Post-ChatGPT, Finance and Tech Job Postings Fell 13% While Augmentation Roles Rose 20%

The public launch of ChatGPT in November 2022 did not uniformly threaten jobs. It split the labor market in two.

A working paper coauthored by Harvard Business School professor Suraj Srinivasan tracked nearly all US job postings from 2019 through March 2025. The researchers found that postings for occupations heavy on structured, repetitive tasks — the kind generative AI can automate — fell 13% after ChatGPT. Postings for roles that blend analytical, technical, or creative work with AI tools rose 20%. The largest reductions hit finance and technology.

The asymmetry is the story. It is not that AI eliminates jobs broadly. It eliminates some and reshapes others. The question for builders, workers, and policymakers is which side of the divide a given role falls on.

Srinivasan, along with Wilbur Xinyuan Chen of the Hong Kong University of Science and Technology and Saleh Zakerinia of Ohio State University, used OpenAI’s ChatGPT to classify over 19,000 tasks across more than 900 occupations. They assigned each occupation two scores: one for automation exposure, based on how many of its tasks can be replaced by generative AI, and one for augmentation potential, based on the share of tasks that require human judgment alongside AI tools.

The methodology matters. The researchers did not ask what jobs might be lost in theory. They measured what employers actually changed in their hiring. The data set covers nearly all US vacancies, making it one of the most comprehensive looks at the early labor-market response to generative AI.

Microbiologists, financial analysts, and clinical neuropsychologists landed on the augmentation side. Their work involves tasks that AI can handle — data processing, pattern recognition, literature review — but also tasks that require human judgment, social skills, or hands-on technical work. Investment managers use AI to evaluate market data, but their decisions still depend on experience and context.

On the automation side, the researchers found that the number of skills required in job postings for exposed roles shrank by 7%. Fewer new skills emerged in those occupations. The message is clear: employers are narrowing what they ask for in roles they expect AI to replace, not expanding them.

Meanwhile, postings for augmentation-prone roles began demanding more AI-related skills: prompt writing, tool usage, human-AI collaboration. The skill set for these jobs is expanding, not contracting.

“Rather than solely eliminating jobs, generative AI creates new demand in augmentation-prone roles, suggesting that human-AI collaboration is a key driver of labor market transformation,” Srinivasan said in the article.

The paper, first released in December 2024 and updated in August 2025, is titled “Displacement or Complementarity? The Labor Market Impact of Generative AI.” It focuses on short-term US effects. The researchers note that long-term impacts and effects on other regions “remain uncertain as adoption scales.”

That caveat is important. The data runs through March 2025, roughly two and a half years after ChatGPT launched. That is enough time to see early signals but not enough to see the full arc of adoption. Model capabilities have continued to improve. Costs have fallen. Enterprise deployment has accelerated. The second-order effects — how workers retrain, how companies restructure, how entire job categories evolve — take years to play out.

Srinivasan recommends that companies invest in reskilling programs to move workers from automation-prone roles into augmentation-prone ones. “Retraining is essential for jobs where generative AI is reducing skill diversity,” he said. In automation-prone occupations, workers may face displacement unless they develop non-automatable skills such as judgment and interpersonal communication.

For augmentation-prone roles, he recommends continuous upskilling in generative AI. “Generative AI is broadening skill requirements, increasing the demand for AI literacy, human-AI collaboration, and domain-specific AI applications.”

The recommendation sounds like standard corporate advice. But the data behind it is specific. The 13% drop in automation-prone postings and the 20% rise in augmentation-prone postings are not projections. They are measured outcomes. Companies are already behaving as if the divide is real.

The finance sector is a telling case. Investment analysis, risk assessment, and compliance reporting all involve structured data processing that generative AI handles well. The paper found that finance saw some of the largest reductions in automation-exposed postings. At the same time, demand for financial analysts who can use AI tools alongside their judgment grew. The same pattern appeared in technology, where coding tasks that follow clear patterns are being automated while roles requiring system design, architecture, and human interaction are expanding.

The core insight for AI builders is this: the market rewards tools that augment human judgment, not just those that replace human labor. A model that writes a compliance report is useful. A model that writes the report and surfaces the three judgment calls the analyst needs to make is more valuable. The augmentation score in the paper captures this — occupations with high augmentation potential have a mix of automatable and non-automatable tasks.

The paper also warns that how companies integrate generative AI is decisive for job loss or growth. Treat it as a cost-cutting tool, and the automation side dominates. Treat it as an augmentation tool, and the expansion side grows. The choice is not technical. It is organizational.

The 13% and 20% numbers are early signals. They will change as models improve, as deployment spreads, and as workers adapt. But they are already telling builders something concrete: the jobs that survive and grow are the ones where AI handles the structured part and humans handle the judgment part. That is the shape of the market.