I’m analyzing what happens to the US economy in the short-term aftermath of the typical job being replaced by AIs and robots. Will there be a financial crisis? Short answer: yes.
This is partly inspired by my dissatisfaction with Tomas Pueyo’s analysis in If I Were King, How Would I Prepare for AI?.
Let’s say 50% of workers lose their jobs at the same time (around 2030), and they’re expected to be permanently unemployed. (I know this isn’t fully realistic. I’m starting with simple models and will add more realism later.)
I’ll assume that AI starts making the world more productive around the same time that this job loss occurs, but that big innovations such as a cheap cancer cures or the ability to conquer the world are still far enough in the future that financial markets aren’t ready to price them in.
These assumptions are designed to help me analyze the effects of job loss with minimal complications from other effects of AI. I’m focused here on the short-term financial and political consequences of job losses. There will be some radically different longer-term consequences, but I’m only analyzing those here to the extent that I expect markets to reflect them at the time of the job losses.
This post is merely an outline of what a rigorous analysis would look like. It’s good enough for informing my investment strategies, but not for persuading politicians to adopt better policies.
Note that this will be one of my least readable blog posts. Most of you should start by reading the conclusion, and only reading the rest if you’re tempted to argue with my conclusions.
If you still think my conclusions are wrong, you can find some more detailed explanations of my reasoning in this conversation with Gemini.
Note that I’m targeting this at readers with a significant background in finance. Please question the details of my analysis, and produce competing guesses based on answering similar questions.
Conclusions
I expect turmoil similar to that of the pandemic. My median guess is that it will be somewhat less sudden than the crash of March 2020, and that markets will mostly recover in one to two years (assuming we have years before something more dramatic happens).
The financial turmoil is likely to induce political instability. I find that hard to predict.
The US government will need to be more competent than it was during the pandemic in order to avoid hyperinflation or defaulting on its debt.
The magnitude of the turmoil will likely be heavily influenced by hard-to-predict expectations.
Maybe a bright spot is that a financial crash could slow capability advances at roughly a time of near-maximum risk related to AI alignment. But that might be offset by politicians being too distracted to do anything competent about alignment.
I’m surprised at how much my outlook fluctuated while writing this post, between optimism and despair, before settling on an intermediate mood.
The process of writing this post convinced me to (slowly) start selling my remaining (small) positions in bank stocks. I’ll be less willing to sell my stocks in gold mining companies. I’ll probably be more willing to sell some of my other stocks when I’ve guessed that they’ve reached bubble levels, rather than hoping to sell close to the peak.
Model 1: A Simplistic Utopia
Automation would increase productivity. Productivity drives increased wealth creation. The high level view suggests that in principle everyone could be wealthier and on average people would have dramatically more leisure time.
Model 2: Redistribution Problems
What happens to government budgets?
I’ll assume for Model 2 that $5 trillion that would have gone to annual wages goes instead to profits of US companies.
If those increased profits can be taxed at a 60% rate, that produces enough revenue ($3 trillion) to fund the resulting increased costs of unemployment insurance payments (assuming these are extended well past the usual 26 weeks), Medicare, and SNAP (food stamps).
Instead of the semi-utopian result I got in Model 1, this looks like a situation where mobs start lynching robots. And that’s with an overly optimistic assumption about the ability to collect taxes. A 60% rate would induce plenty of tax evasion. The limits on society’s ability to redistribute the money will cause significant headaches. Wages are taxed at a higher rate than capital because capital is more able to route around taxation.
Model 3: A More Careful Look at the Government Budget
Under existing tax rules, the vanishing wages will reduce tax revenues by roughly $1700 billion per year. About $1 trillion would be replaced by corporate income taxes, dividend taxes, and capital gain taxes.
I get that $1 trillion by assuming:
- an 18% effective tax rate on the $5 trillion increase in corporate income.
- I’ll guess that stocks held by US investors will appreciate by $10 trillion, but that only yields $40 billion.
- I’ll guess that dividends increase by $1 trillion, but that only yields $60 billion in increased taxes.
One reason that those last two look surprisingly low is that only 25% of stocks held in the US are held by taxable entities. The rest is in 401ks, IRAs, Pension Funds non-profits / universities. Another reason is that the capital gains tax is only collected on a small fraction of stock that’s actually sold in a given year – typically 4%, but I’ve assumed 6%.
So if we assume only minor political changes, the default outcome is a $700 billion decrease in tax revenues, and a goal of $3 trillion in increased safety net spending.
Model 4: Increased Deficit Spending
Surely the increased economic growth that results from the increased productivity will make it easy to finance a temporary deficit increase of $3 trillion to $4 trillion?
The answer seems to be an emphatic maybe.
Assume that government spending, other than interest payments on the debt, grows at 5%/year. Assume also that the economy grows at 30%/year, and financial markets are able to predict those growth rates. Tax revenues grow at the same rate as the economy after the initial plunge due to lack of wages.
See Could Advanced AI Drive Explosive Economic Growth? for where I got the guess of 30% / year growth. The 5%/year number is an even cruder guess as to what would be politically acceptable.
If the government could borrow at current interest rates, revenues would quickly grow to produce a surplus. But I expect businesses to be eager to borrow enough to drive interest rates to at least the economic growth rate, and maybe a bit higher.
I played around with a simulation (see the appendix). With a 31% interest rate, the budget achieves a surplus 16 years later, after spending several years with a deficit that’s more than twice revenues. That looks like a fragile scenario where one negative shock could cause bond investors to give up hope of bonds being a safe investment. So I’ll guess there’s more than a 50% chance of a debt spiral that prevents the government from continuing to borrow.
Model 5: Tax Increases
How much can tax increases narrow this deficit?
I’ll start by adding a modified version of a VAT. I’ll set the rate at 30%. Rates much above that level would trigger more of a black market than I’m comfortable with.
That yields $4.5 trillion/year. Alas, much of that comes from the newly unemployed, impairing my goal of ensuring that they mostly maintain their standard of living. I’ll have the government rebate $3 trillion of that revenue to everyone who files a tax return with income below a certain level – designed so the tax is only collecting money from 40% to 50% of the population. That nets $1.5 trillion.
Next: energy and compute excise taxes. There’s plenty of money that could in principle be collected by taxing the “robots” that replace human labor. But it’s hard to create a precise enough meaning of robot to avoid massive disputes as to what qualifies.
It should be possible to approximate a robot tax via excise taxes on electricity and compute FLOPS.
I suggest a $0.20/kWh tax on electricity that the grid delivers to commercial/industrial users, offset by a $5k rebate per human employee.
Also, a 200% excise tax on GPUs and similar hardware over some FLOPS threshold. Note that this requires unusually strict tracking of computing hardware in order to minimize tax evasion, with almost as drastic enforcement as would be needed to stop further capabilities progress.
Roughly $1.3 trillion could be collected this way. This seems like the right ballpark. There are many remaining questions that I will sweep under the rug.
How much can corporate income taxes be increased? History suggests that a 35% rate can be collected without drastic evasion, and we can minimize flight to lower-tax jurisdictions by promising to lower the tax rate as economic growth increases the tax base. On a tax base of $8 trillion, increasing the effective rate from 18% to 35% yields $1.36 trillion.
Finally, we can add a Land Value Tax. A 2% annual tax on roughly $25 trillion worth of land yields an extra $0.5 trillion. Why 2%? This tax causes declines in the value of land. Banks depend on land values to back up their mortgage holdings. A rate much over 2% would likely destroy the banking system’s solvency (unless it’s phased in quite gradually).
Adding those revenue increases all up:
- $1.5 trillion VAT (net) +
- $1.3 trillion excise taxes +
- $1.36 trillion corporate income +
- $0.5 trillion Land Value
equals $4.66 trillion, or nearly $1 trillion more than needed to cover the minimal needs of the safety net.
There are probably a few more ways that taxes could be increased. But it’s already pushing the limits of political feasibility and the limits on whether the taxes can be collected. The feasibility of this model depends heavily on improved governance.
Model 6: Bailing Out Banks
The limits of the Land Value Tax reminded me of another concern. Rising interest rates cause problems for the banking system. This isn’t directly connected to job losses, but both are the result of AI productivity changes that become pervasive enough to have economy-wide effects. So they’re likely to happen at roughly the same time. Interest rates won’t rise as fast as I’m assuming the job losses will happen, but this model is not too sensitive to how fast rates rise.
Banks have risks because of duration mismatch: they borrow from depositors with a promise that money can be withdrawn on short notice, and invest that money in mortgages and government bonds with a duration averaging maybe 6 years. That’s a good business model if interest rates are stable.
It fails if rates quickly increase from 4% to 30%. In this scenario, the market value of the banks’ mortgages and bonds drops at least 60%, leaving them very insolvent.
Recent history suggests that some banks can get away with paying depositors at a rate that’s 5% less than the prevailing interest rate. So maybe they can pay 25% on deposits when the prevailing rate is 30%. But they can’t afford to pay that much if their income averages 7% (and slowly rising) from their loans. So they’ll end up trying to pay depositors something like 5%, while money market funds pay close to 30%. That gap of 25 percentage points is likely to cause many depositors to try to move their money. Banks won’t be able to handle those withdrawals.
I don’t know what happens if deposit insurance runs out. I’ll bet that politicians aren’t willing to find out, and will bail out the banks. They’ll have, say, the Fed buy up to $18 trillion of banking assets at roughly face value, even though those assets have a market value closer to $6 trillion.
If the crisis happens slowly enough, the bailout could be spread over 12 years, the extra $1 trillion per year from Model 5 would pay for it: buy $1.5 trillion per year and sell it at $0.5 trillion. Alas, 12 years looks wildly optimistic to me.
We should be prepared for a scenario more along the lines of the interest rate increase happening in two years. That implies a roughly $10 trillion cost (18 – 6 for the bailout, offset by 2 from Model 5). A consolation is that it would be clearly a one-time cost (assuming banks don’t create more duration mismatched assets; I suppose that might be a somewhat optimistic assumption).
With 30% interest rates, borrowing that $10 trillion would mean an additional $3 trillion in annual costs added to the regular budget deficit. That sounds like a debt spiral.
What would likely happen instead is that the Fed would “print” enough money to monetize most of it. Creating money to finance a one-time expense is much less disastrous than monetizing a continuing budget deficit.
Monetizing $10 trillion would increase the monetary base by a bit under 200%. In comparison, the Fed increased the monetary base a bit over 150% between 2008 and 2010, while creating less inflation than they wanted. OTOH, doubling the monetary base during COVID caused an annoying burst of inflation. Gemini attributes the difference to the 2008 expansion being designed so that the money stayed in the banks, whereas in COVID it was designed to go to the average wallet. Whereas I’ll mainly emphasize differing expectations as to whether longer-term monetary policy would err in the direction of too little or too much inflation.
Neither of those two episodes will provide an adequate model for predicting an AI-related bank bailout. Several other effects complicate the analysis.
There would be an initial flight to safety, similar to what we saw in March 2020, of people wanting to hold more dollars. That has a deflationary effect – people don’t want to spend the dollars that the Fed prints. Alas, that didn’t last long in 2020, and I don’t see it lasting long after job loss turmoil.
My big concern is the Cagan effect: with growth driving nominal interest rates up to at least 30%, the cost of holding money increases a lot. So the velocity of money increases dramatically, implying that a non-inflationary monetary policy would be the opposite of printing money.
If the economy were in equilibrium, the 30% growth rate would mean that the Fed would be increasing the money supply by 30% to prevent deflation. That partly mitigates the Cagan effect.
A more speculative effect is that the Fed will become more able to credibly commit to zero inflation in a not-too-distant future. This would increase willingness to hold wealth in forms such as savings accounts, at least compared to a monetary policy that aims for 2% inflation. Or maybe it will become appropriate for monetary policy to aim for deflation. Currently, sticky wages create a need for inflation to exceed zero. That will cease being valuable in an economy where AI has substantially reduced the importance of wages.
There are significant limits to our ability to quantify any of those effects, so we’re left with only weak guesses about what they add up to.
Gemini initially guessed that the bailout would produce a one-time inflation of 15% to 20%. Then I reminded it of the Cagan effect, and it changed its guess to closer to 50%. I respect those guesses, but I see enough uncertainty about expectations that I’ll forecast a wider range of possibilities: anywhere from 5% to 75% seems plausible.
We should be able to manage if the financial turmoil is the only serious problem at the time. But it’s easy to imagine that other problems will be competing for political attention.
Model 7: 401k / IRA Effects
I now notice that withdrawals from 401k and IRA accounts of the newly unemployed will create more tax revenue for a couple of years. The newly unemployed might initially withdraw over 20% per year. That probably adds a bit under $0.5 trillion per year. I don’t have time to rewrite my earlier estimates to reflect that.
Sensitivity to Assumptions?
I’ve used plenty of numbers that are clearly just crude guesses. I’ll now outline what I expect if I vary those assumptions.
- Slower job losses.
I’ve assumed an implausibly fast job loss. It’s more likely that the fastest 50% of job losses will happen over a period of one to five years. The one year scenario looks pretty similar to my models. The 5 year scenario offers a modest hope that economic growth will produce more wealth fast enough to mitigate most of the financial troubles. Five years seems like it would require new regulations that are effective at preventing AIs and robots from competing with existing labor. That doesn’t look like the default, but it doesn’t seem too far-fetched.
- Benefits of Productivity.
I’ve been assuming a lag between the initial job losses and any large decline in the cost of living. I see plenty of uncertainty there.
My analysis would change a good deal if safety net spending could be reduced at rates of 25% per year without reducing recipients’ lifestyles. To elaborate on my prior analysis, 25% seems quite reasonable for families that were spending a lot on putting children through college. But there are probably more people whose costs are dominated by housing and medical expenses. I expect productivity gains there to lag significantly. So I lean toward near-term pessimism here.
- Lag in interest rate increase.
Interest rate changes tend to lag changes in economic growth. That potentially yields large short-term benefits, as tax revenues (other than on wages) increase before the government needs to pay high rates to borrow. In a highly optimistic scenario, that could reduce the budget problems by a couple of trillion dollars.
- Economic Growth Rates
My estimate of 30% growth rates is pretty arbitrary. I’m pretty sure I’d get roughly equivalent results from 20% growth or 50% growth. At much over 50%, $10 trillion costs start seeming small enough that I should probably focus on more important issues.
It’s somewhat hard to imagine a technology that could make 50% of jobs obsolete in a few years, but not cause growth above 15%. Such a scenario would reduce the cost of any bank bailouts, but make the other financial problems worse. Probably the net effect would be worse.
- Governance Quality
I want to emphasize that my analysis strongly depends on fairly optimistic assumptions about both actual government competence and market expectations of future competence.
I assume that politicians either anticipate the problems, or react quickly and near-optimally to them. I chose this assumption partly to simplify the analysis, and partly because I wanted to explore best-case possibilities. My median prediction is more pessimistic than this.
This optimistic scenario may require some new reliance on mildly-better-than-human forecasting by AIs. I’m fairly confident that AIs will have the necessary forecasting abilities. I’m weakly confident that politicians will pay some attention to those forecasts.
What worries me is that voters will be slow to learn the value of listening to those forecasts. It seems likely that voter pressure on politicians will reflect a confused mess of mediocre heuristics. I see medium-sized chances that the system will misjudge problems related to tax evasion, or try to get by without tax increases.
One way in which the outcome might be better than I expect is if AI improves government’s ability to collect taxes. AI auditing of corporate income tax returns could in theory enable a much more intricate tax code, where loopholes are more carefully targeted. It’s hard to see how that would be politically feasible at the right time, but maybe AIs are persuasive enough in their advice to enable it.
- Capital Gains?
One of my least confident assumptions is that stocks will appreciate by $10 trillion due to the $5 trillion in expected additional profits. I wouldn’t bet too heavily against the possibility of stocks appreciating more like $100 trillion, and the possibility that the government will find ways to tax that at much higher rates than is the current default. So I have a real, but very blurry, hope for trillions of dollars here, enough to solve most of the problems.
[Added 2026-01-11:
I asked for a more rigorous policy-oriented analysis, and now I’ve got one: Public Finance in the Age of AI: A Primer.
It’s dated a day before mine, but apparently published online a week later. That was a fast response.
It’s more theoretical than I want, with a somewhat excessive focus on measurable economic efficiency. I’m more concerned about risks associated with political turmoil. If we don’t kill ourselves, or get locked into a tyranny, we’ll be wealthy enough to only care a little how close we come to optimal efficiency.
They do a good job of recommending changes in taxation that governments should be preparing for.
They also suggest sovereign wealth funds. Those would be great if governments could establish them before markets bid up AI company stocks and interest rates. But the default outcome is that markets react sooner, leaving unclear benefits.
]
Appendix
Related Work:
I haven’t been able to find much work that seems similar to my analysis. The best is Robots: Curse or Blessing? A Basic Framework, by Jeffrey Sachs, Seth Benzell, and Guillermo LaGarda. The main limitation of their work is that it doesn’t look at the difficulty of collecting taxes. They focus somewhat on the possibility that robots are just barely more productive than human labor, which can lead to economic contraction. I expect robots to be more productive than humans by a significant margin.
The next closest work seems to be Should Robots Be Taxed? by Guerreiro, Rebelo, and Teles.
That paper seems somewhat confused about what jobs will be automated first. They assume that routine jobs will be automated first, and that non-routine jobs will continue to be done by humans (more productively) for at least a little while. They imply that the job loss will be slow enough that economic growth will enable tax revenues to increase naturally.
Whereas now, the CEO of OpenAI expects to automate his unusually-far-from-routine job at about the same time that I expect the main wave of job losses. I expect that the jobs that remain after this wave will be heavily regulated jobs and jobs that are subject to little competition – not the kind of jobs where wages will increase significantly due to productivity gains.
Python code for the simulation in Model 4:
def simulate_deficit(interest_rate, growth_rate, years=10):
# --- Initial Constraints (Trillions USD) ---
revenue = 2.5
primary_spending = 6.0 # Spending excluding interest
primary_spending_growth = 0.05
# Debt Structure
legacy_debt = 35.0 # Existing US Debt
legacy_rate = 0.03 # Avg rate on old debt (3%)
new_debt = 0.0 # Debt accumulated during simulation
# Maturity Constraint: How much old debt reprices every year?
rollover_fraction = 0.20
print(f"\n--- SIMULATION: Rates {interest_rate*100:.1f}% | Growth {growth_rate*100:.1f}% ---")
print(f"{'Year':<5} {'Revenue':<10} {'Spend':<10} {'Interest':<10} {'Tot Deficit':<12} {'Total Debt':<10} Legacy Debt")
print("-" * 65)
for year in range(1, years + 1):
# 1. Update Economics
revenue *= (1 + growth_rate)
primary_spending *= (1 + primary_spending_growth)
primary_deficit = primary_spending - revenue
# 2. Handle Debt Rollover
# A portion of legacy debt matures and becomes "New Debt" at the new rate
maturing_amount = legacy_debt * rollover_fraction
legacy_debt -= maturing_amount
new_debt += maturing_amount
# 3. Calculate Interest Bill
# Old debt pays 3%, New debt (plus rolled over debt) pays the new market rate
interest_bill = (legacy_debt * legacy_rate) + (new_debt * interest_rate)
# 4. Total Deficit & New Borrowing
# If Revenue > Spending, primary_deficit is negative (surplus), reducing debt need
total_deficit = primary_deficit + interest_bill
# All deficit must be funded by NEW debt
new_debt += total_deficit
total_debt = legacy_debt + new_debt
# Formatting for readability
print(f"{year:<5} ${revenue:<9.2f} ${primary_spending:<9.2f} ${interest_bill:<9.2f} ${total_deficit:<11.2f} ${total_debt:<10.1f} ${legacy_debt:<4.1f}")
# Stop if debt explodes to prevent infinite loop/messy numbers
if total_debt > 1000:
print("... [Debt Explosion: Simulation Stopped]")
break
if total_deficit < 0:
print("... [Budget Surplus Achieved!]")
break
Pingback: Compute Tax | Bayesian Investor Blog