How has the labor market
changed since AI arrived?

After ChatGPT arrived in November 2022 as the first conversational generative AI in wide public use, two debates emerged.

• Has AI reduced white-collar jobs? Or, contrary to popular belief, is there no clear sign of that?
• Does AI favor junior workers, because it makes it easier to learn new knowledge and build skills? Or does it favor senior workers, because it lets people who already have expert knowledge handle more work at once?

There is also substantial pushback, because hiring in these occupations may have started falling before ChatGPT (AI) arrived.

There are counter-rebuttals too. The main argument is that even if the drop in total hiring is explained by rate hikes and the business cycle, the shift in the share of young and junior hires began after ChatGPT or continued after rates peaked.

Studies are also coming out in Korea, and the debate there looks much like the one abroad.

These debates persist mainly because an unusually large number of confounding factors appeared within a year on either side of ChatGPT's arrival.

• Pandemic-era surge in IT hiring and the correction that followed: the COVID-19 digital transformation boom of 2020-2021 drove a surge in developer hiring, which then declined from the second half of 2022.
• Interest rate hikes: the Bank of Korea began raising its base rate in August 2021, and the U.S. Federal Reserve in March 2022.
• Investment slump and restructuring: as an extension of those rate hikes, a slowdown in venture investment followed in 2022-2023.

Underscore collected 15,073,974 job postings from February 2017 to April 2025 from one of Korea's largest job platforms. We excluded 2017 through the first half of 2019, a period when many old postings have since been deleted, and analyzed the 10,413,271 postings with a posting date (application start date) between July 2019 and January 2025.

From each posting we extracted variables including the title, company name, experience requirement, education requirement, employment type, salary, work location, qualifications, preferred qualifications, application start and end dates, and job and region tags. The variables below were central to this analysis.

• Job classification: beyond occupational groups such as "IT Development & Data, Backend/Server Development," the job tags go down to detailed keywords such as "languages and frameworks (Java, Python, embedded, firmware, etc.)." This let us look in more detail at software development, the occupation at the center of the debate over AI and job losses.
• Experience requirement: requirements such as "Entry-level," "Any (incl. entry-level)," "Entry-level or 3+ yrs" and "3-10 yrs experience." From these we calculated entry eligibility (whether entry-level applicants can apply) and the average required years of experience. (e.g., 3-10 years → 6.5; 4 years or less → 2)
• Firm identification: using the firm identifier attached to each posting, we obtained the industry, business description, founding date and firm type for 337,158 companies, 77.6% of the 434,688 unique firms behind the 10,413,271 postings (87% of postings). We also matched business registration numbers (BRN) using the Korea Fair Trade Commission (KFTC) registry of mail-order businesses and the Financial Supervisory Service (FSS) DART electronic disclosure system, among other sources. This let us confirm the business details of 96.7% of the firms with firm information (75.0% of all firms) in administrative records, beyond the identifiers the job platform provides.

To measure each job's exposure to AI, we followed the method of Eloundou et al. (2023). That study scored whether about 19,000 detailed tasks in the U.S. Occupational Information Network (O*NET) could be done in at least half the time at the same quality using a large language model (LLM). We linked 1,920 job keywords in the postings data to O*NET occupations and gave each a job exposure score β.

• The roughly 1,900 β values cover 99.8% of all postings.
• We linked keywords to O*NET occupations in two ways.
  • 800 were linked to O*NET occupations through the KNOW occupations of the Korea Employment Information Service (KEIS) (681 of them directly via KNOW occupation codes). The remaining 1,120 were matched directly to O*NET occupations with help from generative AI, then reviewed and confirmed by people.
  • Keywords with no matching O*NET occupation (119 via the KNOW route, 35 via direct matching) were assigned the average β of the 2-3 keywords closest in meaning.
• The scores were high for backend development (0.87), app development (0.87) and bookkeeping (0.78), and low for basic production work (0.08) and nursing assistants (0.11), in line with intuition.
• Unless noted otherwise, the tables and charts in this report measure a posting's exposure using the first keyword with a β value in its list of job keywords, because earlier tags better represent the job in a posting. In a random sample of 1 million postings, the share of keywords that appear in the posting title was 28% for the first tag, 16% for the second and 12% for the third, falling with each position. As described later, we also checked that the results hold when exposure is aggregated in other ways, such as the mean, median or maximum across all job keywords.
• We defined occupations by job keywords. For each of 16 occupations, including software developers (16 keywords such as Backend/Server Development, Frontend and Java), IT infrastructure (H/W, SE), physical design (interior, fashion, exhibition), digital design and bookkeeping, we set a keyword list (Appendix) and split each posting across occupations in proportion to each occupation's share of the posting's job keywords.

Separately, we quantified firms' exposure to AI, that is, "Can generative AI directly produce the core output of this firm's industry?" Two models, OpenAI's GPT-5.6 Astra and Anthropic's Claude Opus 5, scored each industry on the same criteria, giving us firm exposure A.

• The two models' scores had a correlation of 0.983, which is very high.
• As with job exposure, the scores matched intuition: application software development (0.93) and advertising agencies (0.89) scored high, while general clinics (0.07) and restaurants (0.05) scored low.

To check the posting data against actual employment outcomes, we also collected and analyzed the Employment Insurance Statistics that the Korea Employment Information Service (KEIS) publishes monthly (new enrollments by broad industry and insured workers by age). Because these figures are broken down only by broad industry, they show the information and communications industry as a whole but not the detail available in the posting data, so we used them only as supporting evidence.

1Since 2022, the share of entry-level postings has fallen in IT and across office and professional jobs more broadly.

• For software developers, digital design and bookkeeping, both entry-level and experienced-worker postings fell, while for IT infrastructure, physical design, marketing and sales, experienced-worker postings rose and only entry-level postings fell. From 2022 to 2024, the entry-level share changed by -10.3 pp for software developers, -13.8 pp for IT infrastructure, -12.7 pp for physical design, -9.7 pp for sales and -2.6 pp for blue-collar jobs (the comparison group).
• Up to this point, the entry-level hiring rate for blue-collar jobs is almost unchanged, while other office occupations appear to turn in 2022, the year ChatGPT appeared. That makes it look like a causal effect of AI. But...

2Entry-level developer postings began a sustained decline in summer 2022.

• Compared with 2022, average monthly entry-level postings in 2024 changed by -43% for software developers, -34% for digital design, -24% for IT infrastructure and -4% for blue-collar jobs.
• Because a clear decline shows up relative to blue-collar jobs, the comparison group, it is tempting to read this as an AI effect. But the turning point overlaps with the rate-hike period, so it is hard to say clearly how much of the change comes from the business cycle and how much from AI. On top of that, the turning point, summer 2022, came a full six months before ChatGPT appeared.
• That question aside, our data does not support the claim that "developer hiring fell because of the startup investment slump and tech firms' funding troubles, not AI". Relative to blue-collar jobs, entry-level developer postings at tech firms (e.g., Naver, Kakao) fell by about 48% (p<.001), while those at non-tech firms (e.g., GS E&C, MBC, SM Entertainment) also fell by 24% (p<.1).

3Relative to blue-collar jobs, only two occupations saw a significant drop in entry-level postings: software developers and digital design.

The chart below compares the change in postings for each occupation after ChatGPT's release with blue-collar jobs, the comparison group, using difference-in-differences estimated by Poisson pseudo-maximum likelihood (PPML). Each block is 5%. Solid blocks are statistically significant changes; dotted blocks are not significant.

• Only two occupations showed a significant difference: software developers (-33.5%, p<.01) and digital design (-20.6%, p<.05). Both also passed the placebo test, a statistical test that uses July 2021, before ChatGPT appeared, as a "fake post date" (developers -9.8%, p-value 0.352; digital design +3.7%, p-value 0.783).
• Education (-16.6%, p<.1) is borderline, but the placebo test already shows a decline of -21.7% (p<.001), so this looks like a trend that began before ChatGPT.
• For experienced-worker postings, no occupation differed significantly from blue-collar jobs.

4The drop in entry-level developer postings appeared regardless of firm size or age, and was clearest at small and medium-sized firms.

• The chart above breaks down the decline in entry-level developer postings by firm type and compares each with the comparison group (blue-collar jobs).
• A decline appeared at every type of firm, but the only statistically significant ones were small and medium-sized firms (-35%) by size and firms founded less than 20 years ago (-30%) by age.
• Contrary to the common belief that "startups that struggled to raise money stopped hiring juniors," startups' entry-level developer hiring changed by -11%, not significantly different from blue-collar jobs.

5Employment insurance data on actual jobs also show a decline in youth employment.

• Employment of people aged 29 and under in the information and communications industry peaked in the second half of 2022 and has declined steadily for 3 years since.
• In the same industry, by contrast, employment of people in their 30s and those aged 50 and over rose.
• Employment of people aged 29 and under in manufacturing, on the other hand, showed no change.

There were two main reasons we could not clearly separate the effects of these confounding factors using only the number of job postings firms published.

• The conclusion depends on the comparison group. Compared with blue-collar jobs, entry-level developer postings fell by a significant 33.5% more, but compared with IT infrastructure (H/W, SE), which is also part of IT, the difference (-18.9%) is not significant.
• The confounding factors appeared close together in time and all pushed down the incentive to hire entry-level workers. Rate hikes, the correction after the 2021 hiring boom and the venture investment slump all, like AI, reduce the preference for entry-level relative to experienced workers.

With these problems in mind, we set two directions for the analysis.

• First, while controlling for the various confounding factors, we set out to measure accurately the causal effect of generative AI, represented by ChatGPT, on the labor market. Instead of simply comparing differences across occupations or industries over time, we held the firm, month and occupational group fixed (fixed effects) and examined the data statistically.
• Second, instead of the number of postings, we used the years of experience each posting requires as the main variable. If the hypothesis that AI replaces junior workers holds, required experience may be a more important indicator than the simple count of postings.

Below are six major changes in Korea's labor market that we found on this basis.

1Required experience in major office jobs changed little in the 3 years before AI, but began to rise after ChatGPT arrived.

• Software developer postings required about 3.7 years of experience on average from 2019 to 2022. That figure barely moved, even as entry-level postings fell by more than half during COVID and then more than doubled during the hiring surge.
• After ChatGPT arrived, it rose to 4.07 years in December 2022 and an average of 4.82 years in 2024, reaching 5.18 years in January 2025. The pattern is even clearer compared with blue-collar jobs, the comparison group (+0.66 years; p<.001).
• Other office and professional jobs also rose from a low in 2022, but IT infrastructure and physical design were already on a different trend from blue-collar jobs before ChatGPT, and the difference for sales was not significant.

• But looking at occupations one at a time, or defining "major occupations" arbitrarily, can lead to conclusions distorted by the researchers' cognitive biases. So instead of occupation averages, we analyzed the exposure to AI of each job appearing in each posting, which gave us a more neutral and clearer conclusion.

2The more of a job's work AI can do, the more its required experience rose.

• As mentioned, we assigned job exposure values to about 2,000 job keywords. The analysis showed a very clear correlation between exposure to AI and the rise in required experience in the 2 years after ChatGPT appeared.
• One could still question this result. Highly AI-exposed jobs are concentrated at firms that are themselves highly exposed to AI, such as IT companies, so one could argue there is no way to tell whether the "job" was affected by AI or the "industry" that employs many people in that job. Our analysis below rejected this scenario as well.

3Even comparing postings from the same firm in the same month, the rise in required experience grew with exposure to AI.

• Comparing only postings from the same firm, in the same month and within the same occupational group (fixed effects) statistically controls for the firm's finances and restructuring, the effect of interest rates and the economy on that firm, and trends across the whole industry.
• The result held: jobs with 1 standard deviation higher exposure to AI saw required experience rise by an additional 0.073 years after ChatGPT (p<.01). By 2024, the gap in required experience between jobs in the top (5th) and bottom (1st) exposure quintiles had widened to about 5 months.
• As a robustness check, a placebo test that uses July 2021, before ChatGPT, as a "fake post date" found no such difference (+0.009 years), and excluding IT occupations and IT-industry firms left the gap in place at +0.055 years.
• The timing is also clear. As the chart above shows, the gap in required experience by AI exposure became significant at the 10% level from spring 2023, a few months after ChatGPT's release, and at the 5% level from September–November 2023, widening to +0.08 to +0.11 years in 2024. Since firms take time to adopt AI, this lag makes sense.
• For a job one step (1 SD) higher in AI exposure, the share of postings open to entry-level applicants fell by 0.77 percentage points (p<.01), and among postings that require experience, required years rose by 0.088 (p<.01). The hiring market shifted toward experienced over entry-level workers, and, among experienced workers, toward more years of experience.

4Interest rates, the hiring boom, skill levels and industry differences did not explain this gap.

5These changes were clearest in IT development and data, and also appeared in office and professional jobs outside IT, such as sales, retail and trade.

• Ranking the 167 job subcategories on the same criteria (required experience up 0.5 years or more, entry-level share down 5 pp or more, entry-level postings falling at least 20 pp more than experienced-worker postings, no change before ChatGPT), the 7 subcategories that passed every test (SE systems engineering, S/W, information security, backend/server development and others) were all in IT.
• The effect was simply largest in IT jobs: the drop in junior hiring and the rise in demand for experienced workers appeared across office and professional jobs with high exposure to AI.

6In timing, the rise in required experience in postings came before the decline in youth employment.

• The chart above plots "insured workers aged 29 and under in the information and communications industry," from the employment insurance statistics by industry introduced in the previous section, alongside "required experience in developer postings" from our job posting data.
• Required experience in developer postings rose first, from 3.76 years in the second half of 2022 to 4.10 years in the first half of 2023, while employment of people aged 29 and under in the information and communications industry peaked in the second half of 2022 and fell from 2023. With so few data points, we recommend treating this as a reference only.