Transformative AI is an impact-defined category: AI advanced enough to precipitate a societal and economic transition "comparable to (or more significant than) the agricultural or industrial revolution," the formulation used by Winter & Bullock and broadly shared across the governance literature. The term was popularized by Open Philanthropy (Holden Karnofsky), which introduced it in 2016 as an alternative to "AGI" (Some Background on Our Views Regarding Advanced Artificial Intelligence (Karnofsky, 2016)). It defines the technology by its effects on the world rather than by its level of cognitive capability.
Origin
Holden Karnofsky set out the definition in "Some Background on Our Views Regarding Advanced Artificial Intelligence," published by Open Philanthropy on May 6, 2016, as the organization prepared to make potential risks from advanced AI a major priority. Starting from the observation that the agricultural (neolithic) and industrial revolutions "brought about fundamental, permanent changes in the nature of civilization," he gave two definitions (Some Background on Our Views Regarding Advanced Artificial Intelligence (Karnofsky, 2016)):
- Definition 1 (conceptual): "transformative AI is AI that precipitates a transition comparable to (or more significant than) the agricultural or industrial revolution."
- Definition 2 (operational), offered because the first "leaves a great deal of room for judgment": anything fitting one or more of three descriptions, namely AI systems able to perform tasks that in 2016 accounted for the majority of full-time jobs worldwide, and/or over 50% of total world wages, unaided and at costs comparable to employing humans; AI systems "capable of fulfilling all the necessary functions of human scientists, unaided by humans," in developing a technology that becomes the most significant driver of such a transition; or surveillance, autonomous weapons, or other AI-centric technology that itself becomes that driver.
Karnofsky intended the term to be "for the most part, a less restrictive term" than superintelligence, artificial general intelligence or high-level machine intelligence, since transformative AI could come from "a relatively small number of domains in which AI systems achieve better-than-human performance," and he tied his timeline view to there being "a large number of possible paths to transformative AI" (Some Background on Our Views Regarding Advanced Artificial Intelligence (Karnofsky, 2016)). Karnofsky wrote that anything meeting the second definition would "quite likely" meet the first, and that the definitions leave open AI systems that count as transformative "despite not having human-like consciousness or emotions." In the same post he put the likelihood of transformative AI within 20 years at "at least 10% with moderate robustness, and at least 1% with high robustness" (Some Background on Our Views Regarding Advanced Artificial Intelligence (Karnofsky, 2016)). Karnofsky later framed the stakes of transformative AI in his "Most Important Century" blog series (Open Philanthropy). Open Philanthropy rebranded as Coefficient Giving in November 2025, and the 2016 post is now hosted under that name (Open Philanthropy).
Impact definition versus capability definition
The distinguishing feature of the term is that it specifies an outcome rather than a threshold of ability. AGI and superintelligence are capability definitions: they specify a level of cognitive ability, such as human-level or beyond across domains, and invite unresolved debates over what counts as "general." Transformative AI is an outcome definition: it brackets the capability question and asks only whether the economic and institutional effects rise to revolution scale.
This framing allows a policymaker to reason about TAI's consequences without first settling whether a system is "really" AGI. Winter & Bullock explicitly bracket the capability question, writing that "whether 'AGI' or 'superintelligence' or 'powerful AI' or 'transformative AI' will ever arrive, and when ... is beyond the scope of this paper," and build their radical optionality argument on the mere possibility of TAI within roughly 10 years.
Later users have adapted the definition rather than adopting it verbatim. Toby Ord defines transformative AI as AI powerful enough to take over the world if misaligned, or to double the rate of technological progress (Broad Timelines), which combines an outcome criterion with a risk criterion. Dario Amodei describes transformative AI as "a country of geniuses in a datacenter" (Dario Amodei — The Urgency of Interpretability), a capability-flavored image that the Digitalist Papers essay by Agrawal and Gans and the scenario AI 2027 both borrow.
Use in the Digitalist Papers Vol. 2
The Digitalist Papers Volume 2 (Stanford, 2025) is organized around the concept of transformative AI; its essays ask how TAI reshapes labor, fiscal systems, information, market power, and geopolitics. Several sub-themes recur across the literature.
One is a labor-to-capital value shift: if TAI moves value creation from labor to capital, prosperity-sharing requires new mechanisms (AI Dividends (Universal Basic Capital, Digital Dividend, Global Dividend), AI Labor Disruption). A second is what Agrawal & Gans frame as a "genius supply shock," in which TAI provides abundant cheap cognitive labor and the binding constraint shifts to deciding what to solve. A third is fiscal disruption: Korinek & Lockwood argue that a labor-to-capital shift erodes the tax base. A fourth is governance under uncertainty, in which Bostrom's Open Global Investment, Winter & Bullock's radical optionality, and CSET's race-dynamics game theory all treat TAI as a possible future to be prepared for.
Institutional uptake
Beyond Open Philanthropy and the effective-altruism research community, the term has been adopted by frontier developers and economists. Anthropic's "Core Views on AI Safety" (around March 2023) sets out the company's view that transformative AI is likely within the coming decade (Core Views on AI Safety: When, Why, What, and How). The National Bureau of Economic Research organized an Economics of Transformative AI volume, for which chapters such as "The Coasean Singularity?" by Shahidi, Rusak, Manning, Fradkin and Horton were written in 2025 (The Coasean Singularity? Demand, Supply, and Market Design with AI Agents — Shahidi, Rusak, Manning, Fradkin, Horton (2025)). On July 13, 2026, more than 200 economists, executives and researchers released the 88-word statement "We Must Act Now," organized by Erik Brynjolfsson's Stanford Digital Economy Lab, calling on economists, policymakers and technology leaders to "act now to understand the economics of transformative AI" (We Must Act Now: A Statement on AI's Transformation of the Economy (July 2026)).
Debates and positions
Whether TAI is coming, and when, is contested. Estimates range from Acemoglu-style skepticism, which puts total factor productivity gains at no more than 0.66% over 10 years (see The Simple Macroeconomics of AI), to frontier-lab forecasts of transformation within 2 to 5 years (San Francisco Consensus). The term itself is agnostic on timing, though its users are not. Amodei has argued that transformative AI could arrive by 2026–2027 (Dario Amodei — The Urgency of Interpretability); Avital Balwit of Anthropic, writing in the Digitalist Papers, views it as "likely to arrive in the next 2–4 years" in the sense that it "exists in the lab and works well enough for customer use," with diffusion slower (The Digitalist Papers (Stanford, Volumes 1–2)); and Ord gives a median of 2038 with an 80% interval of 3 to 100 years from the time of writing (Broad Timelines).
A second debate concerns rhetorical inflation. Because "transformative" is evocative and unfalsifiable in the short term, the label can lend speculative scenarios weight that the underlying evidence does not establish. On this view, TAI functions as a frame for organizing analysis rather than as a forecast.
Empirical reference points for the debate have begun to appear in official statistics. The Federal Reserve published a FEDS note on July 17, 2026 assembling publicly available data to assess the AI buildout's economic impact (Source: federalreserve.gov). James Pethokoukis's July 22, 2026 essay drew on it to map the distance between San Francisco expectations of an economic discontinuity and the more gradualist forecasts of major banks, the Fed, and the CBO (Source: fasterplease.substack.com). See San Francisco Consensus, AI Bubble Debate.
Relationships
- related: AGI Timelines, Scaling Laws, Recursive Self-Improvement (RSI), Compressed 21st Century, AI Race Dynamics, AI Existential Risk
- instance-of-this: AI Labor Disruption, AI Dividends (Universal Basic Capital, Digital Dividend, Global Dividend) (downstream consequences debated under the TAI frame)
- discussed-in: The Digitalist Papers (Stanford, Volumes 1–2), Radical Optionality: Governing Transformative AI Under Uncertainty, AI 2027, Broad Timelines, Core Views on AI Safety: When, Why, What, and How, Dario Amodei — The Urgency of Interpretability, We Must Act Now: A Statement on AI's Transformation of the Economy (July 2026), The Coasean Singularity? Demand, Supply, and Market Design with AI Agents — Shahidi, Rusak, Manning, Fradkin, Horton (2025)
- related: Artificial General Intelligence (AGI) (capability-defined counterpart), Open Philanthropy (originator of the term), Toby Ord
- depends-on: Some Background on Our Views Regarding Advanced Artificial Intelligence (Karnofsky, 2016) — origin text of the term and its two definitions; author Holden Karnofsky