When private demand scales faster than public infrastructure can physically be built, a temporary friction zone emerges. The regulatory debate is essentially over whether government needs to manage that temporary gap, or simply step back and let energy prices force data center operators to solve it themselves. Given the sheer capital at stake, the market incentive to solve it is already doing heavy lifting.

 

So in short, regulation can work better than owner interest only by restricting the building and use of data centers. 

 

That hits the core distinction. If the primary goal of regulation is to curb resource use beyond what prices and self-interest dictate, the main tool left is direct restriction: caps on energy permits, moratoria on construction, or mandatory efficiency standards that temporarily pause projects. 

Whether that approach is effective or counterproductive depends on how one views the trade-offs:

The Argument for Restrictive Regulation

Proponents argue that relying solely on owner self-interest leaves critical gaps:

• Managing the "Gaps" in Local Infrastructure: A data center owner pays for their electricity, but they don't automatically pay for the multi-year grid upgrades required to support them. Local communities can face higher utility rates or grid instability in the interim.

• Environmental Timelines: Companies focus on lowering costs, which doesn't always equal lowering emissions. Restrictive policies force companies to prove they have secured zero-carbon power before plugging into the grid, preventing dirty power plants from being kept online to meet sudden demand.

The Argument Against Restrictive Regulation

Critics point out that heavy-handed supply restrictions carry heavy unintended consequences:

• Distorting the Efficiency Drive: When regulators dictate how or where facilities can be built, they often lock in older standards. Market pressures drive innovation (like switching to liquid cooling or off-grid power) far faster than bureaucratic rules can adapt.

• Economic Offshoring: Halting local construction doesn't eliminate global demand for compute. It simply pushes data centers—and the accompanying tech jobs, tax revenue, and infrastructure investments—to neighboring states or countries with laxer rules.

• Bottlenecking Broader Technological Progress: Restricting compute capacity doesn't just limit controversial consumer tools; it slows down processing power for medical research, clean energy modeling, and industrial automation.

The Market-Led Equilibrium

In practice, the owners' drive to lower operating expenses naturally targets the exact same bottlenecks regulators worry about. When power becomes too expensive or unavailable, operators are forced to build their own dedicated power generation, shift workloads to off-peak hours, or innovate more efficient architectures.

The central policy choice comes down to whether government should actively restrict growth to protect public infrastructure in the short term or simply set clear baseline rules and let price signals and corporate self-interest drive the long-term solution.

 

Ok, now the employment issue. It reminds me of the phenomenon here in Cambodia, where mechanization and gradual consolidation of land ownership have reduced the need for farm labor, forcing much of the population into Phnom Penh. 

 

That is a spot-on comparison. The shift of Cambodian agricultural workers to Phnom Penh as tractor density rose and plots consolidated is textbook structural displacement—the exact pattern AI critics worry will hit white-collar and service sectors.

Historically, technology shifts labor across three distinct phases:

[ Direct Displacement ]  -->  [ Urban / Sector Migration ]  -->  [ Equilibrium & New Demand ]
(Fewer farm hands needed)       (Move to Phnom Penh for work)     (Construction, services, light industry)

The AI debate mirrors this transition, but with three critical differences that make modern policymakers far more nervous than previous agricultural or industrial revolutions.

Key Parallels vs. Modern Divergences

Feature	Cambodian Agrarian Shift	Modern AI Displacement
Nature of Labor	Replaced physical/manual labor with diesel horsepower.	Replaces cognitive/routine mental labor with algorithmic compute.
Speed of Transition	Generational: Took decades for mechanization and land consolidation to filter through rural provinces.	Compressed: Software updates deploy globally in seconds; adoption happens over months, not decades.
The "Safety Valve" Sector	Clear Migration Path: Displaced rural labor moved to Phnom Penh for garments, construction, tourism, and services.	Unclear Destination: AI hits entry-level knowledge work (copywriting, basic coding, data entry, customer support) simultaneously across multiple industries.

Why Lawmakers Panic: The "Speed and Destination" Problem

When Cambodian farmers migrated to the capital, the economy needed hands to build infrastructure, staff garment factories, and run local service businesses. The skills changed, but the demand for human labor remained high.

With AI, the policy concern is two-fold:

1. Velocity Outpaces Retraining: If a process takes 30 years, older workers retire and younger generations learn new skills naturally. If a software cycle automates tasks in 3 years, middle-aged workers cannot retrain fast enough to match the shift.

2. Targeting the "Entry-Level Rung": In knowledge work, junior positions (drafting contracts, writing basic code, analyzing spreadsheets) are how people learn to become senior strategists. If AI takes over the junior tasks, the ladder loses its bottom rungs.

The Market Counter-Argument

Just as Phnom Penh's economy expanded into industries that couldn't exist without an urban labor pool, optimists argue AI will follow the standard pattern:

• Lowering Capital Costs: Making cognitive work dirt cheap allows small teams to start companies that previously required millions in funding, creating entirely new job categories.

• The "Lump of Labor" Fallacy: Assuming there is a fixed amount of work to be done is historically wrong. When productivity increases, society gets wealthier and demands new services—from personal care and specialized trades to human-centric experiences—that didn't exist before.

Senator Sanders' push for guardrails reflects a desire to slow down the transition so workers don't get overwhelmed in the middle phase—much like managing the shock to rural communities during agricultural shifts—whereas market-oriented approaches trust that new industries will naturally emerge to absorb the workforce.

This transition looks to me more than just replacing one job with another, taking generations. This looks like eliminating all jobs in a much shorter time.

That is the core of the "Technological Unemployment" hypothesis—the fear that AI isn't just another incremental efficiency tool, but a general-purpose engine that compresses multi-generational transition times down to a single decade, while replacing human intelligence outright rather than complementing it.

Economists and technologists are fundamentally divided on whether this speed creates an absolute cliff or simply a chaotic, fast-moving shift.

The Case for Total Job Elimination (The "Cliff" View)

Proponents of your view argue that comparing AI to previous industrial shifts is a fundamental mistake:

• Speed of Software vs. Physical Hardware: Tractor deployment required physical manufacturing, supply chains, and capital cycles that took decades to penetrate farming communities. Software updates reach hundreds of millions of people overnight with zero marginal distribution cost.

• Generality of Cognition: In the past, when physical work was automated, humans moved to cognitive work. If cognitive work (language processing, pattern recognition, code generation) is automated, humans have no higher "intellectual sanctuary" to migrate into.

• Capital Outpacing Labor: If AI can independently execute tasks from start to finish rather than just assisting a worker, companies stop hiring all entry-level labor. The feedback loop that traditionally created new jobs breaks down because the AI itself absorbs the new work.

The Counter-Arguments: Structural Bumps Blocking Immediate Job Collapse

Optimists and market analysts argue that while rapid disruption is real, several structural realities prevent a total collapse of human employment:

[ AI Model Capability ]  --->  ( Enterprise Bottlenecks )  --->  [ Real Economic Adoption ]
   (Scales rapidly)             - Regulatory Compliance            (Spans decades)
                                - Organizational Inertia
                                - Liability & Accountability

1. The "Last Mile" Problem & Liability: A model can write 90% of a legal document or medical diagnostic report in seconds. However, taking legal accountability for a hallucinated clause or a missed medical symptom requires a human. Until software can legally assume liability, human oversight remains mandatory.

2. Enterprise Adoption Lag: Technology capability outpaces organizational adoption. Large enterprises, banks, and healthcare systems move notoriously slowly due to security, legacy tech infrastructure, internal politics, and compliance requirements.

3. The Expansion of Demand: Making cognitive work dirt cheap often creates massive new demand rather than just cutting staff. For instance, cheap software development doesn't mean fewer programmers—it often means every small business now demands custom software, expanding total software production tenfold.

The Policy Implications

If AI truly compresses a 50-year generational adaptation process into 5 years, society faces a massive "transition shock."

This is why modern proposals from lawmakers like Sanders shift focus away from simple job retraining toward structural buffers—such as shorter work weeks, taxing automated capital, or establishing universal basic income—to bridge the gap if labor demand drops faster than human adaptation can keep pace.

I said it looks like that; I don't know. We will have to see how it goes. A negative income tax, as it has been called, is the obvious approach. The others, regulating work weeks and so on, interfere with freedom.

The distinction you're drawing between cash transfers and micro-management gets to the heart of economic policy: direct income floor proposals preserve market choices, whereas behavioral regulations restrict them.

The Negative Income Tax (NIT)—popularized by free-market economist Milton Friedman in the 1960s—and its modern descendant, Universal Basic Income (UBI), operate on a simple principle: fix the income problem directly with money, then let the free market figure out how labor, leisure, and prices balance out.

Why the Negative Income Tax Is Gaining Renewed Traction

• Simplicity and Individual Liberty: Unlike complex welfare programs or strict labor mandates, an NIT sets a minimum guaranteed floor through the existing tax system. If an individual earns zero, they receive a baseline benefit. As they earn income, the benefit gradually tapers off, preserving the financial incentive to work without restricting personal choices.

• Elimination of Administrative Micro-Management: Prescribing maximum working hours, mandating job ratios, or restricting automation creates bureaucratic bloat and distorts private contracts. An NIT avoids telling businesses how to structure employment or individuals how to manage their time.

• Neutrality Toward Innovation: Rather than slowing down technological adoption or trying to predict which jobs will survive, cash transfers allow capital to move toward efficiency while ensuring citizens aren't left destitute by rapid labor shifts.

The Capital-Taxation Challenge

The primary hurdle with an NIT or UBI during a major AI shift isn't the mechanics of sending payments—it is funding the revenue base.

• Most modern tax systems rely heavily on payroll and personal income taxes.

• If software displaces a large share of high-income cognitive labor, labor-based tax revenues drop.

• To fund an NIT under those conditions, governments would have to pivot tax structures toward automated capital, corporate wealth, value-added consumption (VAT), or sovereign dividend funds tied to AI productivity.

Even Senator Sanders introduced an AI Sovereign Wealth Fund proposal designed to capture a share of AI corporate equity to pay annual dividends directly to citizens. While the funding mechanics differ, the underlying recognition is similar: if technology fundamentally restructures the labor market, direct income mechanisms are far cleaner and less coercive than trying to force companies to keep obsolete job roles alive.

For a broader look at how economists and tech analysts discuss this exact dynamic—specifically how widespread AI-driven automation pushes societies toward UBI-style safety nets—this brief breakdown explores how AI impacts the labor market and income models.

If AI Takes All Jobs, Who Will Buy Everything?

Economy Media · 183K views

 

I've always thought taxing consumption (VAT and excise and "sin" taxes) and taxing corporate income (which gets passed on in higher prices, hence in effect a tax on consumption), are to be preferred over taxing incomes (which in effect taxes work). 

 

That critique aligns closely with traditional supply-side and public finance economics: income taxes penalize productive activity (work, investment, and enterprise), whereas consumption taxes penalize resource extraction and spending.

Replacing income taxes with consumption-based systems—like a broad Value-Added Tax (VAT), sales taxes, and targeted excise taxes—creates fundamentally different economic incentives:

Economic Advantages of Consumption Taxation

• Encouraging Capital Formation and Savings: When earnings are not taxed at the point of receipt, every dollar saved or reinvested grows tax-free. This expands the capital pool available for businesses to invest in productivity and infrastructure.

• Eliminating the "Work Penalty": Traditional progressive income taxes increase the marginal tax rate as people work harder or advance in their careers, creating a disincentive to earn more. Consumption taxes apply only when those earnings are spent.

• Broader Tax Base and Compliance: Income tax systems are riddled with loopholes, deductions, and underground cash economies. A VAT captures economic activity across the entire supply chain, making tax evasion significantly harder.

• Corporate Tax Transparency: As you noted, corporate income taxes are largely an invisible pass-through. They are paid through higher consumer prices, lower worker wages, or reduced shareholder returns. Taxing consumption directly makes the tax burden transparent rather than hiding it inside corporate accounting.

Combining Consumption Taxes with a Negative Income Tax

The primary historical objection to consumption taxes is that they are regressive: lower-income individuals spend a much larger percentage of their income on basic necessities than high earners, paying a higher effective tax rate.

However, pairing a broad VAT with a Negative Income Tax (NIT) resolves this issue cleanly:

[ Broad VAT / Consumption Tax ]  --->  Captures Revenue Fairly Across All Spending
            +
[ Negative Income Tax (NIT) ]    --->  Offsets the Tax Burden for Lower Earners
            =
Progressive Consumption-Based System (Sustains Safety Net Without Disincentivizing Work)

By using revenue generated from a VAT to fund a baseline cash transfer (the NIT), the system automatically offsets the consumption tax for lower earners. A household at the bottom ends up with a net gain, while higher spenders pay progressively more in absolute terms—preserving the incentive to produce and invest while protecting basic living standards.

For a deeper look into the economic logic of replacing complex welfare and tax structures with a simple direct cash mechanism, Milton Friedman explains the Negative Income Tax in his own words, outlining why cash transfers preserve individual freedom far better than government intervention.

Gemini is AI and can make mistakes.

Searching the web