METR explains how task substitution changes measures of AI uplift

METR explains how task substitution changes measures of AI uplift

METR’s article was published May 8, 2026. It argues that estimates of AI’s productivity impact depend on which tasks are counted. METR distinguishes uplift on old tasks, uplift on new tasks, and uplift in value after people shift between tasks. Under simplifying assumptions, uplift on old tasks is no greater than uplift in value, which is no greater than uplift on new tasks. The distinction matters because people may choose tasks that AI makes much faster, even when those tasks add less value than their apparent time savings suggest. METR calls a task a “Cadillac Task” when someone does it because AI made it cheap and its uplift in value is much lower. Applying this framework, METR says an estimate based on old tasks might be a lower bound, but argues that the Claude-based estimate it discusses is likely an overestimate: speedups on selected tasks may not represent all tasks in the same task group. METR’s conclusion is that high uplift on new tasks need not mean equally high uplift in value. The inequalities depend on assumptions about people’s beliefs and choices; practical factors can make them fail. The article does not measure how much substitution within task groups occurs or how much it changes estimates of value uplift.

Publication: f9256cbb-4cd2-b95b-5f3d-b8cb471db8d3

attribution — pass: The reading attributes the framework, inequalities, critique of the Claude estimate and the term Cadillac Task to METR, the publisher of this single blog post. The arguments stay framed as METR's claims.

We have found it important to distinguish between three measures of AI’s uplift:

summary — uncertain: The summary is accurate but leaves central terms hard for a non-specialist to follow, and no definitions are supplied. It never says that old tasks means a typical pre-AI workday task mix and new tasks means the mix chosen once AI is available. It does not explain that task group refers to O*NET task categories. It also leaves the Claude-based estimate unidentified; the source attributes it to Tamkin and McCrory (2025), at roughly 17%. Smallest repair: one plain-language clause for each of these terms and a short identification of that estimate.

“New tasks” means the set of tasks you’d do in a typical day after AI is available.

date — pass: The reading gives May 8, 2026, which matches the UTC day of evidence.originalPublished and the URL path. The evidence kind is publisher-page, so no archive disclosure is needed. The reading presents no later events, uses no hindsight framing and does not rely on the source's present-tense remark about coding agents.

support — pass: Each claim matches the source. The inequalities are tied to simplifying assumptions and to maximizing behavior with accurate beliefs. The source says the old-task estimate is an expected lower bound but likely an overestimate because of within-group substitution, and the reading reports both points. Cadillac Task is defined as in the source, and the reading notes that practical factors could reverse the inequalities. The stated limitation about an unmeasured extent of substitution fairly reflects the article.

However it seems likely there is substantial substitution within task groups, towards those which have higher time-savings.

media — pass: No media was selected and no verified media candidates were supplied, so no media identity or relevance issue arises.