Asking for a friend … which jobs are safe from AI?

Support for NPR and the following message come from SAP Concur, a leading brand for integrated travel, expense, and invoice management solutions.

SAP Concur delivers AI breakthroughs that help your business move forward faster.

Learn more at concur.com.

This is Planet Money from NPR.

Last summer, Charlie Baker was very bored.

He was a Rising College senior, had an internship at the New Jersey Department of Community Affairs.

Entering data into spreadsheets, that was what he did on an exciting day.

And one day, he's in the break room.

I'm picturing beige, everything beige.

Yeah.

Or gray.

It's like, and it also, it has this sort of smell of like a thrift store,

if that makes sense.

Yep.

I do know that smell.

It's not, you know, I don't love it for you, but I know it.

And in that pungent break room, he sees on the table something that in other circumstances would not be exciting.

Someone had left out like a LSAT studying book, and I was like, oh, maybe I should check this out.

He starts working through this book, doing practice questions for the law school admissions exam for fun.

And it is like the perfect law school meet-cute.

LSAT book and Charlie run into each other in the break room, and the rest is history.

It's all weird little like puzzles, the most convoluted riddles like anyone has ever written.

And I was like, oh, I should really do this.

Do this, meaning take the LSAT and go to law school.

But almost immediately, it's like the soundtrack shifts in Charlie's mind.

Maybe he's not in a fun law school rom-com.

Maybe he and everyone he knows is actually living in some kind of dystopian technological horror movie where there's an evil robot on the prowl going after every last job.

I don't know.

I don't know if it's worth the investment now to go to law school for three years, if I'm potentially going to just be replaced by an AI chat bot.

AI.

A lot of people are worried about this.

I have no idea how to plan for the future.

It's so uncertain and scary.

What would AI not automate out?

Anna Wynne, like Charlie, is worried about the robots.

Anna is in her 30s.

She's been doing product design in tech for 10 years.

Do you like your job?

Yeah, I do.

It's great.

But then there's also, you know, all the looming layoffs that are happening across the industry.

She suspects that some of these layoffs are already driven by AI.

And she's worried that she could be next.

There's already software that can make designers like her a lot faster.

She thinks it's possible that AI could eventually cut her out of the loop entirely.

So Anna has started paging through a list of jobs in her mind, trying to imagine which of them might survive AI.

And it's anything right now physical, maybe a plumber.

I mean, it's going to take a while before you could automate that.

I was thinking maybe electrician next to look at.

Or there's her mom's job, nail tech.

She's always like, you know, you can come back here.

There's a lot of work.

Like, I'm ready to start a business with you if you want.

But also, who knows?

Maybe welding.

I saw like a video with someone who trained for it.

I was like, huh.

She's still in the research phase, but she's taking it really seriously.

Getting down to brass tacks.

What the school costs, how much maybe I would be making the first year, second year as an apprentice.

So really kind of, you know, doing the numbers before making a leap.

Charlie is doing the numbers too.

Is law school going to be worth it?

If I'm going to graduate with, say, whatever, $100,000 in debt to a legal field where they're decreasing the jobs, I mean, That's a really bad situation.

So, despite his love affair with the LSAT, he has decided to delay law school for now.

Wait and see.

Anna is also in a sort of holding pattern, both of them just bracing for the AI future.

Charlie and Anna are so, so not alone.

I mean, I am worried about this, and so are lots of my friends.

Worried about which jobs they should steer towards or away from.

Worried about what direction their kids should go or not go.

Dozens of you, our listeners, have written into us about this, saying things like, Maybe my yoga teacher side gig is actually my safest bet now, and my parents were in real estate, and I never thought I'd say it, but maybe that's what I should do.

It feels to me like we all have no idea how to think about this.

Like, even if you can really quickly remember all the jobs that exist, which of them might be your safe harbor?

How do you figure that out?

Hello, and welcome to Planet Money.

I'm Amanda Oronchik, And I'm Sally Helm.

Asking for a friend, which jobs are safe from AI?

Today on the show, we talked to two researchers who have come up with some first drafts of the future.

Some potential blueprints for people like Charlie and Anna and me.

And Sally.

Two frameworks for thinking about how AI will affect jobs.

Which might disappear, which might be more likely to stay, and which will change in ways we haven't even imagined.

I love these conversations knowing more more about the machines, but also about what it actually means to be human.

This message comes from Capital One with the Venture X card.

Earn unlimited double miles on everything you buy.

Plus, get premium benefits at a collection of hotels when booking through Capital One travel.

What's in your wallet?

Terms apply.

Details at capital1.com.

This message comes from Capital One.

The Capital One Venture X Business Card has no preset spending limit, so the card's purchasing power can adapt to meet business needs.

Plus, the card earns unlimited double miles on every purchase.

So the more a business spends, the more miles earned.

And when traveling, the VentureX Business Card grants access to over a thousand airport lounges.

The VentureX Business Card.

What's in your wallet?

Terms and conditions apply.

Find out more at capital1.com/slash venturex business.

I had kind of a secret mission as I set out to report this story, one that I hadn't even fully articulated to myself.

What I really wanted to find, if I'm being honest, is a list of jobs that are just going to be immune to AI.

The super intelligent thinking machines will not be able to do them.

And for a moment, I thought I'd found it.

I have the list pulled up in front of me right now.

It takes almost a thousand jobs and ranks them by something called AI exposure.

And before I understood precisely what that means, I was like, jackpot, this list is going to tell me and Anna and Charlie everything we need to know about our AI future.

On the list are all kinds of jobs, midwives, detectives, pesticide handlers, sprayers, and applicators, comma, vegetation.

Oh, there's lots of cool ones there.

Dredge operator.

That one's pretty cool.

This is Daniel Rock.

He is the man behind the list.

His co-authors on his paper about this were researchers at OpenAI.

They actually used AI as a tool in this study.

And what they did is they took these thousand or so jobs and looked at each job as a bundle of tasks.

Yeah, things you do if you are a midwife or a detective or a pesticide handler.

20,000 tasks that people do in the economy.

The source of these task lists is an amazing government database called O-Net.

If you go look at it, which I recommend you do, I also recommend that you set a timer or you may find yourself, as I did, looking up after like half an hour and realizing that you have just read the entire task list for baristas.

Daniel and I looked at the task list for him, an economist.

According to this, you have 16 tasks.

Is that right?

Yeah, last I counted.

Yeah, that sounds about exactly right.

Okay, explain economic impact of policies to the public.

Supervise research projects and students' study projects.

Have you ever done that?

That sounds, yeah, I was doing that yesterday.

Daniel's paper looked at 19,265 tasks listed in O-Net.

The paper took those tasks and evaluated how exposed each one is to AI.

Daniel's measuring exposure, which means basically how much these large language models can help us do our tasks.

If the AI can help a human complete a task in at least half the time, Daniel labels it E1.

If AI can't really help at all, it's E0.

Then there's E2, which is a sort of in-between score.

E2 is, yeah, you could get some benefits, but you have to build systems around it.

Like, AI can't just do this one out of the box.

It'd need some extra software or something tacked on in order to help.

So we pulled up the task list for an acute care nurse.

They have 26 tasks.

Let's say

administer blood and blood product transfusions.

Right.

So in the horrifying future world nightmare where AI systems do this, it's probably not a large language model or like this vintage of technologies doing that.

So we're going to call that an E0.

So not exposed.

Not exposed.

So you go task by task.

So yeah, so here I have document data related to patient's care.

Yeah, that seems like something a large language model could help.

So yeah, that would be like an E1 task.

And then you give an exposure score to the job as a whole.

And voila, the exposure list.

When I first opened it, there was like a drumroll in my mind because it is a concrete way to look at this big question about the future, about how AI is going to start reaching into the labor market and shaking things up.

I'm looking at this list.

I've put it in order.

Down at the bottom, we've got...

wellhead pumpers.

Yeah.

Our favorite dredge operators.

Love the dredge operator.

Porers and casters, common metal.

Also on the low end, less exposed to AI, there were athletes, dancers, short-order cooks, and as Adawin suspected, a lot of physical blue-collar jobs.

Meanwhile, at the top, a lot of knowledge workers, translators, writers.

We have public relations specialists.

Why are they so high?

Oh, wow.

Yeah.

So I've seen this one in person when a public relations specialist used GPT-4 for the first time, and I saw the light bulb go off.

She, you know, had it write a press release for her in her tone.

And she said it did an absolutely great job.

Now, there was a little bit of fear in there too, because she said, wow, this is like the first few years of my career, like just in a machine.

Yeah.

So this is the thing that feels scary about Daniel's list.

The idea that this machine has read up on everything we've ever done, and now maybe it doesn't need us.

It feels like the jobs at the top of this list are going to disappear, killed by AI.

I mean, like, is this

basically an automation hit list?

No, it's absolutely not an automation hit list.

It's instead a, what is the potential for this work to change list?

That is admittedly not as catchy a way to describe it.

But that is really the big point that Daniel wanted to stress to me.

Exposure to AI is not the same thing as this job will be automated.

So Anna, Charlie, and you Sally.

We are not looking at a list of safe and unsafe jobs.

The main question that Daniel Rock is asking in this paper is not, can I come up with a list of jobs that are safe from AI so that Sally Helm can sleep easier at night?

Daniel has a much bigger question about AI as a whole.

He wants to figure out how far-reaching is the change we're talking about here?

Is AI the kind of technology that will seep into basically every corner of the economy?

Economists call that a general-purpose technology.

So, is it that, or is it something more limited?

Is this like electricity, or is it, you know, like Instagram?

They're very different, right, in terms of the implications.

Instagram is obviously a general-purpose technology, and electricity was okay if it looks like

Instagram changed at all.

Just kidding.

Obviously, electricity is the general purpose technology.

It changes life and work so much that almost no job today doesn't have something to do with electricity.

At least it feels that way to me.

And when you look at these exposure scores, it's really clear.

AI is going to touch a lot of sectors.

Daniel and his co-authors find, yeah, seems like a general purpose technology.

And what that means for us is something that I found simultaneously sort of deflating and kind of hopeful.

Daniel told me that because AI appears to be this big new general purpose technology, the changes to the economy will be so vast that they are very hard to imagine from where we stand now.

Like, you know, before electricity, there was no job electrician or electrical engineer or lighting technician, all of which are today listed in O-Net.

So much is going to change that we really can't say which jobs are going away, which jobs are going to become more important.

We're kind of saying, you know, let's cool it with all of the prognostication about how jobs are going away.

And this is frustrating because it means I don't have something very concrete to bring back to Charlie and Anna.

But this is one of the big takeaways from Daniel's paper.

Like, if you take really seriously that we're talking about something on the order of electricity here, you have to admit that the changes to the labor market might not be what you first imagine.

They might be bigger and weirder.

They might be better or worse.

Basically, you cannot plan around them.

But what you can do is see from this list which jobs are likely to change the most, like at least at first.

And Charlie and Anna are right.

By that measure, lawyers are going to see a lot of changes and welders will see fewer.

But change in this case is a value-neutral word.

Daniel is adamant.

We should not hear this job will change and think this job will go away.

If you're really exposed, it could be great for you.

If you could use AI to make yourself a thousand times more productive,

let's say you're an AI researcher, right?

They're highly exposed.

If you can use these tools to be a really high quality AI researcher,

you might do really well and companies are going to be really excited to hire you at higher wages.

Yeah, if workers get more productive and companies and consumers want more of what they are producing, then everyone wins.

So it could be that some of these highly exposed fields see an explosion of growth, that they're a really good place to be.

Economists would say that demand is elastic.

Of course, if workers get more productive and the world doesn't want even more of what they're producing, demand is inelastic, that leads to job loss.

Like maybe we only need so many news articles or logos.

And so if newswriters and graphic designers get way more productive, there are fewer of those jobs available.

Like if I'm a company and I look at this list and I think, okay,

well, it looks like various people that I employ are pretty high on this list.

Maybe I should think about automating those jobs.

Like, does that make sense?

They might think that way, but they should not think of it that way on the basis of our data.

Daniel thinks that organizations will need to experiment.

He gave me an example of a study where a company gave an AI tool to two groups of paralegals.

One group was told, just use these tools to get more productive.

And the other group was told, use these tools to do the parts of your job that you hate.

The office where they said, use this tool to get rid of the things you don't like doing, the paralegal role changed.

They really flourished.

They started working on some work that even seemed like junior attorney work.

And the other office, there was limited adoption.

It didn't really make as much of a dent.

So, what Daniel's paper does tell us is which jobs AI might change.

What it doesn't tell us is which jobs will live and which will die.

But in my quest to answer that question, I did find another paper that gave me a whole new way of looking at all of this, using that same list of 19,000 plus tasks that workers are doing all across the economy.

That's after the break.

Support for NPR and the following message come from USPS.

With USPS Ground Advantage Service, it's like your shipment has a direct line to you.

It leaves the dock, you know about it.

It's on the road, you know.

And when it reaches your customer, you guessed it, you're still in the know.

Here's the real game changer.

It's one journey, one partner, total peace of mind.

Check out USPS Ground Advantage Service at usps.com slash in the know because if you know, you know.

This message comes from Granger.

If your job at a healthcare facility includes disinfecting against viruses, you know prevention is the best medicine and maintaining healthy spaces starts with a healthy cleaning routine.

Granger's world-class supply chain helps ensure you have the quality products you need when you need them, from disinfectants and cleaning supplies to personal protective equipment to help deliver deliver a clean bill of health.

Call 1-800-GRANGER, clickgranger.com, or just stop by.

Granger for the ones who get it done.

This message comes from Donors Choose.

Before there are leaders, there are learners.

Before progress, there's potential.

Students are showing up ready to grow, but they need fully stocked classrooms to thrive.

Learn how to donate at donorschoose.org/slash back to school.

Isabella Lawisa is a researcher at MIT.

And when she started to hear her friends talk about AI a couple of years ago, she had a Franklin Delano Roosevelt moment.

I don't think I was ever afraid of AI taking over my job.

I was more perhaps afraid of the fear that people were feeling.

You were afraid of the fear.

The only thing you fear is fear itself, Isabella said.

No, yes, because I saw that there was a lot of anxiety in folks around me.

And I was like, ooh, this isn't good.

She wanted to figure out, are those fears justified?

Isabella is a computational social scientist, meaning basically that she incorporates computer science techniques to help answer social science-y questions.

She teamed up with a well-known MIT economist, Roberto Rigabon, and what they ended up doing was kind of turning Daniel's O-Net research inside out.

Instead of looking at what tasks AI can help do, they asked, What are humans good for?

Let's look at what humans can do because we're here.

There's billions of us on the planet right now.

And even if AI came and automated all the jobs that exist, then what are we going to do?

Right?

So that's what really sparked that kind of question of like, hey, let's look at what is complementary that humans can do very well that machines still can't do that well, at least for now.

To answer that question, Isabella and her co-author talked to a lot of people.

Her co-author, Roberto Regabon, has actually been thinking about this for years.

They consulted psychologists and philosophers, and they ended up condensing things down into a single score.

It's called the EPOC score.

It's an acronym, and each letter stands for an area where they think that humans will be especially needed to complement AI.

It's kind of a human-ness score.

E stands for empathy, pretty human trait.

AI can maybe simulate it, but arguably the whole point of empathy is that another human is feeling your pain.

P is presence.

Do you need to physically be there to do the task?

Or does your work benefit from face-to-face collaboration?

Then we have O for opinion, judgment, critical thinking.

But here we also really want to emphasize all the moral and ethical judgments that humans have to do, right?

So it's kind of like ethics, but you didn't want another E.

Yes.

C is for creativity.

Isabella emphasizes that AI is trained on a bunch of existing data.

So even if they can, like, write a poem, they're arguably not as good as humans at imagining entirely new possibilities.

And then H is one of my favorites, actually.

H is for hope.

So tell me about that one.

Yes, that is also my favorite.

Because when it came up, I was like, really?

Hope is hope.

Is hope something that we need for work?

And then when you actually look at the data, there's a lot of occupations that require to have hope in the future.

The full name of this category is hope, vision, and leadership.

So it's things that involve planning and like envisioning a goal and rallying people to get there.

One example might be a substance abuse counselor.

You got to have hope for your client's recovery.

In fact, in some ways, that is the very thing you're hired to have.

Next, Isabella wanted to figure out how much of these various skills are involved in any given occupation.

So she used essentially a computer program to read all of those O-Net tasks, and then she she would assign each job an overall epoch score.

The result is, Sally, for you, a list.

A list.

A list of jobs that essentially score higher or lower on humanness.

Yes, imagine my excitement.

A list.

And near the top of the list, we have, for example, emergency management director.

These are people who prepare for disasters and then come in after disasters to help manage the fallout.

The job requires lots of judgment, lots of empathy, lots of presence.

In fact, managers of all kinds scored high on epoch.

Even something like information technology project managers.

That was surprising to me.

It kind of sounds like a computer-y job.

But if you look at their list of tasks, it's a lot of planning, a lot of leading teams, managing people, and in general, more jobs than you might think have a lot of Epoch going on.

Construction workers, for example, scored higher than Isabella expected on empathy.

I was very surprised and I was like, hmm, what's happening here?

And it turns out that there's one or two tasks in their occupational description, which says they are mentoring others or teaching less experienced construction workers, for example.

Now, there were some weird things that happened in assigning epoch scores, because some tasks are so obvious to us that they actually aren't explicitly written down in O-Net.

Like the task list for barber doesn't say, I have to physically be at the salon holding the scissors in my hands.

So a lot of physical, manual labor jobs actually scored pretty low on Epoch, even though absent some kind of like major robotics boom, those are jobs where you do in fact really have to be there.

But the other thing that sticks out is that clerical jobs tended to score low.

Tax preparers, insurance appraisers.

So it's possible that those jobs could be most at risk from AI.

Now, of course, all of this is just a theory.

Maybe AI will get a lot more human-like, or maybe we just won't care that it's only simulating empathy.

But importantly, unlike Daniel and his co-authors, Isabella and her co-author actually did try to break down the risk for different jobs based on these humanness scores.

The question is basically, is AI likely to swoop in and steal this job?

Or is the job still going to exist, but AI is just going to help humans out?

Is the job likely to be automated or augmented?

Isabella's paper looks at that question in an interesting way.

It takes those task lists again from O-Net, and it zeroes in on the fact that some tasks tend to occur together.

So if you have a bunch of clerical tasks, but also some connected tasks that are highly human, then your job might be safer.

AI might end up augmenting you, not replacing you.

Think of it not as a robot taking your job, but as your own personal bionic arm.

Like for a professor, one of their tasks might be making slides for a lecture.

AI can probably do that.

But a linked task, giving the lecture.

That's pretty human.

Or take lawyers.

The task that is delivering the argument in front of the judge requires a lot of presence, so it's really hard to automate that task.

But then writing the brief about it, you know, that task might be very automatable.

You know, this is actually making me think of a listener who wrote into us.

His name's Charlie.

I told Isabella about Charlie Baker, our listener who has decided to delay law school.

And Isabella agrees with Daniel Rock.

The legal field is likely to be affected by AI.

The more clerical type of jobs can be more easily automated, but there's another great number of occupations in the legal field which are not going to be as impacted.

All the different occupations that require critical thinking, judgment, even creativity, that is not going to go away.

So you're kind of telling Charlie you can go to law school and think about like the more interesting parts of the law.

Like try to get good at judgment, try to get good at

argument.

Don't worry about clerical tasks so much because they might be done by machines.

Yes, exactly.

Like learn how to think.

It is kind of similar to what Daniel Rock told me.

Well, it sounds like Charlie's already off to a very clever start.

It sounds like Charlie is thinking about the discount rate on his expected future cash flows for being a lawyer is being a little bit higher, riskier cash flows there.

Sorry.

No.

Thank you.

Lawyers in particular are the group of people I'm least worried about.

They will find a way to change the rules of the game that help, you know,

as a field, right?

He did say, you know, remember those paralegals who made their jobs more interesting and think about how you could do that as a lawyer.

Like, try to imagine what what might be possible for a young lawyer in the future that isn't possible now.

Daniel also had an interesting thought for Anna Wynne, the tech worker who's thinking about becoming a plumber or a welder.

He pointed out, if everyone decides to be a welder right now, there just might end up being too many welders.

Maybe wages would go down.

So you're not necessarily safe.

And also you're not necessarily in danger no matter where you are along the spectrum.

We just can't know.

Daniel and Isabella both had one very concrete piece of advice, which is to learn to use AI so that you can be ready to kind of roll with what's coming, hopefully shape it to your advantage.

And that's not my favorite ever piece of advice.

I think because it's just hard.

Figuring out how to use these tools well takes work, let alone how to use them ethically.

But I did talk to one person who helped me see how this could go well.

Her name is Kat Reardon.

She is a veterinarian.

She told me that one of her favorite zoo animals is the Kawadis.

They have these long noses and these and these incredible stripey tails.

There was one, his name was Jacob, and his favorite thing in the world was dryer sheets.

So if I like put dryer sheets in my pocket, he would come and like put his nose in my pocket and get all excited about the dryer sheets.

So why was I talking to Kat about this Kawadis and dryer sheets?

Because Kat has made the previously unknown veterinarian to AI career jump.

Here's how it happened.

She was posting online one day about about how she'd started using ChatGPT to help her with her patient notes.

Taking these notes is a huge drain on her and other veterinarians.

And ChatGPT was making this really annoying part of Kat's job way faster.

So she posted about this, and then she heard from an AI startup saying, actually, we're trying to make a tool like that to sell to vets.

Do you want to try it?

She did.

And now she works there, doing things like helping the AI learn veterinary terms that it needs to know.

She also still works as a vet, and she uses the tool to automate parts of her job, like listening to her appointments and doing a first draft of her notes.

Yeah, she told me it's helped in some surprising ways.

Honestly, I'll get bit less often because I have my hands on the animal, both hands, and I can kind of feel if they're kind of starting to get upset and I can feel the little muscles tensing or whatever that I was distracted and not paying attention to previously because I was worried about getting my notes done.

And because you had one hand like on a pen.

Literally.

Yeah.

Yeah.

So I, yeah,

I feel like it's a safety issue as much as anything else.

It's the AI augmentation story that Daniel and Isabella are hoping for, and not one that I would have imagined.

And as I had these conversations, I kept thinking that that is in fact the trait we all need to be applying here.

Imagination.

I went in looking for a list, a concrete guide to help me navigate what's coming.

But But I learned there really is no list.

Not yet, maybe not ever.

We are in for a weirder ride than that.

Today's episode was produced by Eric Metal and edited by Marianne McCune.

It was backtracked by Sierra Juarez and engineered by Robert Rodriguez.

Alex Goldmark is our executive producer.

Special thanks to Arvind Karunakaran.

He wrote that paper about apparel equals using AI.

I'm Sally Helm and I'm Amanda Aronchik.

This is NPR.

Thanks for listening.

This message comes from Warby Parker.

What makes a great pair of glasses?

At Warby Parker, it's all the invisible extras without the extra cost, like free adjustments for life.

Find your pair at warbyparker.com or visit one of their hundreds of stores around the country.

This message comes from EasyCater, a business tool for food, helping organizations order food for meetings and events from favorite restaurants, with search tools for dietary needs, budget per person, and more.

EasyCater also offers employee lunch programs, tools for managing food spend, flexible payment options, and simplified expense reporting.

Plus 24-7 live support, all on one platform.

Learn more at easycater.com.

This message comes from LinkedIn Ads.

Target your buyers by job title, company, role, and skills with LinkedIn Ads.

All the professionals you need to reach in one place.

Get a $250 credit on your next campaign at linkedin.com/slash results.

Terms and conditions apply.