Page 101 - ISC PROCEEDINGS 21.4

improves through increasing data and algorithmic refinement; and enables downstream
innovations in workflow design, organizational structure, and business model
reconfiguration (Goldfarb et al., 2019). A critical policy implication of GPT theory is the
productivity J-curve, articulated by Brynjolfsson, Rock, and Syverson (2018) at MIT. GPTs
typically generate measured productivity stagnation before benefits materialize, because
realizing a GPT's full potential requires complementary investments in organizational
restructuring, workforce retraining, and new business processes. The electrification
analogy is instructive: it took two to three decades after the widespread installation of
electric motors before factory layouts were sufficiently reorganized to capture
electricity's full productivity potential (David, 1990). The same lag appears to be operating
with AI, implying that short-term pilots are insufficient; sustained, systemic
transformation is required (Brynjolfsson et al., 2021). This paper argues that Vietnam is
currently positioned in the early, pre-productivity-realization phase of the AI GPT cycle,
and that deliberate policy action is required to accelerate passage through the J-curve.
2.2. The productivity paradox and its relevance to developing economies
Brynjolfsson (1993) introduced the productivity paradox to describe how rapid
computer proliferation in the 1980s-1990s failed to accelerate aggregate productivity
statistics, a paradox partially resolved by the mid-to-late 1990s productivity boom, which
Brynjolfsson and Hitt (2000) attributed to eventual IT-organizational change
complementarity. Brynjolfsson et al. (2018) argue that a second productivity paradox is
now emerging with AI: transformative capabilities coexisting with sluggish aggregate
productivity statistics.
MIT economist Daron Acemoglu offers an influential counterargument. In The
Simple Macroeconomics of AI (2024), he develops a task-based model demonstrating that
AI's macroeconomic effects are bounded by the fraction of tasks AI can actually automate
versus augment, projecting only a modest 1.1-1.6% GDP increase over a decade and
annual TFP growth of approximately 0.05% (Acemoglu, 2024). His model implies that AI
enthusiasm must be tempered by rigorous attention to whether AI augments worker
capabilities or merely automates and displaces. The tension between optimistic GPT-
based projections and Acemoglu's more cautious task-based analysis has direct policy
implications for Vietnam: AI deployment must be structured to maximize human-AI task
complementarity rather than labor displacement, particularly critical given that
approximately 61% of Vietnam's employed population is engaged in informal sectors
where automation effects risk being regressive.
2.3. The analytical core of total factor productivity
Total Factor Productivity (TFP), conceptualized as the Solow residual, the portion of
output growth unexplained by capital and labor inputs, is the standard metric for
measuring technology's and organizational innovation's contribution to productivity
(Solow, 1956). In growth accounting terms:
where (Y) denotes output, (K) capital, (L) labor, and (A) the TFP index capturing
technological efficiency. For Vietnam, CSIRO Aus4Innovation research found that TFP
contributed only approximately 20% of total output-per-worker growth, with the
remainder attributable to capital accumulation and labor expansion (CSIRO Research,
2025). This is consistent with Dinh et al. (2023), who found a declining average trend in
TFP's contribution to Vietnamese manufacturing output, accompanied by widening
productivity gaps between technology-frontier firms and industry laggards.

100