Why the Remote Labor Index jump doesn't mean AI replaces jobs

The Center for AI Safety reported that Claude's Fable 5 increased the Remote Labor Index from 2.5% in October 2025 to 15.8%, meaning AI can now handle 15.8% of freelance jobs end-to-end. However, the author argues this jump does not indicate AI replacing broader knowledge work, as the index measures only static freelance projects that differ significantly from full-time jobs with human messiness and accountability.

Why the Remote Labor Index jump doesn't mean AI replaces jobs ... about real-world automation of knowledge work and enterprise workflows Two weeks ago, Center for AI Safety https://safe.ai/blog/significant-increase-in-digital-labor-automation announced that Claude’s Fable 5 increased 6-fold how many remote “jobs” LLMs can perform, going from 2.5% in October 2025 to, now, 15.8% 1 footnote-1 . It’s a major update to their Remote Labor Index - a measure how much of the freelance jobs market can be handled, end-to-end, by AI. I find the jump to 15.8% noteworthy but unsurprising, and the details of how the index is composed will tell you why. The Remote Labor Index is one of the most useful benchmarks to watch, because it’s one of the closest ones to answer the question “what does LLM progress mean for human jobs and real-world businesses?” To build it, researchers pulled 240 projects, amounting to about ~$140k in economic value, off of sites like Upwork. The specific projects are unknown to the public, so benchmarks can’t be gamed. It then tracks what percentage of these projects that can be done autonomously by an agent, end-to-end; it’s important there’s no partial credit for “some completion”. The approach is prudent and discriminative to not-really-real progress, and yet - there’s a big gap between climbing the RLI and taking over large swaths of business work. Since RLI measures freelance remote work only, there’s already some removal of the human element that exists in regular jobs. Even if you set aside the explicit carve-outs that RLI calls out: requiring physical labor, long-term evaluation, or direct client interaction, there are nuanced differences between freelance and, well, loadbearing work. Jobs and projects posted on sites like Upwork are materially different from full-time jobs, in-person or remote. They are easy to slice up to contract out, straightforward to evaluate, require limited amount of trust and accountability. Nobody is losing sleep at night if they go wrong, they just get another contractor 2 footnote-2 . In contrast, most knowledge jobs or business workflows are not like that. As I’ve mentioned in a few comments on Import AI https://substack.com/@kamilas/note/c-289334214?r=5zwg7&utm source=notes-share-action&utm medium=web or Don’t Worry About the Vase https://substack.com/@kamilas/note/c-292716085?r=5zwg7&utm source=notes-share-action&utm medium=web , my experience as a technology practitioner vs. researcher puts me in a skeptical seat when it comes to the promises that AI will take over most of the economy, or even knowledge work or business activities: yes, we might see human-light and AI-heavy companies; we will see “work” augmented on a massive scale. But 1 humans and economy will evolve quickly in response, moving our labor and knowledge to more effective uses, and 2 even in the existing economy, there’s enough muck layer https://loadbearingtech.substack.com/p/every-emerging-tech-has-to-yield in most jobs - the interhuman messiness - that the agent-consumption of broader economic activity will be resistant to major changes 3 footnote-3 . On one hand, the RLI is the closest benchmark we have to tracking real world impact, so increases here are notable. On the other hand, “Freelance Contractor Index” would be a more precise name, and 15.8% of that is a very different number than 15.8% of labor. Lastly, I’m sure this will change in the future, but for the time being, the projects captured in the Remote Labor Index are static, meaning they were captured at the end of 2025 and LLMs are evaluated against the same set. This gives us a frozen reference point, but of course the remote labor landscape will change precisely as quickly as LLMs evolve, and projects will materially change. So even if 2030’s LLM can do 100% of 2025’s remote jobs, that number will be different for 2030’s remote jobs. Jack Clark https://open.substack.com/users/44606-jack-clark?utm source=mentions wrote about this recently https://open.substack.com/pub/importai/p/import-ai-464-fables-writes-gpu-kernels?r=5zwg7&selection=99f07361-7172-4f7b-bebc-e5afe7384c96&utm campaign=post-share-selection&utm medium=web&aspectRatio=instagram&textColor=%23ffffff&bgImage=true taking the opposing position: But I’m more optimistic about human ambition and hunger to evolve, and more bearish on the “ mucky https://loadbearingtech.substack.com/p/every-emerging-tech-has-to-yield ” part of jobs and labor: ongoing relationships, ambiguous long-term horizons, and the right-hemisphere or instinctive parts of doing business. PS. Muck layer https://loadbearingtech.substack.com/p/every-emerging-tech-has-to-yield is a concept I will return to often, which is why I wrote a post about it; I update it often. 1 footnote-anchor-1 When I started writig this post about a week ago, the number was at 16.1% - there’s a chance it’ll get updated further. 2 footnote-anchor-2 That’s not to say they’re immaterial, since the median human-completion time for tasks in the index is 11 hours. 3 footnote-anchor-3 In BlueDot’s AGI Strategy course, the facilitator pointed out to me that my argument assumes humans stay relevant as the “labor” part of the economy, but capital and technology might detach from the equation, leaving labor behind. There’s some convenience to the sweepiness of this argument but I’d love to think through the minutiae of it. Which parts of the economy? Because as soon as we have one human involved, we need a host of humans to support their various human flaws, both executive and emotional. But I am curious about engaging in this argument, so feel free to reach out.