AI hasn’t – More than four years after AI entered mainstream life, “automatable” roles like radiology and software engineering are still very much staffed. Benjamin Todd, president of the nonprofit 80,000 Hours, argues that AI so far automates parts of jobs—not their mess
On paper, the target list looked simple: radiologists, software engineers, pilots—jobs built around tasks that could, in theory, be broken down and automated. After nearly four years of AI in everyday society, those roles are still very much alive.
Benjamin Todd. president of the nonprofit 80. 000 Hours. said the reason is brutally practical: AI often manages only the defined. “clean” segment of a job. not the messy parts that require judgment. coordination. and human presence. “We should caveat all of this with ‘survived so far,’” Todd told Business Insider.
That caveat matters because several high-profile predictions promised faster disruption. In March, Anthropic CEO Dario Amodei predicted AI could be writing all of the code in software within a year. In 2016, pioneering computer scientist Geoffrey Hinton predicted AI would render radiologists obsolete within five years. But Todd’s view is that the automation story has played out differently in real workplaces—one task at a time.
Todd described a pattern he says fits many white-collar jobs: the tidy, easily automated slice is the minority. “I actually think this is what a lot of white-collar jobs are like,” Todd said. “The defined, clean bit is actually a minority of the work, and most of it is this coordination type stuff.”
He pointed to radiology as a clear example. A 2013 observational study published in the “Journal of the American College of Radiology” found radiologists spent just 36.4% of their time interpreting images. The remaining time went to tasks including consulting with physicians. supervising studies. caring for patients. and administrative work—work Todd said machines can’t yet replicate fully.
The same logic shows up in software engineering, one of the professions AI has disrupted most aggressively. Todd said AI tools have taken over coding, and hiring for junior software engineers has weakened. But he added that overall employment in the field has still increased since ChatGPT launched.
Tech hiring analytics firm TrueUp reported that software engineering job openings at tech companies climbed to more than 67. 000 in 2026—its highest level in more than three years—with listings roughly doubling since mid-2023. Todd argued the shift is less about replacement and more about a productivity change: “I think for software engineers. it’s more the productivity effect. ” he said. “Most software engineers I know now say they’re just producing a lot more code than they did before.”.
Translators represent another variation on the same theme. Even after Google Translate rose. Todd said translator employment remained slightly up overall from 2023 to 2024. citing an analysis by Virginia University professor Basil Halperin based on Bureau of Labor Statistics Occupational Employment and Wage Statistics data.
Under Todd’s framework, the real danger arrives when AI can handle enough of a role that a worker becomes unnecessary—when the “part” AI automates grows large enough to replace the entire function.
He also warned that wages are more likely to come under pressure once AI systems can automate a large enough share of a job’s tasks. That’s part of why he urged workers to shift their mindset away from chasing “safe jobs” and toward building “safe skills.”
In an updated version of his book “80. 000 Hours: How to Have a Fulfilling Career that Does Good. ” Todd lays out four characteristics of skills most likely to increase in value in the age of AI. He said the most valuable skills are those that are hard for AI to do. complement AI systems. produce things society wants much more of. and are difficult for other people to master.
Todd told Business Insider that roles most likely to survive tend to involve work that is “data-poor, long-horizon, messy, and where we want a person in the loop.” Examples he gave include strategy, long-term research, social coordination, and relationship-heavy work.
Still, he didn’t suggest this is a permanent shield. Todd warned that the balance could shift quickly if AI systems become capable of handling larger, more autonomous projects. “The process takes longer than people expect,” Todd said. “But it really is a matter of degree.”
In other words, the jobs that remain today may not be immune—they may just be earlier in the timeline of what AI can do. The question for workers now is how long “survived so far” remains accurate.
AI jobs automation 80 000 Hours Benjamin Todd radiologists software engineering ChatGPT Anthropic Dario Amodei Geoffrey Hinton TrueUp Google Translate translator employment labor market safe skills
So basically AI is just doing the easy parts? I mean yeah, my job is still here.
I don’t buy it. Radiology is probably just gonna be “automated” later anyway, like they always say. They said software engineers would be gone years ago and now everyone’s acting shocked.
Wait reply to Mark—so the article is saying AI can’t replace pilots or whatever because of “messy parts”? But isn’t navigation already basically computer stuff? Feels like they’re moving the goalposts. Like, it will happen, just not on their schedule.
“Survived so far” is such a weird phrase, like oh cool it’s surviving, got it. But radiologists only spend 36% reading images?? that sounds low, unless the rest is all paperwork and politics. Also coordinating is still work, people keep forgetting humans get tired and need decisions. So yeah, AI isn’t replacing everyone, it’s just making everyone do the messy parts faster.