What OpenAI’s own numbers say
OpenAI’s own research puts 19% of US workers in the zone where large language models could affect at least half their tasks. Its internal framework classifies 18% of occupations as “high-risk” for task automation. Neither figure comes from a critic. The GPTs are GPTs paper (Eloundou, Korinek, Rock, et al., 2023) found that roughly 80% of US workers could see 10% or more of their tasks affected by LLMs, and 19% face 50%-plus exposure. With LLM-powered software layered in, 47-56% of all tasks could theoretically be completed faster at the same quality. These are OpenAI’s co-authors publishing on OpenAI’s watch.
The company’s four-archetype labor framework, as summarized by AICerts.ai, splits occupations into buckets: 18% high-risk for task automation, 24% “reorganize,” 12% growth, and 46% low-change. The framework classifies task exposure, not employment outcomes. But task exposure is the necessary precondition for displacement.
What the labor market has actually done
Thirty-three months after ChatGPT’s launch, the Yale Budget Lab found no discernible economy-wide labor market disruption from generative AI. The occupational mix has shifted slightly faster than during past periods of technological adoption, but the difference is not large and predates generative AI. There is no upward trend in AI-exposed tasks among the unemployed, regardless of how long they have been out of work.
A finding from OpenAI’s own data compounds the puzzle: since 2024, unemployment rose more in the low-change occupation bucket than in the high-risk one, per the AICerts.ai summary. If exposure determined displacement, the arrow should point the other way.
The caveat worth noting: Yale’s analysis references other recent studies that provide nascent evidence of possible AI impacts on early-career workers. But the trend predates ChatGPT, and small sample sizes make the signal unreliable. Nobody has clean causal identification here yet.
The capability-deployment gap
The most revealing number in the dataset is not the 19% exposure ceiling or the 18% archetype classification. It is the adoption rate. OpenAI’s framework reports that actual task adoption sits at 23.8% against a 90% capability ceiling, according to the AICerts.ai writeup. ChatGPT usage is three times higher in the high-exposure group than elsewhere, but three times a small number is still a small number.
That gap matters because it restructures the policy question. The argument is no longer “will AI displace workers?” vs “will it augment them?” The argument is: vendors are publishing capability ceilings that assume near-universal deployment, while actual deployment sits at roughly a quarter of that ceiling. The exposure numbers describe a world that does not exist yet, and may not exist for a long time, if the adoption curve flattens.
Why the IPO timeline matters
OpenAI is widely reported to be preparing for a 2026 public offering. Its CFO, Sarah Friar, has publicly expressed concern about funding future compute agreements if revenue growth continues to slow. CNBC reported in April 2026 that OpenAI’s revenue growth estimates fell short of projections. Oracle signed a $300 billion five-year computing deal with the company, and OpenAI expanded an Amazon agreement by $100 billion.
That capital structure creates a specific incentive problem. OpenAI is publishing labor-impact research that describes very large potential effects on the workforce while simultaneously telling investors that adoption is still in early stages and the addressable market is enormous. Both statements can be true at the same time. The tension is that the same numbers serve opposite rhetorical purposes: 19% task exposure is alarming in a policy hearing and bullish in an IPO roadshow.
Disclosure requirements that accompany a public filing are the likeliest forcing function for whether OpenAI will have to reconcile those two framings in a single document.
What policymakers should ask for next
The evidentiary baseline has shifted. A vendor has published its own task-exposure numbers, and an independent lab has published a 33-month empirical null. The gap between them is the actual subject of the policy debate, not the raw numbers on either side.
Three things would narrow that gap:
Internal usage data at the occupation level. OpenAI knows which occupational cohorts use ChatGPT and at what intensity. Publishing usage broken down by the same occupation categories used in the exposure analysis would let researchers correlate capability ceilings with actual deployment patterns.
Occupation-level time series. Yale’s aggregate null is useful but coarse. A time series that tracks employment, hours, and wage changes in the 18% high-risk bucket against the 46% low-change bucket, month by month, would show whether the two unemployment trends are converging or diverging.
Vendor disclosure standards for displacement claims. If a company publishes a capability ceiling, it should disclose the corresponding adoption rate in the same document. The 90%-vs-23.8% gap is the kind of thing securities regulators will eventually ask about anyway.
None of this requires OpenAI to say something it has not already published. It requires the company to publish the adoption data that sits next to the exposure data it has already released. The gap between those two numbers is where the policy fight will be fought. Right now, one side of the gap is public and the other is not.
Frequently Asked Questions
Does the 19% exposure figure apply outside the US?
The GPTs are GPTs paper studied US workers specifically. Extrapolating to other labor markets assumes similar occupational structures and technology access. Countries with a different service-sector weight or lower LLM penetration would see different exposure ceilings entirely.
How big is the observed labor shift compared to the internet transition?
Yale quantifies the occupational-mix change at roughly 1 percentage point above the internet-era baseline. The internet transition took over a decade to produce measurable labor reallocation, and generative AI is 33 months in with a smaller observed shift so far.
Does OpenAI treat the 18% high-risk classification as a layoff forecast?
OpenAI explicitly describes it as “not a layoff forecast.” The framework identifies which tasks could change, not whether employers will choose to automate them. A task being technically automatable and a worker actually being displaced are separated by adoption decisions, cost-benefit math, and regulatory constraints the model does not capture.
What legal and regulatory events could force disclosure before the IPO?
Musk’s lawsuit challenging OpenAI’s corporate restructuring was dismissed in May 2026, removing a legal obstacle to the filing. If OpenAI submits an S-1, SEC rules require disclosure of material risks. The company would then face a framing choice: list the 19% exposure figure as a labor-market risk, or cite it as evidence of addressable market size. Using the same number for both purposes in a regulated filing is the specific disclosure problem the IPO creates.