AI and careers
When answers are cheap, what makes a person valuable?
In the 1990s, a computer studying hospital records learned that asthma lowers a pneumonia patient’s risk of dying. The pattern was real, and acting on it could have killed people. AI has since made answers almost free. The evidence points to the people who can tell when an answer is wrong and whom others trust to act on it.
- The price of an answer fell more than 280-fold in two years, and the need to check it stayed. In an experiment at Boston Consulting Group, 84.5% of consultants working alone solved a tricky problem. With GPT-4, 60% to 71% did.
- People often cannot feel when a tool stops helping. In a 2025 trial, experienced programmers took 19% longer with AI and still believed they had been 20% faster.
- Social skills were paying off before ChatGPT. From 1980 to 2012, US jobs that demand both math and social skill grew the most. Jobs heavy in math and light on people shrank.
- No single skill is safe by itself. GPT-4 out-argued people 64.4% of the time when it knew a few facts about them. When a chatbot invented court cases, the lawyers paid the $5,000 penalty.
- Time with the boss pays, and it can be unfair. At a bank in Southeast Asia, smokers who got a boss who smoked were promoted faster, with no difference in how well they worked. Men there rose faster under male bosses in a similar way, and that helps explain why women fell behind.
In the mid-1990s, a large research project set out to teach computers which pneumonia patients needed a hospital bed and which could safely recover at home. Rich Caruana and his colleagues later told the story of one program in that project. It studied real patient records and learned a rule. Pneumonia patients with a history of asthma were less likely to die.
The program hands you its rule. A patient with asthma and pneumonia is waiting. What do you do?
The record the rule learned from left out the most important fact. These patients went straight to intensive care, and that care is why fewer of them died.
The rule was true of the data. Asthma patients who arrived with pneumonia usually went straight to the intensive care unit. The care they got there worked so well that they died less often than other pneumonia patients. Sent home, they would have faced a much higher risk. The program saw the result of the treatment and gave the asthma the credit.
This program wrote its rules in a form people could read, so the researchers could spot the problem and remove it. Their most accurate models were neural networks, programs whose predictions are hard to trace back to a reason. The team reasoned that if the simple program had learned the asthma rule, the neural networks had learned it too, and they judged those models too risky to try on patients. Catching the error took someone who knew what hospitals do with asthma patients.
What is left to get paid for?
Three decades later, machines answer far more than questions about pneumonia, and people are asking what is left for humans. On 22 October 2026, the biographer Walter Isaacson put that question to the investor Peter Thiel. Thiel talked about watching a sunset, an experience he doubts a computer will ever have. He also talked about the way people relate to each other “in families and companies in our larger society.”
“If there were no humans in the world, I don’t think the computer would experience the sunset in the same way… It’s the way we relate to other people, you know, in families and companies in our larger society… So I think these relational dimensions maybe are very core and will become more important.”
Peter Thiel, 54:45 to 56:00
For a student about to graduate, this is a practical question. If a machine can supply what used to take years to learn, what is left to get paid for? The answer starts with what has really become cheap.
The price of an answer really did collapse
Stanford’s AI Index tracks the cost of getting answers from a system as capable as GPT-3.5, the kind of model behind the first ChatGPT. In November 2022 the price was $20.00 per million tokens. A token is a small piece of text, about three-quarters of a word. By October 2024 the price was $0.07, more than 280 times cheaper.
Three answers come out of this machine. Each one looks fine. You will meet all three again, and none of them holds up.
Cheap answers also hand experience to beginners. In a study of 5,172 customer-support agents, an AI assistant raised the number of problems solved per hour by 15% on average. The least experienced agents gained the most. The most experienced gained a little speed and lost a little quality. If the job were only knowing the answer, that job would be worth close to nothing. In When the answer is cheap I followed where the money goes for firms. This essay follows the people.
Nobody can see where the tool stops working
The job also includes knowing when the answer is wrong. An experiment with 758 consultants at Boston Consulting Group tested this. Researchers randomly gave some consultants GPT-4 and left others without it, so a large gap between the groups could be credited to the tool. On 18 tasks, such as proposing ideas for a new shoe, the consultants with AI finished 12.2% more of the work and worked 25.1% faster. Then came one more task, which asked them to combine spreadsheet figures with clues buried in interview notes. The researchers picked it because GPT-4 would give a wrong but convincing answer.
Three of these four tasks sit inside what GPT-4 could do. Which one sits outside?
Task 4 sat outside. It needed spreadsheet figures combined with clues in interview notes, and GPT-4 gave a wrong answer that sounded convincing.
The researchers call the edge of what AI can do a jagged technological frontier. Two tasks can look equally hard to a person, and AI can be excellent at one and misleading on the other. The edge carries no warning sign, and it moves every time models improve.
On that last task, 84.5% of the consultants working alone got it right. With GPT-4, 70.6% did. With GPT-4 and an overview of how to write prompts, only 60% did. Harvard’s summary of the study blames over-reliance on the AI’s output. Consultants like these could solve the problem by themselves, and once they had the tool, many took its answer.
You might expect to notice when a tool stops helping. METR, a research group that evaluates AI, tested this in early 2025. Sixteen experienced programmers worked on 246 real tasks from their own projects, and each task was randomly assigned to allow AI tools or ban them.
How did allowing AI change the time these programmers took? Drag to guess.
Measured, the programmers were 19% slower with AI. They expected to be 24% faster and still felt 20% faster afterward.
The programmers took 19% longer with AI. They had expected to be 24% faster, and after doing the work they still believed they had been 20% faster. Tools improve quickly, and METR’s 2026 follow-up leaned toward a speed-up, although METR calls that data “only very weak evidence.” The gap between how the tool felt and what it did is the part to remember.
Trust seems to feed that gap. In a survey of 319 knowledge workers, whose jobs mostly involve handling information, people with more confidence in AI reported less critical thinking when using it. People with more confidence in their own ability reported more. A survey cannot show which causes which, but the pattern fits the consultants’ results.
This is why cheap answers do not make knowledge worthless. AI has made answers cheap, which leaves two things to pay a person for, knowing when an answer is wrong and being someone others trust to act on it. The first depends on knowing a subject well. The second depends on other people, and that is the part of Thiel’s answer with data behind it.
Social skills were paying off before ChatGPT
In 2017, the Harvard economist David Deming sorted US jobs by how much math and how much social skill they demand. He then followed each group’s share of employment from 1980 to 2012, a decade before ChatGPT.
Four kinds of job, 1980 to 2012. Which grew the most as a share of US employment?
Jobs heavy in both math and social skill grew the most. Jobs heavy in math and light on people shrank.
Jobs heavy in both math and social skill grew the most, by 7.2 percentage points of US employment. A percentage point is the plain difference between two percentages, so going from 10% to 12% is a rise of 2 points. Jobs heavy in math and light on social skill shrank by 3.3 points. Among the fastest-growing knowledge jobs Deming lists are physicians and lawyers, who need both. Nothing in these data supports trading math for networking.
His explanation is about trade. A team gets more done when each person does what they do best and hands the rest to someone else, and every handover has a cost. Social skills lower that cost. In 2021, Deming and Ben Weidmann tested the idea by randomly assigning people to several teams and predicting each team’s score from its members’ individual skills. Some people’s teams kept beating the prediction, whichever team they joined. These team players matched everyone else on IQ and personality and scored higher on a test of social intelligence. Social skill raised a team’s performance about as much as IQ did.
A popular answer to what people should learn now names seven skills, and golf is one of them. Here they are, each on its own card. A press turns a card over to show what the evidence says, and the tray in the corner collects each card as you reach its section below.
Networking. 20 million LinkedIn users tracked for five years. Moderately weak ties moved people into new jobs most.
Holds upRead the testPersuasion. GPT-4 out-argued people 64.4% of the time when it knew a few facts about them.
AI does this tooRead the testCuriosity. An algorithm predicted cost when patients needed it to predict illness. Removing the bias would move one share from 17.7% to 46.5%.
Holds upRead the testTaste. Image tools raised output 25% and the average artwork got less novel. Someone has to filter.
Holds upRead the testStorytelling. Theranos raised more than $700 million on a story its evidence did not support.
Needs checkingRead the testGrit. People who succeed do not try more often. They keep more from each failure.
Needs a fixRead the testGolf. Shared breaks with the boss led to 0.70 pay grades, with no difference in performance.
Holds up, unfairlyRead the testNetworking
Job news travels through people you barely know
In 2022, researchers tested networking on LinkedIn by changing whom its “People You May Know” feature suggested to more than 20 million users over five years. Those users made 2 billion new connections and took 600,000 new jobs. The connections most useful for finding a new job were moderately weak ones. Usefulness peaked at about 10 mutual connections, and closer ties than that helped less. I went through the older evidence for this in Your career is a startup. Close friends tend to hear the same news you hear. An acquaintance sits in another circle.
Your friends all know the same two openings. One weak tie reaches a circle that knows four others.
Written applications are telling employers less. Anaïs Galdin and Jesse Silbert studied coding jobs on Freelancer.com. Before large language models, the kind of AI behind ChatGPT, employers paid more for applications tailored to the job. Afterward they did not, because anyone could produce a tailored application. The authors simulated a market where written applications tell employers nothing about ability, and workers in the top fifth by ability were hired 19% less often. Those in the bottom fifth were hired 14% more often. A professor who has watched you solve a hard problem can now tell an employer what your cover letter cannot.
Persuasion and storytelling
A machine can now out-argue you
Persuasion is the skill on the list that AI has reached most clearly. In a 2025 experiment, 900 people debated a randomly assigned opponent online, either a person or GPT-4, and reported how far they agreed with the topic before and after. Some opponents also received a few facts about the person they were debating, such as age and political affiliation.
Pick who debates you. The rope shows how often they out-argue a plain human opponent.
GPT-4 with a few facts about you was more persuasive 64.4% of the time. This is the one result here that clearly differs from a plain human opponent.
Given those facts, GPT-4 was more persuasive than a human opponent 64.4% of the time, counting only the pairs where one was more persuasive than the other. Short online debates differ from closing a sale, and this study says little about trust built over years. It does rule out treating persuasion as a permanent human specialty.
A machine can make the argument. It cannot answer for it. In 2023, the New York lawyer Steven Schwartz needed cases to support a client’s lawsuit against an airline. He turned to ChatGPT, which he told the court he “falsely assumed was like a super search engine.” It gave him court decisions that did not exist. When he grew suspicious, he asked it to check its own work.
Judge P. Kevin Castel imposed a $5,000 penalty on the lawyers and their firm. He wrote that there is “nothing inherently improper about using a reliable artificial intelligence tool for assistance,” and that “existing rules impose a gatekeeping role on attorneys to ensure the accuracy of their filings.” That gatekeeping is what a client pays a lawyer for once the text itself costs almost nothing.
Storytelling carries the same risk. A good story tells people what the problem is and what has changed, and it can travel faster than the evidence. Theranos told a story about blood testing that, in the words of the US Securities and Exchange Commission, raised “more than $700 million from investors through an elaborate, years-long fraud.”
Curiosity and taste
The right answer to the wrong question
Curiosity can change what a whole project is measuring. Ziad Obermeyer and his colleagues examined a commercial algorithm that US health systems used to pick patients for extra care. By some measures of accuracy it did its job well. What it predicted was next year’s health-care costs, and people treated cost as a stand-in for how sick a patient was. Because of unequal access to care, less money is spent on Black patients with the same needs as white patients, so the algorithm scored Black patients as healthier than they were.
With the bias removed, 46.5% of those picked would be Black patients, up from 17.7%.
Correcting this would raise the share of Black patients among those automatically chosen for extra help from 17.7% to 46.5%. A sharper prediction of cost would not have fixed it. Someone had to ask what the number measured. That is curiosity as a working skill.
Taste comes next, once a machine has produced twenty options. Eric Zhou and Dokyun Lee studied more than 50,000 artists and 4 million artworks as text-to-image AI arrived. The tools raised output by 25% and favorites per view by 50%, and the average artwork became less novel. The authors read this as evidence that the skills that count are coming up with ideas and filtering the results. Filtering well depends on having seen enough good and bad work to tell them apart.
Grit
People who succeed do not try more often
Grit sounds simple, since it seems to mean trying again. Yian Yin and his colleagues measured failure directly, using 776,721 grant applications to the US National Institutes of Health and 58,111 startups. They also studied 170,350 terrorist attacks, a grim third case of repeated attempts. People who eventually succeeded did not try more often than people who stopped. The two groups’ runs of failure were about the same length.
What differed was how much each attempt kept from the ones before. Their model treats an attempt as many parts, like the budget and the data sections of a grant application. For each part, you either write a new version or reuse the best version from your last few attempts. A single setting, called k, decides how many past attempts you look back on.
You look back two attempts. That is enough to hold on to good versions, and quality keeps climbing.
With k at 0, every attempt starts from scratch and quality never improves. At k of 1 it climbs for a while and stalls. At k of 2 it keeps climbing. Two people who work equally hard can land on different sides of that line. In the data, the people who eventually succeeded got faster from one attempt to the next, as the model predicts, and the people who stopped did not. The two groups were already distinguishable by the second failure.
Golf
The joke on the list has real evidence behind it
Golfing with clients sounds like a joke, and the best test of it involves cigarettes. Zoë Cullen and Ricardo Perez-Truglia studied a large commercial bank in Southeast Asia where managers rotate between teams. The rotation works like a natural experiment, because employees do not choose when a new boss arrives. Some smokers got a new boss who smoked, and others got a new boss who did not.
Hold the button to charge, and let go to putt. Each ball in the hole adds a line to the scorecard.
| Hole | Smokers who got a smoking boss, compared with those who did not | Result |
|---|---|---|
| 1 | Share of their breaks spent with the boss | +24 points, a 63% increase |
| 2 | Pay grades gained after ten quarters | +0.70 |
| 3 | Difference in effort or performance | None found |
Smokers who got a smoking boss spent an extra 24 percentage points of their breaks with that boss, a 63% increase. Ten quarters later they were 0.70 pay grades ahead of smokers who got a non-smoking boss. Non-smokers gained nothing from the boss’s habit. The researchers found no difference in effort or performance, so the extra promotions do not appear to come from extra output. They suggest the breaks let employees win favor with the boss and pick up useful information, such as which tasks to ask for.
The same bank shows the cost. Men who moved to a male manager were promoted faster, and women did no better under female managers than under male ones. The authors estimate that this male-to-male advantage could explain about a third of the gender gap in promotions at the bank. Informal time rewards whoever already gets invited, and in this bank it tracked no difference in the work.
An undergraduate, a burnt scroll and the whole list
In 79 AD, the eruption of Vesuvius buried a library under 20 meters of volcanic mud and turned its papyrus scrolls to charcoal. They crumble if anyone tries to open them. Brent Seales at the University of Kentucky spent nearly 20 years on the problem, and in 2019 he scanned some of the scrolls in a particle accelerator. In 2023, the tech entrepreneurs Nat Friedman and Daniel Gross turned that work into an open contest with a $700,000 grand prize.
That August, the contestant Casey Handmer stared at the scans for hours until he noticed a cracked-mud texture that turned out to be ink. Luke Farritor, a 21-year-old computer science student and SpaceX intern, saw people discussing it on the contest’s Discord chat server. He spent his evenings and late nights training a machine-learning model on that texture. Before long the model was finding traces “invisible to his own eye.” One result was a word. When papyrologists, scholars of ancient papyrus, saw it, “they gasped.” It read porphyras, the Greek for “purple.”
porphyras, “purple”
Networking. Farritor heard about the contest on a podcast and saw Handmer’s finding on its Discord server. He later teamed up with two other contestants.
Persuasion. Each papyrologist, working independently, recovered more text from the winning entry than from any other. The evidence did the persuading.
Curiosity. Casey Handmer stared at the scans for hours and found ink nobody had yet seen inside the scrolls.
Taste. Farritor picked out a few dozen ink strokes to label and train on. Choosing what a model learns from is curation.
Storytelling. The contest describes itself as “Resurrecting an ancient library from the ashes of a volcano.”
Grit. Seales spent nearly 20 years on the problem. Farritor gave it his evenings and late nights during an internship.
Golf, or equivalent. Seales did not answer Friedman’s emails, so Friedman had his assistant phone the department. He then invited Seales to California to “hang out,” and “to my shock he did.”
Farritor then joined Youssef Nader, an Egyptian PhD student in Berlin who had found the same word on his own, and Julian Schilliger, a Swiss robotics student whose mapping tools made the scans readable. On 5 February 2024, the three won the grand prize for 15 columns of text, more than 2,000 characters. Papyrologists reviewed the entries without knowing whose they were, and a technical team audited and reproduced the code. The model could see what no person could. The prize was paid once people had checked it.
Farritor could do this because he had already learned to train models, and that learning does not come free. In an earlier essay I described high-school students in Turkey who practised maths with GPT-4. Their practice scores rose 48%, and then they scored 17% lower on the exam than classmates who had used only a textbook.
Each study here covers one setting. The bank is one bank in Southeast Asia. The debates were short and online. The consultants used GPT-4, which newer models have passed. These results describe where value sits today, and they cannot say which skills will pay in 2040.
About 95% of that scroll is still unread. In the columns already read, the author, a follower of the philosopher Epicurus, asks whether things available in small quantities give more pleasure than things available in abundance. The answer, on a scroll no one could read for 2,000 years, is “we do not right away believe things that are scarce to be absolutely more pleasant than those which are abundant.”
Further reading
- Walter Isaacson, The Builders, interview with Peter Thiel (22 October 2026), with a searchable transcript. The question comes at 53:17.
- Dell’Acqua et al., Navigating the jagged technological frontier (Organization Science, 2026), and Ethan Mollick’s plain-language account by one of its authors.
- Caruana et al., Intelligible models for healthcare (KDD, 2015), the source of the asthma rule.
- David Deming, The growing importance of social skills in the labor market (Quarterly Journal of Economics, 2017), and Weidmann and Deming, Team players (Econometrica, 2021).
- Obermeyer et al., Dissecting racial bias in an algorithm used to manage the health of populations (Science, 2019).
- Yin et al., Quantifying the dynamics of failure across science, startups and security (Nature, 2019), the model behind the grit figure.
- Cullen and Perez-Truglia, The old boys’ club (American Economic Review, 2023), the smoking-break study.
- Judge Castel’s opinion in Mata v. Avianca (2023), which reproduces the ChatGPT screenshots.
- Vesuvius Challenge, 2023 Grand Prize announcement, with images of the recovered columns.