The 2026 Prize in Economic Sciences recognised Robert E. Hall and Michael Woodford for the microeconomic foundations of macroeconomic stabilisation. This article asks what their framework implies for artificial intelligence (AI) in the ten largest economies by 2026 nominal GDP (IMF World Economic Outlook, October 2026): the United States, China, Germany, Japan, the United Kingdom, India, France, Russia, Italy and Brazil. I build a country index of occupational AI exposure from ILOSTAT employment by ISCO-08 major group and the AI Occupational Exposure scores of Felten, Raj and Seamans, and combine it with Stanford HAI capacity data, the IMF AI Preparedness Index, IMF fiscal data and BIS policy rates. Three findings emerge. First, AI capacity is far more concentrated than output: the United States and China produce 64% of the group's GDP but account for 90% of its AI private investment, and occupational exposure is unrelated to AI capacity (Spearman ρ = 0.02, exact permutation p = 0.97). Second, two pre-specified hypotheses are not supported: the most exposed economies do not have systematically less fiscal or monetary space (ρ = 0.14, p = 0.71), and lagged AI investment does not predict labour-productivity growth in 2018–2025 (randomisation p = 0.86). Third, a calibrated exercise based on the consumption Euler equation shows that, when the intertemporal elasticity of substitution is below one, an AI-driven rise in trend growth raises the interest–growth differential. With an elasticity of 0.5 and the mid-range OECD productivity scenario, the debt-stabilising primary balance rises by 1.1 percentage points of GDP in Japan and by 0.7–0.8 points in the United States, Italy and France. The gain would widen monetary room above the lower bound while narrowing fiscal room where public debt is highest.

Whether Europe has fallen behind the United States and China is central to European economic policy, yet the debate pools very different activities into one verdict. This study separates three stages of technological competitiveness: knowledge (the science base), creation (turning knowledge into firms, capital, frontier technology and exportable manufacturing) and deployment (adoption and diffusion). It measures the European Union's position in business, artificial intelligence and renewable energy against both rivals with 14 indicators from six public sources. Each indicator is expressed as a log ratio and analysed with HAC-robust trend tests, a pre-specified test for a change after 2020 and leave-one-out pillar indices. In 2023 the EU stood at 1.34 times the United States in knowledge and 1.23 times in deployment, but at 0.33 in creation. Against China it stood at 0.55, 0.29 and 1.52, respectively, and all three ratios declined by 4-12% a year from 2017. The EU's share of R&D among the world's 1,000 largest corporate investors fell from 24.7% to 16.4% between FY2010 and FY2024, and its share of ICT corporate R&D from 16.2% to 6.0%. Europe has fallen behind, but mainly in conversion rather than in science or adoption.

Between January and September 2026, Fund E, the most conservative of Chile's five mandatory pension fund types, lost about ten per cent in real terms, its second-worst result on record, while the funds' positions in foreign interest-rate derivatives swung to a reported valuation loss of about US$7.7 billion. This article asks why standard Value-at-Risk (VaR) measures did not warn of the loss. Using daily unit values for the seven pension fund administrators (AFPs) from January 2015 to October 2026, Chilean and US government yields and the exchange rate, I estimate the funds' effective duration to US long-term yields and backtest six VaR models at daily and monthly horizons. Four hypotheses were stated before estimation. First, Fund E's sensitivity to the US ten-year yield rose from about 0.2 years in 2016-2019 to 7.2 years in 2026 (change 6.3 years, z = 11.4), rising from late 2022 and accelerating from late 2023, consistent with the build-up of receive-fixed dollar swaps; US rates account for-7.9 of the-6.5 percentage points of nominal return in January-September 2026. Second, contrary to the hypothesis, daily 99% VaR models broadly met their coverage in 2026 (three to six exceptions against about two expected); the loss was a persistent drift, not a sequence of tail days. Third, an exposure-adjusted VaR, which applies the current factor exposure to a long history of rate moves, rose to 7.4% of the fund for a one-month horizon in June 2026, more than twice the 3.5% given by historical simulation scaled by the square root of time, which was then at its lowest level since early 2024; only the exposure-adjusted measure covered the September 2026 loss of 5.1%, in all seven AFPs and in nine window combinations. Fourth, differences in pre-shock exposure across AFPs did not significantly predict the size of the loss (Spearman ρ = 0.50, exact p = 0.13). The gap between what a volatility-based VaR measures and what the portfolio is exposed to supports the stress-based derivatives limit that the regulator introduced in April 2026.

Between January and September 2026, Fund E, the most conservative of Chile’s five mandatory pension fund types, lost about ten per cent in real terms, its second-worst result on record, while the funds’ positions in foreign interest-rate derivatives swung to a reported valuation loss of about US$7.7 billion. This article asks why standard Value-at-Risk (VaR) measures did not warn of the loss. Using daily unit values for the seven pension fund administrators (AFPs) from January 2015 to October 2026, Chilean and US government yields and the exchange rate, I estimate the funds’ effective duration to US long-term yields and backtest six VaR models at daily and monthly horizons. Four hypotheses were stated before estimation. First, Fund E’s sensitivity to the US ten-year yield rose from about 0.2 years in 2016–2019 to 7.2 years in 2026 (change 6.3 years, z = 11.4), rising from late 2022 and accelerating from late 2023, consistent with the build-up of receive-fixed dollar swaps; US rates account for − 7.9 of the − 6.5 percentage points of nominal return in January–September 2026. Second, contrary to the hypothesis, daily 99% VaR models broadly met their coverage in 2026 (three to six exceptions against about two expected); the loss was a persistent drift, not a sequence of tail days. Third, an exposure-adjusted VaR, which applies the current factor exposure to a long history of rate moves, rose to 7.4% of the fund for a one-month horizon in June 2026, more than twice the 3.5% given by historical simulation scaled by the square root of time, which was then at its lowest level since early 2024; only the exposure-adjusted measure covered the September 2026 loss of 5.1%, in all seven AFPs and in nine window combinations. Fourth, differences in pre-shock exposure across AFPs did not significantly predict the size of the loss (Spearman ρ = 0.50, exact p = 0.13). The gap between what a volatility-based VaR measures and what the portfolio is exposed to supports the stress-based derivatives limit that the regulator introduced in April 2026. JEL classification: G23; G28; G32; G17; E43

Background. The graviton, the hypothetical quantum of gravity, is often invoked in speculation about future artificial intelligence (AI) and immersive virtual worlds, yet a single graviton is widely regarded as undetectable and its only quantified observable property is an upper bound on its mass. Whether massivegraviton physics could even be represented inside the computational systems that run virtual worlds and AI has not been tested. Methods. Two computational experiments with known truth were run. Experiment 1 asks whether the Yukawa correction implied by a graviton at the current 90% credibility bound (mg ≤ 2.42 × 10-23 eV/c 2) is representable in six floating-point formats, using a 60-digit reference, a naive in-format evaluation, and a velocity-Verlet orbit integrator of the kind used in physics engines. Experiment 2 injects the massive-graviton dispersion phase into noise-free inspiral waveforms and measures, for each format, the ratio between the time-and-phase-aligned dispersion residual and the rounding error, and the test AUC of a neural-network discriminator trained on rounded inputs (90 training runs). Results. The correction is unrepresentable at every virtual-world scale: it first becomes storable in float64 at 5.4 × 10 8 m and in float32 at 1.25 × 10 13 m (about 84 AU). Naive in-format evaluation manufactures phantom corrections up to 9 × 10 16 times the true value, and a float32 orbit at 1 AU deviates by 834 km where the true effect is 8 m. In AI formats the opposite holds: at the bound every format, including 8-bit FP8, preserves the residual (representational SNR 9.7 to 2 × 10 7), with representational floors 3 to 4,500 times below the bound for formats from FP8 to float32; classifier accuracy did not degrade with precision. Conclusions. Precision is the binding constraint for virtual-world engines but not for AI, where noise and learnability bind first. Any claim that a graviton-scale effect appears in a simulated world should be accompanied by an explicit representability check.

Robots driven by artificial intelligence now fail in public: robotaxis block ambulances, surgical instruments break inside patients and humanoid robots lose control at exhibitions. What society learns about these failures depends on which incident regime records them. This article compares three regimes for embodied AI between 2021 and 2026: the mandatory crashreporting order of the US National Highway Traffic Safety Administration (8,135 incidents involving automated driving systems or Level 2 driver assistance), the Food and Drug Administration's adverse-event register for computer-assisted surgical systems (127,627 reports since 2015) and the media-derived AI Incident Database (1,734 incidents). Fifty-four road-vehicle incidents in the media-derived database were hand-coded and linked to mandatory reports. Twenty of the 21 US crashes in the media-derived database were also reported to the regulator, but only 21 of 8,135 mandatorily reported incidents (0.26%) reached the mediaderived database. Contrary to the pre-specified expectation, media visibility rose with severity: 3.1% of fatal or serious incidents against 0.37% of the rest (odds ratio 8.7; Fisher exact p < 0.001), a gradient that keeps its direction across nine robustness checks. Almost half of the media-recorded US events (19 of 40) were operational failures without a crash, which a crashtriggered order cannot capture by design. In the surgical register, a fourfold rise in malfunction reports in 2025 is concentrated in two wear-related problem codes filed by the manufacturer, while injury reports barely changed. Humanoid robots, the most visible failures of 2026, fall under no reporting regime. Each regime sees a different class of failure, and none sees the robots now entering public space.