How input assumptions shape predictions of economic growth from AI
We created this tool to let you be part of the discussion. Change the inputs according to what you think is most relevant and determine TFP growth within the task-based model from Acemoglu (2024). As AI research advances, estimates may change. Therefore, this tool is built to be updated, and we invite you to share new studies with us.
We also present other models and estimates of TFP, Labor Productivity, and GDP Growth, demonstrating the vast array of opinions on these subjects.
| Component | Acemoglu | Aghion & Bunel | Korinek et al. | ||
|---|---|---|---|---|---|
| Modest | Substantial | Extreme | |||
| GDP Share of Exposed Tasks | 20% | 60% | 20% | 30% | 50% |
| affected mass of tasks | |||||
| Share of Exposed Tasks That Are Profitable to Automate | 23% | 50% | 20% | 40% | 60% |
| diffusion share of instances | |||||
| Average Cost Savings from AI | 27% | 40% | 0.30 | 0.45 | 0.80 |
| log unit-cost decline per instance | |||||
| AI-Exposure-Adjusted Labor Share of Value Added | 53.5% | 57% | 60% | 60% | 60% |
| base-period labor share, unadjusted | |||||
| Resulting TFP Gain | 0.07pp/yr0.66% / 10 yrs | 0.68pp/yr6.8% / 10 yrs | 0.12pp/yr0.7% / 6 yrs | 0.51pp/yr3.1% / 6 yrs | 2.10pp/yr13.4% / 6 yrs |
The labels in orange are Korinek et al.’s inputs for three scenarios of AI growth. The differences with Acemoglu’s and Aghion & Bunel’s shared definitions are explained below. This section uses Korinek et al.’s first-order approximation, which yields cumulative results close to their main model; the equation the simulation actually runs on is discussed in Alternative Scenarios Considered by the Economists.
Korinek et al. do not publish an annualized rate. They report cumulative gains over six years, so the annual figures shown here are ours, computed as ln(1 + gap)/6:
— Applied to their published cumulative figures of 0.7 / 3.1 / 13.4%: 0.12 / 0.51 / 2.10 pp per year.
— Applied to the first-order product of the four inputs, 0.72 / 3.24 / 14.40%: 0.12 / 0.53 / 2.25 pp per year.
The published figures come from the exact simulation and sit slightly below the first-order product.
The dials below can be set at Acemoglu's, Aghion & Bunel's, Korinek et al.'s or your own preferred values. You can change between Acemoglu / Aghion & Bunel's and Korinek et al.'s frameworks by choosing your preferred definition's button. We also include an estimate horizon dial that allows you to choose over what period the changes in TFP accumulate.
How do we measure an occupation's exposure to AI?
ACEMOGLU
20%
From Eloundou et al. (2024) and combined into occupations using wage bills from BLS, resulting in bill weighted share of exposed occupations. Interpreted as GDP share of tasks exposed to AI.
AGHION & BUNEL
60%
From Pizzinelli et al. (2023), who measure AI exposure by abilities. Abilities are weighted by prevalence and importance and aggregated to occupations, resulting in AI occupational exposure (AIOE) by occupation and employment shares by country. The 60% is US-specific and includes both high and low complementarity employment.
KORINEK ET AL.
20%/30%/50%
Modest / Substantial / Extreme
affected mass of tasks
Korinek et al. define this figure as “the fraction of the economy's tasks that AI affects, at pre-AI wage-bill weights.” The range 20%, 30%, and 50% represent the three different measures of task exposure in Eloundou et al. (2024) from least to most intense.
Acemoglu uses task exposure while Aghion & Bunel use ability exposure. Aghion & Bunel flag occupations above the AIOE median as exposed to AI; Acemoglu uses each occupation's share of exposed tasks, then weights by wage bill to get a GDP share. Aghion & Bunel apply no weighting, assuming the GDP share of tasks exposed to AI is similar in magnitude to the share of tasks exposed to AI (despite using abilities).
In the first-order form above, Korinek does not weight the share of tasks with BLS data like Acemoglu, but defines the values as wage-bill weighted, acknowledging this as a step.
In Korinek et al.'s main specification, the share of affected tasks starts from the mid-2026 reading in all scenarios, follows a logistic path, and ends in different 2030 values depending on the scenario.
How many tasks are worth automating?
ACEMOGLU
23%
From Svanberg et al. (2024).
AGHION & BUNEL
50%
From Svanberg et al. (2024) cost-decline scenarios under a 22% annual decline from Besiroglu & Hobbhahn (2022).
KORINEK ET AL.
20%/40%/60%
Modest / Substantial / Extreme
diffusion share of instances
Represents the percentage of task instances that AI affects. The numbers are picked by the researchers but the upper bound is informed by Bick et al. 2024 and the lower bound is informed by (Griliches 1957; David 1990; Comin and Hobijn 2010).
Aghion & Bunel incorporate projected computing-cost declines. Acemoglu excludes them from his baseline, saying a doubling of GPU capacity “will not necessarily translate into a 50% decline in costs, in general,” citing diminishing returns, bottlenecks from other inputs, and current architectural limits.
We present Korinek et al.'s diffusion share in this section because it occupies the same place as Acemoglu's and Aghion & Bunel's share of exposed tasks profitable to automate. However, it should be noted these inputs represent different considerations. Diffusion focuses on the amount of people using AI in a task for whatever reason, whereas Acemoglu and Aghion focus on a task's profitability with AI.
We note that Korinek et al. consider diffusion at the task instance level, not the task level.
In Korinek et al.'s main specification, the diffusion starts from the mid-2026 reading in all scenarios, follows a logistic path, and ends in different 2030 values depending on the scenario.
How much does AI save in labor costs?
Korinek et al.'s values are log points, not percentages. They are equivalent to productivity gains of 35%, 57% and 123%.
| Include | Study | Value | Used by | Article |
|---|---|---|---|---|
Noy and Zhang (2023) Writing tasks assisted with Chat GPT |
40%speed | AcemogluA&BPWBM | Link → | |
|
Brynjolfsson et al. (2023) Customer support agents assisted with generative AI tool |
14%Issue resolution rate | AcemogluA&BPWBMOECD | Link → | |
Peng et al. (2023) Software developers implementing HTTP servers assisted with AI pair programmer |
56%speed | A&BPWBM | Link → | |
Cui et al. (2025) Software developers using AI-based coding assistant |
26%Task completion rate | PWBM | Link → | |
|
Dell'Acqua et al. (2023) Consultants using GPT-4 to complete tasks |
12%Task completion rate | PWBMOECD | Link → | |
|
Haslberger, Gingrich & Bhatia (2023) Experiment with sample of UK Working Age Population using ChatGPT for work tasks *Calculated by Filippucci et al. (2024), not reported in paper |
48%Average gain in time and quality | OECD | Link → | |
Gambacorta et al. (2026) Software developers using CodeFuse, LLM for coding tasks. |
22%Task completion rate | Link → |
Mean of 2 selected: 27%
ACEMOGLU
27%
Labor cost reductions from AI averaged from two experimental studies: Noy & Zhang (2023) and Brynjolfsson et al. (2023).
AGHION & BUNEL
40%
Labor cost reductions from AI averaged from three experimental studies: Noy & Zhang (2023), Brynjolfsson et al. (2023), and Peng et al. (2023).
KORINEK ET AL.
0.3/0.45/0.8
Modest / Substantial / Extreme
log unit-cost decline per instance
Represents the unit-cost decline per AI-performed instance, whether automated or augmented, also referred to as the log productivity gain per AI-performed instance.
We display it in two forms: 1.) in log points (as Korinek et al. have it in their paper) and 2.) converted increase in productivity (%).
The numbers are picked by the researchers but informed by three studies: (Brynjolfsson et al. 2025; Huang et al. 2025; Demirer et al. 2026)
Aghion & Bunel include Peng et al., which Acemoglu excludes because he believes their setting (computer programmers on GitHub Copilot) is “less likely to be relevant to other tasks and occupations.”
We note Korinek et al.'s difference in units, as their model uses log points and not % increase in productivity.
In Korinek et al.'s main specification, the unit cost decline starts from a mid-2026 reading (0.3, 0.35, 0.45 depending on the scenario), grows in the case of the substantial and extreme scenario, and ends at 0.45 and 0.8 for these scenarios respectively.
How do labor cost savings become overall cost savings?
ACEMOGLU
53.5%
Industry-level AI exposure estimates from Eloundou et al. (2024) combined with industry labor shares from BEA.
AGHION & BUNEL
57%
Value comes from Acemoglu's April 2024 draft.
KORINEK ET AL.
60%
base-period labor share, unadjusted
Korinek et al. use a base-period labor share in all scenarios, which they refer to as the conventional value for the US.
No difference in methodology between Acemoglu and Aghion & Bunel. Aghion & Bunel use value from Acemoglu's April draft.
Korinek et al. refers to labor share as labor share of income, not value added. It is also not adjusted for AI exposure.
In the main specification, labor share of income is endogenous and falls over time due to AI.
Over what period do these changes accumulate?
This section allows you to explore other considerations the economists make in their papers.
Acemoglu's variants
Aghion & Bunel's intervals
The two estimates above are part of a growing field of literature. Select growth metrics below to see how other studies predict AI will change the economy. We are updating this page as new studies take place.
Select one to compare
Baseline is weighted average calculated from 2026 BLS News Release Table A.
Send us a study, and we will consider it for the number line above. Only the first two fields are required.