https://arxiv.org/html/2503.04941v2

Both our model and those of Davidson and Korinek and Suh share a core assumption: AI systems trained with more compute become capable of automating increasingly complex tasks. This creates a direct link between advances in computing power and the expansion of AI’s economic impact through task automation.

A third key component of the automation process, is the process of reallocation of human labor from automated to yet-to-be automated tasks. Intuitively, as AI automation proceeds, the automation process can be expected to result in a significant increase in effective labor supply for automated tasks (more on this in Section 5 ). In a setting in which labor tasks are assumed to be complements in production (as we assume in GATE, see Section 5 ) a potentially significant determinant of the macroeconomic impacts of AI automation will be given by the extent to which the economy is “bottlenecked” by the non-automated tasks, in which effective labor supply can be expected to increase at a much slower rate than in automated tasks.

We employ the labor accumulation metaphor for two reasons. Firstly for simplicity, as it is easier to analyze than keeping track of additional factors of production and variable parameters governing the substitutability of different factors of production. Secondly, to emphasize what we believe is a key distinction between AI and previous automation technologies: the ability of AI to eventually automate all labor tasks currently performed by humans, leading to AI becoming a perfect substitute for all types of human labor once full automation occurs. This approach is of course of a metaphor, as AI development depends on processes of compute accumulation that are similar in nature to standard capital accumulation. Alternative modelling strategies, for instance modelling AI development as increasing the substitutability of human and AI labor at the level of each task would produce similar predictions. is the key mechanism through which AI development affects output in GATE, as we abstract away from any effects of AI development on TFP, which is exogenously defined and kept constant over time. 23 23 23

Another drawback of the baseline specification of GATE is that we have not allowed for any uncertainty affecting the mapping between AI system technical characteristics and the set of tasks that the system is able to perform. In other words, we have assumed that the social planner is omniscient about the path of AI labor automation. When this assumption is relaxed, uncertainty about the relationship between compute investment and automation capabilities can significantly alter investment decisions and economic trajectories. In such conditions, uncertainty about future automation capabilities can lead to more cautious investment decisions, thereby depressing the overall level of investment in AI.

The AI development module abstracts away from important non-compute inputs, particularly data availability, cost, and quality. While the compute-based framework captures many key dynamics of AI development, it omits potentially significant constraints around data collection, quantity, and curation requirements. This simplification could lead to overly optimistic automation timelines by not accounting for cases where data scarcity or quality issues become binding constraints, even with abundant compute resources—a concern that appears increasingly relevant given recent projections about the exhaustion of public training data ( Villalobos et al., 2022 ) . This concern may be especially salient for complex tasks that rely on real-world embodiment or physical experimentation, where data can be especially costly to gather. Future iterations of the model could benefit from explicitly incorporating data as a complementary input alongside compute, though this would inevitably introduce additional complexity into modeling the interplay between data requirements, collection costs, and AI automation capabilities.

See Autor et al. (2003) . The distinction of between cognitive and manual tasks is likely to be particularly consequential given the (to date) slower progress of AI systems in the physical realm and the plausible different production function for “digital workers” in the physical domain. Allowing for these additional dimensions of heterogeneity across labor tasks would be attractive, but would be computationally costly and highly speculative given the absence of high quality measurements for the relative economic value of physical tasks and of the key parameters of the production technologies applicable to robotics.

1e33 to 1e41 (1e36.5)

For example, Stewart 2021 reports that global chip revenues were around 0.4% of GWP in 2005. Under the GATE production function, using our median runtime compute requirements, this implies an automated fraction of 10%. Second, we examine the estimates of Acemoglu and Restrepo (2019) who investigated automation by relating the labor share, sectoral productivity changes and labor reallocation. Between 1987 and 2017 they found a net change in task content of production of approximately 10%. As such, we choose 10% as our median estimate for this parameter, such that runtime compute initially has some economic value. We account for around 2 × \times uncertainty in this parameter.