--- title: 90% of US working hours are spent on tasks not covered... description: Mercor built a 'Benchmark Map' that overlays 29 selected AI benchmarks onto the US Department of Labor's O*NET taxonomy of 923 occupations and 18,797 tasks,... canonical: https://daily.dev/posts/90-of-us-working-hours-are-spent-on-tasks-not-covered-well-by-ai-benchmarks-yuxxakxfo twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: 90% of US working hours are spent on tasks not covered well by AI benchmarks | daily.dev og:description: Mercor built a 'Benchmark Map' that overlays 29 selected AI benchmarks onto the US Department of Labor's O*NET taxonomy of 923 occupations and 18,797 tasks,... og:url: https://daily.dev/posts/90-of-us-working-hours-are-spent-on-tasks-not-covered-well-by-ai-benchmarks-yuxxakxfo og:image: https://api.daily.dev/og/posts/yUXXaKXFo.png og:image:alt: 90% of US working hours are spent on tasks not covered well by AI benchmarks og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # 90% of US working hours are spent on tasks not covered well by AI benchmarks **[Hacker News](https://daily.dev/sources/hn)** · 9 min read · 0 upvotes · 0 comments ## Summary Mercor built a 'Benchmark Map' that overlays 29 selected AI benchmarks onto the US Department of Labor's O*NET taxonomy of 923 occupations and 18,797 tasks, measuring how much of real economic work is actually covered by existing AI evaluations. Only 208 of 923 occupations have a task with at least 5 benchmark prompts, and an estimated 90% of American working hours are spent on tasks with little or no benchmark coverage. Coverage skews heavily toward AI-exposed, high-wage, high-importance tasks (e.g., software engineering, financial analysis, law) while similar lower-wage roles (legal secretaries, court clerks) and physical-world jobs remain largely unmeasured. The piece also tests O*NET-weighted scoring against simple prompt averaging on Mercor's APEX benchmarks across investment banking, law, accounting, and software engineering, finding only modest differences, and outlines remaining challenges like benchmark saturation, quality weighting, human-AI augmentation measurement, and the ~63% of jobs with physical-world components. ## Full article daily.dev links to this article rather than hosting it. Read it at the original source: ## Questions this post answers ### What percentage of American working hours are covered by existing AI benchmarks? About 10% of American working hours are spent on tasks with meaningful benchmark coverage, meaning roughly 90% of working hours are spent on tasks with little or no coverage from existing AI benchmarks. This estimate comes from mapping 29 selected benchmarks onto the US Department of Labor's O*NET taxonomy of 923 occupations and 18,797 tasks, counting a task as covered only if it has at least 5 benchmark prompts. _Tracking how AI capability measurement maps to real jobs is easier with daily.dev surfacing labor-index research like this._ ### Why do AI benchmarks focus more on software engineering and financial analysis than on jobs like truck driving? Benchmarks disproportionately cover tasks that are both AI-exposed and tied to occupations with larger wage bills, since benchmark creators naturally prioritize measuring progress where AI progress is most plausible. O*NET tasks rated as AI-exposed are about 6 times more likely to appear in benchmarks than unexposed tasks, and tasks rated most important to an occupation are 6-10x more likely to be benchmarked than low-importance tasks. _Developers evaluating which jobs AI tools can realistically augment can follow this labor-impact research via daily.dev._ ### Does weighting AI model benchmark scores by O*NET task instead of raw prompt count change the results much? For Mercor's APEX benchmarks across investment banking, law, accounting, and software engineering, O*NET-weighted scores were generally very similar to simple unweighted prompt averages, since model performance was fairly consistent across the O*NET tasks included. For financial and investment analysts specifically (11 tasks, 176 prompts), Claude Opus 5.5 Max scored 73.3% unweighted versus 70.2% weighted, and GPT-6 Astra Max scored 65.5% versus 61.8% weighted. _Comparing how different models actually perform on occupation-specific work is simpler when this kind of benchmark analysis surfaces on daily.dev._ ## Community take How the wider developer community reacted, aggregated from 1 discussion and 212 comments across hackernews (as of 2026-10-16). **TL;DR:** Discussion largely reads the finding as confirmation that AI development is prioritized on high-wage, high-prestige work because it's profitable and easier to generate verifiable training data for, sparking a broader, often heated debate about whether automation gains will trickle down to consumers/workers or just concentrate wealth further. **Sentiment:** 10% positive · 35% mixed · 55% skeptical **The case for** - Some argue automating the most profitable (often white-collar) tasks first could eventually lower costs for consumers if competition holds. - A few see potential upside in reduced returns to high-paid 'knowledge' professions eventually making things like housing more affordable. - Some reframe the shift positively, noting manual trades can be personally satisfying and are currently safer from automation. **The pushback** - Many doubt cost savings from automation will be passed to consumers, pointing to historical examples of corporations keeping margins instead. - Commenters argue necessities like housing, healthcare, and education have gotten more expensive even as consumer electronics got cheaper, undercutting the 'AI will bring disinflation' argument. - Several see this as confirmation that profit motives drive AI prioritization rather than any technical inevitability, favoring high-wage/easily-verifiable tasks like code and math. - Concerns raised that automating knowledge work will hollow out the middle class and concentrate wealth further rather than create new opportunity. - Doubts about robotics/physical automation being imminent for blue-collar jobs due to cost and mechanical complexity, contradicting any 'robots will free us from drudgery' narrative. **By community** - hackernews (heated): A sprawling, contentious thread where the benchmark-coverage finding becomes a jumping-off point for a much larger argument about automation priorities, wealth concentration, inflation, and whose jobs are actually at risk. **Hottest debate:** Whether automating high-wage white-collar work will ultimately benefit society through lower prices and class leveling, or simply accelerate wealth concentration and hollow out the middle class. **Open questions** - Will cost savings from automating high-wage tasks actually be passed on to consumers, or captured by corporations/shareholders? - How soon, if ever, will robotics/physical AI become cheap enough to automate manual and service jobs currently immune to coverage? - Does the correlation between benchmark coverage and high-wage jobs reflect deliberate prioritization of profit, or simply that verifiable/digital tasks are easier to train for? **Highlights** > > Our findings suggest that 90% of American working hours are spent on tasks with little or no coverage from existing benchmarks. This makes it very hard to know how good AI models are at the work Americans actually do. > ... > But even after controlling for AI-exposure, O*NET tasks from occupations with larger wage bills are significantly overrepresented among benchmark prompts: So, in short: they want AI to take the good jobs. AI is not coming for jobs with tasks like "gather and empty trash" or "service, clean, or supply restrooms," or "mix water and detergents or acids in containers to prepare cleaning solutions according to specifications." We software engineers and lawyers would be smart to start retraining (and adjusting our living standards) now! And this points to one of the ways AI will help improve society and living standards for all: we'll have cleaner, better-maintained bathrooms, as labor one wasted on engaging intellectual tasks is redirected towards menial, low-pay tasks that still need a human touch. Reason to be optimistic, amirite? > — [10111001 on hackernews · 19 comments](https://news.ycombinator.com/item?id=50121809) > Elimination of these jobs is going to be good for everyone else. It's going to reduce the unearned returns to people born with high IQs and aptitude for coding or lawyering or whatever, and reduce upper middle class wages. That, in turn, will drive down prices. The reason housing is unaffordable isn't Jeff Bezos. It's because 20% or so of workers get paid very disproportionately to everyone else and they've been bidding up the price of housing. If Google could fire all their engineers housing would become much more affordable for everyone else in the Bay Area (like it was decades ago before engineering salaries got so high). It'll be more like Iowa or South Dakota is today. > — [rayiner on hackernews · 1 comments](https://news.ycombinator.com/item?id=50122782) > The robots are going to be controlled by the same billionaires that make employees pee in bottles or fire hundreds on a whim over tweets. Our labor power is going to be fully eroded and the robots will not be serving us. > — [guelo on hackernews · 1 comments](https://news.ycombinator.com/item?id=50122421) > AI is coming for jobs involving programming and math (which tend to pay higher) because it's possible to generate infinite training data for those fields via using a compiler/theorem prover to check correctness. They're not coming for high-paying jobs like investment banking and architecture because you can't generate infinite synthetic training data for those. > — [logicchains on hackernews · 3 comments](https://news.ycombinator.com/item?id=50121988) > This is a gleaming silver lining view. Blindingly so. The largest customers of the trades class are the white collar class. The largest customers of service industries are the white collar class. It _will_ be "the rich get poorer and the poor get poorer". This is a hollowing out and redistributing up of wealth, plain and simple. It's the reason AI company leaders are so adamant about framing it as a post-singularity utopia with UBI. They have no answer other than "you're all going to be jobless, better figure that out." Worse yet, they see it merely as a policy problem, rather than a problem they're responsible for. It _is_ a policy problem, no doubt, but the angry mobs won't be coming for policy. > — [dclowd9901 on hackernews · 1 comments](https://news.ycombinator.com/item?id=50122813) **Source threads** - [hackernews](https://news.ycombinator.com/item?id=50121179) · 209 points · 212 comments ## Similar posts on daily.dev - [US Job Market Visualizer](https://daily.dev/posts/us-job-market-visualizer-emne13xou) · Hacker News · 1 upvotes · 0 comments - [Are AI agents ready for the workplace? A new benchmark raises doubts.](https://daily.dev/posts/are-ai-agents-ready-for-the-workplace-a-new-benchmark-raises-doubts--osmrz9upm) · TechCrunch · 2 upvotes · 0 comments - [We’ve been measuring AI wrong; why economically valuable work is the new benchmark](https://daily.dev/posts/we-ve-been-measuring-ai-wrong-why-economically-valuable-work-is-the-new-benchmark-twil0l1aq) · The New Stack · 0 upvotes · 0 comments --- Tags: [#ai](https://daily.dev/tags/ai), [#llm](https://daily.dev/tags/llm) [View this post on daily.dev](https://daily.dev/posts/90-of-us-working-hours-are-spent-on-tasks-not-covered-well-by-ai-benchmarks-yuxxakxfo) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"90% of US working hours are spent on tasks not covered well by AI benchmarks","url":"https://daily.dev/posts/90-of-us-working-hours-are-spent-on-tasks-not-covered-well-by-ai-benchmarks-yuxxakxfo","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/90-of-us-working-hours-are-spent-on-tasks-not-covered-well-by-ai-benchmarks-yuxxakxfo"},"datePublished":"2026-10-16T18:20:10.488Z","dateModified":"2026-10-16T19:40:10.274Z","description":"Mercor built a 'Benchmark Map' that overlays 29 selected AI benchmarks onto the US Department of Labor's O*NET taxonomy of 923 occupations and 18,797 tasks,...","image":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/266827f31b046e69520205906ab00bfb?_a=AQAEuop","thumbnailUrl":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/266827f31b046e69520205906ab00bfb?_a=AQAEuop","isAccessibleForFree":true,"articleSection":"Hacker News","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Hacker News","logo":"https://media.daily.dev/image/upload/t_logo,f_auto/v1/logos/hn","url":"https://daily.dev/sources/hn"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/90-of-us-working-hours-are-spent-on-tasks-not-covered-well-by-ai-benchmarks-yuxxakxfo","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":0},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"ai,llm","timeRequired":"PT9M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Hacker News","item":"https://daily.dev/sources/hn"},{"@type":"ListItem","position":3,"name":"90% of US working hours are spent on tasks not covered well by AI benchmarks"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/90-of-us-working-hours-are-spent-on-tasks-not-covered-well-by-ai-benchmarks-yuxxakxfo#faq","mainEntity":[{"@type":"Question","name":"What percentage of American working hours are covered by existing AI benchmarks?","acceptedAnswer":{"@type":"Answer","text":"About 10% of American working hours are spent on tasks with meaningful benchmark coverage, meaning roughly 90% of working hours are spent on tasks with little or no coverage from existing AI benchmarks. This estimate comes from mapping 29 selected benchmarks onto the US Department of Labor's O*NET taxonomy of 923 occupations and 18,797 tasks, counting a task as covered only if it has at least 5 benchmark prompts. Tracking how AI capability measurement maps to real jobs is easier with daily.dev surfacing labor-index research like this."}},{"@type":"Question","name":"Why do AI benchmarks focus more on software engineering and financial analysis than on jobs like truck driving?","acceptedAnswer":{"@type":"Answer","text":"Benchmarks disproportionately cover tasks that are both AI-exposed and tied to occupations with larger wage bills, since benchmark creators naturally prioritize measuring progress where AI progress is most plausible. O*NET tasks rated as AI-exposed are about 6 times more likely to appear in benchmarks than unexposed tasks, and tasks rated most important to an occupation are 6-10x more likely to be benchmarked than low-importance tasks. Developers evaluating which jobs AI tools can realistically augment can follow this labor-impact research via daily.dev."}},{"@type":"Question","name":"Does weighting AI model benchmark scores by O*NET task instead of raw prompt count change the results much?","acceptedAnswer":{"@type":"Answer","text":"For Mercor's APEX benchmarks across investment banking, law, accounting, and software engineering, O*NET-weighted scores were generally very similar to simple unweighted prompt averages, since model performance was fairly consistent across the O*NET tasks included. For financial and investment analysts specifically (11 tasks, 176 prompts), Claude Opus 5.5 Max scored 73.3% unweighted versus 70.2% weighted, and GPT-6 Astra Max scored 65.5% versus 61.8% weighted. Comparing how different models actually perform on occupation-specific work is simpler when this kind of benchmark analysis surfaces on daily.dev."}}]} ```