Token Taxes: mitigating AGI's economic risks

This paper proposes "token taxes," a usage-based surcharge on AI inference enforceable through existing compute governance infrastructure, as a targeted mechanism to mitigate the severe economic risks and tax base erosion posed by the development of Artificial General Intelligence (AGI).

The Gist

Imagine the world is about to get a massive upgrade. We are talking about AGI (Artificial General Intelligence), a type of AI so smart it can do almost any job a human can do, and do it better and faster.

While scientists are busy making sure this super-AI doesn't accidentally destroy the world (the "capability" risk), this paper argues we need to worry about something else: the economy.

Here is the simple breakdown of the problem and the paper's proposed solution, using some everyday analogies.

🚨 The Problem: The "Robot Takeover" of Our Wallets

Think of the first Industrial Revolution (when steam engines and factories arrived). It made things cheaper and faster, but for a long time, regular workers didn't get richer. In fact, life got harder for many. This was called "Engel's Pause."

The authors say AGI will make this look like a gentle breeze compared to a hurricane. Here is why:

1. The Government's Empty Piggy Bank: Right now, governments get most of their money by taxing people's paychecks. If robots and AI do all the work, no one has a paycheck to tax. The government runs out of money to build schools, hospitals, and roads.
2. The "Rentier" Trap: Imagine a country that owns a giant oil field but doesn't need its citizens to work. The government gets rich from the oil, so they stop caring about the people. They might ignore poverty or bad living conditions because they don't need the citizens' votes or labor anymore. The authors fear AGI will turn every country into this kind of "oil state," where the government ignores its people because the AI does the work.
3. The Global Divide: Right now, only a few rich countries (like the US and China) own the super-computers needed to run AI. If they tax the AI, they get rich. Poorer countries that just use the AI get nothing. This makes the rich richer and the poor poorer.

💡 The Solution: The "Token Tax"

The paper suggests a new kind of tax called a Token Tax.

What is a "Token"?
Think of an AI conversation like a game of "Words for Money." Every time you ask the AI a question or it gives you an answer, it breaks the text down into tiny chunks called "tokens." You pay the AI company based on how many tokens you use.

The Idea:
Instead of taxing the company's profits (which they can hide) or the robots themselves (which is hard to count), the government should put a small sales tax on every single token used.

• Analogy: Imagine you go to a coffee shop. Usually, you pay for the coffee. Now, imagine the government adds a tiny "cup tax" to every cup sold. It doesn't matter who owns the coffee machine or where the beans are grown; if you drink the coffee, the tax is paid.

🛡️ Why is this better than other ideas?

The authors say this is the "Goldilocks" solution for two main reasons:

1. It's Hard to Cheat (The "Black Box" Problem)
AI companies are like magicians. They know exactly how many tokens they used, but the government only sees the final trick (the answer). Companies could lie and say, "Oh, we only used 10 tokens!" to pay less tax.

• The Fix: The paper suggests a 3-Step Audit Pipeline:
  • Step 1 (The Black Box): The cloud company (like AWS or Google Cloud) that actually runs the AI acts as a referee. They count the tokens for the government.
  • Step 2 (The "Norm" Rule): If the company tries to lie, the government says, "We know the average AI uses 1,000 tokens for this job. If you say you used 10, we will tax you as if you used 1,000."
  • Step 3 (The White Box): If they still argue, the government gets to peek inside the code to count the tokens manually.

2. It Fixes Global Inequality
Currently, if a company in the US builds an AI, the US government gets the tax money, even if a farmer in Kenya uses the AI.

• The Fix: With a Token Tax, the tax is collected where the AI is used. If the farmer in Kenya uses the AI, the tax money stays in Kenya (or goes to a global fund). This ensures that countries using the technology also get a share of the wealth, not just the countries building it.

🤔 What about the critics?

The paper addresses three big worries:

• "Won't this stop innovation?"
  • Response: Maybe. But we don't know yet. The authors suggest using Agent-Based Modeling (basically, a super-advanced video game simulation) to test how this tax would change the economy before we actually do it.
• "Why not tax the computer power (FLOPs) instead?"
  • Response: Taxing computer power is good, but it's hard to track. A Token Tax is better because it taxes the actual usage (the value created), not just the electricity burned. We could actually do both!
• "Won't the US and China just say 'No'?"
  • Response: Big countries might try to block this. But the authors suggest that a group of countries (like the EU) could band together and enforce this tax among themselves. If they are big enough, the US and China can't stop them, just like they couldn't stop the GDPR (privacy laws) from passing.

🏁 The Bottom Line

The paper argues that if we don't prepare for a world where AI does all the work, we risk a future where governments go broke, people lose their power, and inequality explodes.

The Token Tax is a simple, enforceable way to make sure that when AI creates wealth, that wealth is shared with society (via taxes) rather than just disappearing into the pockets of a few tech giants. It's like putting a toll booth on the highway of the future, ensuring everyone pays their fair share to keep the road open.

Technical Summary

Based on the paper "Token Taxes: Mitigating AGI's Economic Risks" by Lucas Irwin, Tung-Yu Wu, and Fazl Barez, here is a detailed technical summary covering the problem, methodology, key contributions, results, and significance.

1. Problem Statement

The paper addresses the severe economic risks posed by the advent of Artificial General Intelligence (AGI), specifically the potential for:

• Fiscal Crises: As AGI displaces human labor across all industries, governments will lose their primary revenue source (income taxes), leading to a collapse in public finance.
• Citizen Disempowerment: If states no longer rely on human labor for economic growth, they lose the incentive to respond to citizen needs, potentially creating "rentier states" where the government is funded by AI capital rather than human welfare.
• Global Inequality: The concentration of AI compute resources (the "Compute North") in developed nations (US, China) versus the reliance of developing nations ("Compute South") on renting these resources exacerbates global wealth divides.

The authors argue that current AI safety research focuses heavily on capability risks (e.g., alignment, existential threat) while neglecting these economic risks. Existing proposals like "robot taxes" (taxing firms based on automation) are deemed insufficient due to enforcement difficulties and their failure to address global inequality.

2. Methodology

The paper proposes a policy framework centered on Token Taxes and outlines a technical enforcement mechanism.

• Core Concept: A Token Tax is defined as a usage-based surcharge applied to model tokens at the point of sale (when the AI is used), rather than a tax on the firm hosting the model.
• Enforcement Architecture: The authors propose a Staged Audit Pipeline involving cloud compute providers as intermediaries between AI model providers and governments. This pipeline operates in three escalating stages:
  1. Black-Box Token Verification: Cloud providers (hyperscalers) act as independent record-keepers. They collect metadata on compute consumption and token-level usage data to verify reported usage against independent logs, preventing AI labs from under-reporting usage to maximize profits.
  2. Norm-Based Tax Rates: If black-box verification fails or data is missing, auditors apply a "norm tax." This involves charging companies a flat rate based on the empirically observed average token usage for that specific model category. This prevents profit inflation through strategic under-reporting.
  3. White-Box Audits: As a final escalation, companies are legally required to share internal generative process data with third-party auditors to confirm exact token counts, eliminating the ability to "game" the system.
• Analytical Approach: The paper advocates for the use of Agent-Based Modeling (ABM) to simulate the economic impacts of token taxes under various AGI growth scenarios, partnering technical governance researchers with economists.

3. Key Contributions

• Proposal of Token Taxes: The authors introduce the "token tax" as a superior alternative to traditional robot taxes. It shifts the tax base from capital ownership (where models are hosted) to consumption (where models are used).
• Technical Enforcement Framework: The paper provides a concrete, three-stage audit pipeline (Black-box $\to$ Norm-based $\to$ White-box) that leverages existing compute governance infrastructure to ensure compliance without requiring immediate access to proprietary model weights.
• Addressing Global Inequality: The paper demonstrates how token taxes can redistribute value. By taxing usage at the point of sale, countries in the "Compute South" (consumers of AI) can capture tax revenue, rather than revenue flowing solely to the "Compute North" (hosts of AI infrastructure).
• Geopolitical Strategy: The authors propose forming "coalitions of the willing" (e.g., the EU) to implement coordinated token taxes, making them harder for AI superpowers (US, China) to veto compared to unilateral national measures.

4. Results and Arguments

While the paper is a policy proposal rather than an empirical study with numerical results, it presents the following logical conclusions:

• Enforceability: Token taxes are technically feasible because cloud providers already possess the necessary metadata infrastructure to track inference usage.
• Tax Neutrality: By taxing token usage, the system restores "tax neutrality," ensuring that AI capital is not taxed at a significantly lower rate than human labor, thereby removing the incentive to replace humans purely for tax avoidance.
• Mitigation of Inequality: Unlike robot taxes which might only benefit the host country, token taxes ensure that the economic value generated by AI usage is captured by the consuming nation, mitigating the "resource curse" dynamic in a post-AGI world.
• Counter-Argument Rebuttal:
  • Innovation: The authors argue that ABM can be used to fine-tune rates to ensure innovation is not stifled.
  • FLOP Taxes: They argue token taxes are not mutually exclusive with FLOP (floating-point operation) taxes; a hybrid approach could leverage the enforceability of FLOP taxes with the equity benefits of token taxes.
  • Superpower Veto: Coordinated regional agreements can bypass the veto power of individual AI superpowers.

5. Significance

• Paradigm Shift in AI Safety: The paper shifts the focus of AI governance from purely technical safety (alignment) to socio-economic safety, arguing that economic destabilization is an immediate and urgent risk comparable to capability risks.
• Policy Innovation: It moves beyond abstract "robot tax" concepts to a specific, technically grounded mechanism (token tax) that aligns with the current architecture of Large Language Models (LLMs) and cloud computing.
• Global Equity: It offers a novel solution to the geopolitical disparity in AI, proposing a mechanism where developing nations can participate in the economic benefits of AGI rather than being left behind.
• Research Roadmap: It establishes a clear agenda for future research, specifically calling for the development of Agent-Based Models to simulate token tax impacts and the refinement of the audit pipeline.

In summary, the paper argues that Token Taxes are the most viable first-line defense against the economic collapse and inequality that could follow the deployment of AGI, offering a technically enforceable, usage-based solution that leverages existing cloud infrastructure to ensure fair distribution of AI-generated wealth.