There’s been a lot of talk lately about an “AI explosion that will automate everything” to “AI will produce huge rents”. While it’s far from clear if any of these predictions will pan out, there’s a more grounded version of such questions we can quantitatively address:
Suppose AI automated every task that’s currently automatable — which is still not all jobs — and didn’t even create new ones. How capable would it need to be before its rents could fund something like a universal basic income (UBI)?
It turns out one can, in fact, give a clean, analytic threshold for this question.
Off the bat, it’s perhaps worth mentioning that this is explicitly a funding result, and not a welfare result. I’m not arguing whether UBI is good/bad, or how it should be distributed, or how society should handle purpose/meaning in a post-work world. In fact, we don’t even have to assume the world is fully post-work in our treatment here – we can base ourselves on current U.S. economic parameters. Our goal is just to identify whether AI rents can become large enough to cover basic needs, and to explicitly identify what societal levers can make this threshold more or less feasible.
Below I give the broad strokes. For full technical details, proofs, and calibrations, see my paper:
In a Solow-Zeira task model where AI boosts the productivity of automated tasks by a factor γ, the minimum capability needed to fund a transfer B (as a share of GDP) from AI rents alone is:
γ≥(B/YΘ(1−c)αˉ1−ρAρκˉρ)σ,ρ=σ−1σ<0.
Where the meaningful levers are:
Θ: public revenue share of AI capital (e.g. via taxes/ownership)
c: AI operating + alignment costs
ᾱ: fraction of tasks that are automated by AI
σ < 1: CES curvature (capturing the Baumol “cost-disease” effect that indefinitely producing more goods doesn’t lead to infinite economic output)
κ̄: capital-output ratio
B/Y: size of the transfer
What happens when we plug in real numbers?
Calibrating to U.S. data (sources in the paper):
UBI ≈ $12k/year/adult → B/Y ≈ 11%
Θ ≈ 14.5% (effective public capture of AI-capital profits)
ᾱ ≈ 0.42, σ ≈ 0.66, s ≈ 0.22, δ ≈ 5.6%, c ≈ 0.6
This gives:
Based on current estimates, AI needs to be ~5-7× as productive as today’s automation on automatable tasks to fund a UBI-sized transfer.
Not 50×. Not 500×. About 5-7× beyond current automation productivity.
And that’s a worst-case upper bound—after all, we assumed:
no new tasks,
no labor complementarity,
no spillover productivity,
no endogenous growth.
In the paper, we analyze all of the above and show they would lower the required UBI AI capability threshold γ.
When might we hit 5-7×?
Figure 1. AI capability doubling trajectories vs. UBI threshold.
Figure 1 shows AI reaching γ in:
~2028 with 1-year doubling of AI capabilities
~2031 with 2-year doubling of AI capabilities
~2038 with 5-year doubling of AI capabilities
~2052 with 10-year doubling of AI capabilities
Even very conservative AI capability growth scenarios cross the bar by mid-century.
In Figure 2, we analyze monopolistic or concentrated oligopolistic markets and show that they reduce the threshold by increasing economic rents, whereas heightened competition between AI firms significantly raises it. We’ll skip this for the sake of keeping this post short, but feel free to consult the paper, especially the theorem that supports it in Proposition 2.
The strongest lever: public revenue share (Θ), not nationalization
The cleanest comparative static is that:
Increasing public revenue share (Θ) of AI profits from ~15% to ~33% cuts γ roughly in half.
Note, this doesn’t require full-scale nationalization; rather, it mirrors the difference between current U.S. and Scandinavian-style corporate-profit capture. As we can see in Figure 3 below, beyond ~50% public ownership, there are diminishing returns — unless operating/AI alignment costs c rise.
This makes Θ the most practical policy knob for reducing how much AI capability is needed to fund a basic transfer.
Figure 3 (from the paper). Public revenue share (Θ) vs. AI capability threshold (γ) under low and high cost regimes (c).
Figure 3 shows:
At Θ ≈ 14.5%, γ ≈ 5× (low cost) or 8× (high cost).
At Θ ≈ 33%, γ drops to 3× (low cost).
High alignment/regulatory/compute overhead (high c) pushes γ up.
Beyond Θ ≈ 50% public ownership, the curve flattens.
The economic message is simple: capturing more of the rents makes an AI based UBI more feasible without requiring nationalization, unless operating and alignment more capable AI systems becomes too costly.
Why this matters even if you don’t like UBI
This result isn’t an endorsement of UBI. It’s a feasibility benchmark.
If AI-generated rents exceed this threshold, society would have the financial slack to fund a basic transfer — or any number of alternative redistributive mechanisms (sovereign wealth dividends, tax credits, negative income tax), depending on political taste.
The point is:
AI capabilities don’t need to grow very far to support a UBI even without creating new jobs.
Whether one likes or dislikes that fact, we think it’s useful to know where the tipping point is.
Try the interactive version
I believe the analytic AI UBI threshold (Proposition 1, and associated comparative statics in Corollary 1) is the more lasting contribution than any specific numbers I plug in here. I’ve made my code available here, so anyone can try their own calibrations.
There’s also a public tool made by the AI+Wellbeing Institute that implements my model based on the code above, connecting UBI feasibility to certain well-being measures.
It’s fun to play with and makes the comparative statics more intuitive. Try it out for yourself!
An AI Capability Threshold for Funding a UBI (Even If No New Jobs Are Created)
There’s been a lot of talk lately about an “AI explosion that will automate everything” to “AI will produce huge rents”. While it’s far from clear if any of these predictions will pan out, there’s a more grounded version of such questions we can quantitatively address:
Suppose AI automated every task that’s currently automatable — which is still not all jobs — and didn’t even create new ones. How capable would it need to be before its rents could fund something like a universal basic income (UBI)?
It turns out one can, in fact, give a clean, analytic threshold for this question.
Off the bat, it’s perhaps worth mentioning that this is explicitly a funding result, and not a welfare result. I’m not arguing whether UBI is good/bad, or how it should be distributed, or how society should handle purpose/meaning in a post-work world. In fact, we don’t even have to assume the world is fully post-work in our treatment here – we can base ourselves on current U.S. economic parameters. Our goal is just to identify whether AI rents can become large enough to cover basic needs, and to explicitly identify what societal levers can make this threshold more or less feasible.
Below I give the broad strokes. For full technical details, proofs, and calibrations, see my paper:
“An AI Capability Threshold for Rent-Funded UBI in an AI-Automated Economy” https://arxiv.org/abs/2505.18687
The core result in one line
In a Solow-Zeira task model where AI boosts the productivity of automated tasks by a factor γ, the minimum capability needed to fund a transfer B (as a share of GDP) from AI rents alone is:
γ≥(B/YΘ(1−c)αˉ1−ρAρκˉρ)σ,ρ=σ−1σ<0.
Where the meaningful levers are:
Θ: public revenue share of AI capital (e.g. via taxes/ownership)
c: AI operating + alignment costs
ᾱ: fraction of tasks that are automated by AI
σ < 1: CES curvature (capturing the Baumol “cost-disease” effect that indefinitely producing more goods doesn’t lead to infinite economic output)
κ̄: capital-output ratio
B/Y: size of the transfer
What happens when we plug in real numbers?
Calibrating to U.S. data (sources in the paper):
UBI ≈ $12k/year/adult → B/Y ≈ 11%
Θ ≈ 14.5% (effective public capture of AI-capital profits)
ᾱ ≈ 0.42, σ ≈ 0.66, s ≈ 0.22, δ ≈ 5.6%, c ≈ 0.6
This gives:
Not 50×. Not 500×. About 5-7× beyond current automation productivity.
And that’s a worst-case upper bound—after all, we assumed:
no new tasks,
no labor complementarity,
no spillover productivity,
no endogenous growth.
In the paper, we analyze all of the above and show they would lower the required UBI AI capability threshold γ.
When might we hit 5-7×?
Figure 1. AI capability doubling trajectories vs. UBI threshold.
Figure 1 shows AI reaching γ in:
~2028 with 1-year doubling of AI capabilities
~2031 with 2-year doubling of AI capabilities
~2038 with 5-year doubling of AI capabilities
~2052 with 10-year doubling of AI capabilities
Even very conservative AI capability growth scenarios cross the bar by mid-century.
In Figure 2, we analyze monopolistic or concentrated oligopolistic markets and show that they reduce the threshold by increasing economic rents, whereas heightened competition between AI firms significantly raises it. We’ll skip this for the sake of keeping this post short, but feel free to consult the paper, especially the theorem that supports it in Proposition 2.
The strongest lever: public revenue share (Θ), not nationalization
The cleanest comparative static is that:
Note, this doesn’t require full-scale nationalization; rather, it mirrors the difference between current U.S. and Scandinavian-style corporate-profit capture. As we can see in Figure 3 below, beyond ~50% public ownership, there are diminishing returns — unless operating/AI alignment costs c rise.
This makes Θ the most practical policy knob for reducing how much AI capability is needed to fund a basic transfer.
Figure 3 (from the paper). Public revenue share (Θ) vs. AI capability threshold (γ) under low and high cost regimes (c).
Figure 3 shows:
At Θ ≈ 14.5%, γ ≈ 5× (low cost) or 8× (high cost).
At Θ ≈ 33%, γ drops to 3× (low cost).
High alignment/regulatory/compute overhead (high c) pushes γ up.
Beyond Θ ≈ 50% public ownership, the curve flattens.
Why this matters even if you don’t like UBI
This result isn’t an endorsement of UBI. It’s a feasibility benchmark.
If AI-generated rents exceed this threshold, society would have the financial slack to fund a basic transfer — or any number of alternative redistributive mechanisms (sovereign wealth dividends, tax credits, negative income tax), depending on political taste.
The point is:
Whether one likes or dislikes that fact, we think it’s useful to know where the tipping point is.
Try the interactive version
I believe the analytic AI UBI threshold (Proposition 1, and associated comparative statics in Corollary 1) is the more lasting contribution than any specific numbers I plug in here. I’ve made my code available here, so anyone can try their own calibrations.
There’s also a public tool made by the AI+Wellbeing Institute that implements my model based on the code above, connecting UBI feasibility to certain well-being measures.
It’s fun to play with and makes the comparative statics more intuitive. Try it out for yourself!