When we talk about the impact of artificial intelligence on the workforce, the conversation often centers on a single, misleading number: the percentage of time saved. We hear that a developer is 'two times faster' at writing code, or that an administrative assistant handles emails 'four times faster' with a chatbot. However, a new study from the research organization METR argues that these numbers are fundamentally incomplete because they fail to account for how humans actually change their behavior when tools become cheaper or faster. The researchers distinguish between three measures of uplift: uplift on old tasks, uplift on new tasks, and uplift in value.

The core issue is 'task substitution'. Imagine a software engineer who previously spent their day balancing writing documentation and creating pull requests. If an AI suddenly makes writing pull requests much faster, the engineer does not simply finish their day early. Instead, they might use the extra time to write even more pull requests or take on new types of work entirely. Because the mix of tasks changes, measuring the speedup on 'old tasks'—the things you did before you had the tool—misses the bigger picture of how the overall value of your output has shifted.

The authors demonstrate that these three metrics follow a specific inequality: uplift on old tasks is often less than the uplift in value, which in turn is often less than the uplift on new tasks. This divergence is critical for any student looking at AI implementation. If you only measure how fast you are at your 'pre-AI' workload, you are likely underestimating the actual value gain. Conversely, looking only at 'new tasks'—which might include projects you never would have attempted without AI—can provide an inflated sense of productivity that ignores the actual time costs involved in the final output.

Perhaps most intriguing is the concept of 'Cadillac Tasks.' This term describes situations where AI makes a specific task so cheap that we suddenly do it all the time, even if the marginal value is low. In these cases, the uplift on new tasks is high, but the actual value added to the organization might be negligible. This suggests that businesses should look beyond simple speed metrics and instead analyze the elasticity of substitution—how easily one task can be swapped for another—when evaluating whether a tool is actually creating long-term business value.

For those entering the workforce, understanding this distinction is crucial. When you are asked to evaluate the success of an AI tool, do not just count the hours saved on your current to-do list. Instead, ask how the tool is changing the composition of your work. Are you simply doing more of the same, or are you utilizing the tool to perform tasks that were previously impossible? Navigating this shift is the true key to understanding the economic impact of the AI revolution, far beyond the hype of simple speed benchmarks found in the ONET database or typical chatbot evaluations.

When we talk about the impact of artificial intelligence on the workforce, the conversation often centers on a single, misleading number: the percentage of time saved. We hear that a developer is 'two times faster' at writing code, or that an administrative assistant handles emails 'four times faster' with a chatbot. However, a new study from the research organization METR argues that these numbers are fundamentally incomplete because they fail to account for how humans actually change their behavior when tools become cheaper or faster. The researchers distinguish between three measures of uplift: uplift on old tasks, uplift on new tasks, and uplift in value.

The core issue is 'task substitution'. Imagine a software engineer who previously spent their day balancing writing documentation and creating pull requests. If an AI suddenly makes writing pull requests much faster, the engineer does not simply finish their day early. Instead, they might use the extra time to write even more pull requests or take on new types of work entirely. Because the mix of tasks changes, measuring the speedup on 'old tasks'—the things you did before you had the tool—misses the bigger picture of how the overall value of your output has shifted.

The authors demonstrate that these three metrics follow a specific inequality: uplift on old tasks is often less than the uplift in value, which in turn is often less than the uplift on new tasks. This divergence is critical for any student looking at AI implementation. If you only measure how fast you are at your 'pre-AI' workload, you are likely underestimating the actual value gain. Conversely, looking only at 'new tasks'—which might include projects you never would have attempted without AI—can provide an inflated sense of productivity that ignores the actual time costs involved in the final output.

Perhaps most intriguing is the concept of 'Cadillac Tasks.' This term describes situations where AI makes a specific task so cheap that we suddenly do it all the time, even if the marginal value is low. In these cases, the uplift on new tasks is high, but the actual value added to the organization might be negligible. This suggests that businesses should look beyond simple speed metrics and instead analyze the elasticity of substitution—how easily one task can be swapped for another—when evaluating whether a tool is actually creating long-term business value.

For those entering the workforce, understanding this distinction is crucial. When you are asked to evaluate the success of an AI tool, do not just count the hours saved on your current to-do list. Instead, ask how the tool is changing the composition of your work. Are you simply doing more of the same, or are you utilizing the tool to perform tasks that were previously impossible? Navigating this shift is the true key to understanding the economic impact of the AI revolution, far beyond the hype of simple speed benchmarks found in the ONET database or typical chatbot evaluations.