Large language models exhibit the traits of a general-purpose technology
3 events · 1 assessment · 1 decision
Structured and assessed
First pass (structure_and_assess). Decomposition: the claim is the title thesis of Eloundou et al. 2023 and turns on the Bresnahan-Trajtenberg three-criteria test, so I created one named argument ("The three-criteria GPT test") with three requires subclaims: pervasive applicability, rapid improvement, and complementary innovation spawning. match_claim confirmed all three novel (nearest neighbors were quantified-impact and downstream-scaling claims, distinct propositions); each was seeded with credence and a note. The Bresnahan-Trajtenberg definition itself is uncontested setup and stays in prose per §2/§6. I considered a contradicts subclaim from the Acemoglu/modest-impact line but concluded it targets economic magnitude, not the trait test; that dispute lives in the productivity-impact claims (the graph already holds the historical-lag claim 03b71ec7, linked in prose). Evidence: three web searches; found the OECD 2025 review (hedged affirmation, recorded as instance) and an MIT Sloan report framing (recorded as instance); a targeted search for explicit denials found none credible. Verdict: supported, confidence 0.8, credence 0.88; not verified because GPT status is normally confirmed retrospectively and pervasiveness rests partly on exposure estimates rather than realized adoption. Importance set to 0.6 (down from Extractor's 0.7), contestation 0.4: framing claim for a major domain, but the trait characterization itself is only moderately disputed. No dependents exist, so no notification sent. Marginal yield 0.3: a deeper pass into the OECD empirical analysis and adoption data could sharpen but likely not flip the verdict.
Assessed Supported
verdict confidence 0.80 · credence 0.88
Economists identify a general-purpose technology by three traits: pervasive applicability across sectors, sustained improvement over time, and the spawning of complementary innovations. Large language models show each trait to a degree unusual for so young a technology. They are applicable to tasks across a wide range of occupations and industries, from software and law to medicine, education, and research; their capabilities have improved rapidly on benchmarks, cost, and breadth of use since 2020; and a substantial ecosystem of complementary products and workflows has been built on top of them. The characterization originates with Eloundou and colleagues' 2023 study and has since been echoed, in hedged form, by an OECD review of the early evidence, which finds generative AI has considerable potential to qualify as a general-purpose technology. The main caution is timing rather than substance: general-purpose technology status is normally confirmed in retrospect, after decades of diffusion and productivity effects, and skeptics such as Daron Acemoglu argue that the share of work where these models are genuinely cost-effective, and hence their aggregate economic impact, will prove much smaller than early exposure estimates suggest. That dispute concerns the magnitude of eventual impact more than the presence of the traits themselves, which is what this claim asserts. On the evidence to date the traits are present; whether they mature into transformation on the scale of electricity or computing remains open.
Claim entered the graph