## wpiea2026180-source-pdf

## Source details

**Canonical URL:** [wpiea2026180-source-pdf](https://www.imf.org/-/media/files/publications/wp/2026/english/wpiea2026180-source-pdf.pdf)

## Other formats

- [Markdown version](/-/media/files/publications/wp/2026/english/wpiea2026180-source-pdf.pdf.md)
- [Structured JSON version](/-/media/files/publications/wp/2026/english/wpiea2026180-source-pdf.pdf.json)

---

### Estimation method and identification
- Identification elements:
  - Firms’ ex ante exposure proxied by pre-2021 EU revenue share and AI hiring growth.
  - Narrow event window around the April 21, 2021 Commission proposal to isolate announcement-driven repricing.
- Baseline cross-sectional regression (Equation 1):
  - CAR_i[τ1,τ2] = α + β1 (EU×AI)_i + β2 AI_i + β3 EU_i + Γ X_i + δ_s + ε_i
  - EU and AI standardized to mean zero and unit variance within the analysis sample; X_i denotes financial controls; δ_s are FF48 industry fixed effects.
- CAR construction:
  - CAR_i[τ1,τ2] = Σ_{τ=τ1}^{τ2} \^ε_{iτ}, where \^ε_{iτ} is residual from Fama–French three-factor model estimated over trading days [−250,−11] with minimum 120 valid days (median in-sample R^2 = 0.34).
- Four CAR windows reported: [0, +2] primary, [0, +1], [−1, +1], [−2, +2].
- Multi-event analysis: CARs re-estimated around subsequent legislative milestones holding exposure measures fixed at April 2021 values.

### Main empirical findings (announcement returns and interpretation)
- EU×AI Growth interaction:
  - Positive and significant at the 1% level in every window; headline: +40 bps on CAR[0, +2].
- Panel A detailed EU×AI Growth coefficients (CAR in bps; standard errors in parentheses):
  - CAR[0, +1]: +31.8 ∗∗∗ (9.0)
  - CAR[0, +2]: +39.7 ∗∗∗ (10.3)
  - CAR[−1, +1]: +43.4 ∗∗∗ (7.2)
  - CAR[−2, +2]: +45.2 ∗∗∗ (8.8)
- Panel A other main coefficients (CAR in bps; SE in parentheses):
  - EU Revenue Share:
    - CAR[0, +1]: +3.2 (9.4)
    - CAR[0, +2]: +2.4 (13.9)
    - CAR[−1, +1]: +15.1 ∗ (8.5)
    - CAR[−2, +2]: +22.5 ∗∗ (11.1)
  - AI Growth (standalone):
    - CAR[0, +1]: −21.9 ∗∗∗ (6.0)
    - CAR[0, +2]: −16.7 ∗∗ (6.4)
    - CAR[−1, +1]: −37.1 ∗∗∗ (8.0)
    - CAR[−2, +2]: −32.2 ∗∗∗ (7.6)
- Panel B: marginal effect of EU exposure on CAR[0, +2] (∂CAR/∂EU = β3 + β1 AI):
  - 1 SD below mean: −37.3 bps (FF48-clustered p 0.092; Wild-cluster p 0.163; N 769)
  - At mean: +2.4 bps (FF48-clustered p 0.862; Wild-cluster p 0.897; N 769)
  - 1 SD above mean: +42.1 bps (FF48-clustered p 0.0005; Wild-cluster p 0.053; N 769)
- Economic aggregation and magnitude:
  - High EU × high AI firms (above median in both): lost 9 bps (−$16 billion loss on a $17.5 trillion base).
  - Low EU × high AI firms: lost 58 bps (−$47 billion on a $8.1 trillion base).
  - Differential: 49 bps.
  - Applying 49 bps differential to high EU group base implies jurisdictional commitment shielded roughly $86 billion (49bps×$17.5 trillion base) in shareholder value at announcement.
- Timing across milestones:
  - EU×AI interaction large and significant at 1% only at the April 2021 Commission proposal; indistinguishable from zero at subsequent milestones.

### Pre-trend, event dynamics, and persistence
- Pre-trend (symmetric 81-day window [−40, +40]; factor loadings estimated over [−291,−41]):
  - Pre-event EU×AI coefficients (bps): CAR[−40,−1]: −25.2 (SE 59.5; t −0.42; p 0.674; N 769); CAR[−40,−21]: −10.9 (SE 41.8; t −0.26; p 0.795; N 769); CAR[−20,−1]: −14.3 (SE 23.4; t −0.61; p 0.545; N 769); CAR[−10,−1]: 4.3 (SE 18.6; t 0.23; p 0.818; N 769); CAR[−5,−1]: 14.6 (SE 11.3; t 1.29; p 0.205; N 769).
- Post-event cumulative path:
  - CAR[0, +2]: 39.7 ∗∗∗ (SE 10.3; t 3.87; p < 0.001; N 769)
  - CAR[0, +10]: 86.7 ∗∗∗ (SE 32.2; t 2.70; p 0.010; N 769)
  - CAR[0, +40]: 174.0 ∗∗∗ (SE 54.4; t 3.20; p 0.003; N 769)
  - Cumulative coefficient path near zero pre-event, rises sharply after day 0 and stabilizes at approximately +100 bps by day +40.
- Long-horizon CARs (OLS of CAR):
  - CAR[0, +1]: 31.8 bps; CAR[0, +2]: 39.7 bps; CAR[0, +5]: 43.3 bps; CAR[0, +10]: 86.7 bps; CAR[0, +20]: 113.0 bps; CAR[0, +30]: 126.3 bps; CAR[0, +40]: 174.0 bps (all significant at 1%).
- Local projections: modest temporary partial reversals at h = 10 and h = 20 but overall no long-run reversal; θ_h estimates noisy and often statistically indistinguishable from zero.

### Randomization inference and placebo evidence
- Placebo distribution from 483 trading dates (Dec 2020–Dec 2022, excluding ±10 trading days around April 21, 2021).
- EU×AI Growth (actual, placebo SDs, RI p-values):
  - CAR[0, +1]: Actual +31.8 bps; Actual t +3.54; Placebo SD 15.2 bps; RI p (coeff.) 0.039; RI p (t-stat) 0.012.
  - CAR[0, +2]: Actual +39.7 bps; Actual t +3.87; Placebo SD 18.5 bps; RI p (coeff.) 0.041; RI p (t-stat) 0.017.
  - CAR[−1, +1]: Actual +43.4 bps; Actual t +6.06; Placebo SD 18.5 bps; RI p (coeff.) 0.031; RI p (t-stat) 0.002.
  - CAR[−2, +2]: Actual +45.2 bps; Actual t +5.15; Placebo SD 24.3 bps; RI p (coeff.) 0.062; RI p (t-stat) 0.000.
- EU Revenue Share alone falls well within its placebo distribution (e.g., CAR[0, +2] RI p (coeff.) 0.882).
- Placebo means for all three coefficients centered near zero.

### AI Growth versus AI Level (horse race)
- Replacing AI hiring growth with AI hiring level or including both:
  - EU×AI Level coefficients (CAR[0, +2]): 0.2 (SE 14.4) — insignificant.
  - When both interactions included, EU×AI Growth remains essentially unchanged and significant:
    - CAR[0, +2]: 39.6 ∗∗ (SE 17.8) for EU×AI Growth (with level included).
  - AI Level main effects insignificant; AI Growth main effect remains negative and similar in pattern.
- Interpretation: effect operates through AI growth (change) rather than AI level (stock).

### Robustness checks (specifications and sensitivity)
- Robustness maintained when:
  - Using Fama–French five-factor model (FF5): EU×AI Growth CAR[0, +2] 40.8 ∗∗∗ (SE 10.5).
  - Replacing EU revenue share with alternative historical windows (e.g., avg 2014–2018: 40.7 ∗∗∗ (10.3)).
  - Excluding potential tax-haven countries (Ireland, Luxembourg, Netherlands): 41.0 bp (p-value = 0.082).
  - Excluding April 22 (US capital gains tax proposal) and controlling for equity duration, yield and VIX news: remaining days positive and significant.
  - Removing largest DFBETA outliers: deleting five and ten largest DFBETA observations yields 37.2 and 34.5 bp, respectively.
- Pre-trend, earnings-announcement contamination, and PEAD-adjusted specifications do not overturn inference; headline [0, +2] EU×AI coefficients remain in a narrow band (roughly +32 to +40 bps across nested specs).

### Leak timing
- April 14, 2021 leak windows:
  - CAR[−5,−3]: EU×AI Growth = 9.1 (8.6) — positive but insignificant.
  - CAR[−5,−1]: EU×AI Growth = 14.6 (11.3) — positive but insignificant.
  - Combined CAR[−5, +2]: EU×AI Growth = 54.3 ∗∗∗ (13.5) — significant and larger than official-window estimate, implying the official-window CAR[0, +2] is conservative.

### Sample and descriptive statistics (main analysis sample N = 771)
- AI_Growth (ppts/yr): Mean = 1.432; SD = 2.819; Median = 0.532; P75 = 2.202.
- EU_Share: Mean = 0.098; SD = 0.089; Median = 0.081; P75 = 0.135.
- CAR[0,+2] (bps): Mean = −18.861; SD = 326.153; Median = −11.787; P75 = 130.964.
- Market Cap (USD bn): Mean = 21.368; SD = 57.590; Median = 4.515.
- Extensive-margin sample N = 1,264 statistics shown in Appendix.

### Jurisdictional commitment versus scale (Section 6.1)
- Core finding: EU revenue share × AI Growth predicts positive abnormal returns; effect is not explained by firm size or broad international exposure.
- Tests replacing EU revenue share with scale proxies (selected coefficients, CAR[0, +2]):
  - EU Share × AI Growth: 39.7 ∗∗∗ (10.3)
  - EU USD × AI Growth: −2.8 (4.6)
  - log(EU USD) × AI Growth: −4.4 (10.0)
  - log(Mcap) × AI Growth (horse race): −13.4 (9.1)
- Scale decomposition (CAR[0, +2], bps):
  - Baseline EU × AI: 39.7 ∗∗∗ (10.3)
  - + log(Mcap): EU × AI = 41.5 ∗∗∗ (10.7); log(Mcap) = −13.4 (9.1)
  - + PP&E/AT: EU × AI = 34.0 ∗∗∗ (9.5); PP&E/AT = −48.8 ∗∗∗ (14.6)
  - + log(Employ): EU × AI = 42.1 ∗∗∗ (11.3); log(Employ) = −12.8 ∗ (7.1)
  - Saturated: EU × AI = 35.4 ∗∗∗ (9.4); PP&E/AT = −47.5 ∗∗∗ (15.5)
- Interpretation:
  - Jurisdictional capital effect requires relative EU importance (EU revenue share), not absolute scale; physical-capital intensity (PP&E/AT) is strongly negative, R&D intensity essentially null.

### Political engagement and lobby tests (Section 6.2)
- Political engagement indicators (pre-event): Commissioner meetings (73 firms), Lobby registration (74 firms), AI White Paper consultation (17 firms), Country-level lobby (86 firms), EU-level trade-association membership (45 firms), Any political engagement dummy.
- Adding engagement controls, horse races, and triple interactions:
  - EU × AI Growth remains virtually unchanged from baseline (~+40 bps on CAR[0, +2]) across specifications (example: Commissioner meetings control: 38.6 ∗∗∗ (9.8); triple: 43.2 ∗∗∗ (10.7)).
  - None of the political engagement variables absorb the EU×AI effect; post-April-2021 lobbying outcomes: EU×AI does not predict additional lobbying (all effects statistically insignificant), while AI hiring growth and firm size strongly predict lobbying.
- Conclusion: political engagement does not substitute for EU presence in generating the jurisdictional capital effect.

### Foreign revenue share versus EU revenue share (Section 6.3)
- Adding Foreign × AI Growth (non-US, non-EU revenue share) to baseline (CAR[0, +2], bps):
  - EU × AI Growth: 34.0 ∗∗∗ (12.0)
  - Foreign Income × AI: 0.3 (10.6) — economically small and insignificant.
  - EU Revenue Share (main effect): 16.4 ∗ (9.5)
  - AI Growth main effect: −22.4 ∗∗∗ (7.9)
- Conclusion: the jurisdictional capital effect is EU-specific; general foreign exposure does not generate the effect.

### Amplifiers of jurisdictional capital
- High-risk AI buildup (HRAI; Section 7.1):
  - Triple interaction EU × AI × HRAI (CAR in bps; std. err. in parentheses):
    - CAR[0, +2]: 15.02 ∗∗∗ (2.81) for EU × AI × HRAI
    - Two-way EU × AI from same regression: 31.69 ∗∗ (2.03)
  - Binary split (Panel B):
    - EU × AI, HRAI > 0 (N = 202): CAR[0, +2] 34.86 ∗∗∗ (3.46) and larger effects at wider windows (e.g., CAR[−1, +1]: 72.17 ∗∗∗ (6.95))
    - EU × AI, HRAI = 0 (N = 554): CAR[0, +2] 27.83 (1.49)
  - Interpretation: each additional SD of HRAI adds ~+15 bps to baseline effect on CAR[0, +2]; effect concentrated among firms scaling into high-risk AI exposure.
- Stability of EU commitment (CV of EU share; Section 7.2):
  - Low-CV (stable) firms: EU×AI coefficient significantly positive and close to full-sample estimate.
  - High-CV (volatile) firms: EU×AI insignificant across windows.
  - Interpretation: stable EU presence signals accumulated jurisdictional capital.
- Geographic concentration (Section 7.3; HHI and Top-1 splits):
  - High-concentration (HHI above median, N382) EU×AI coefficients (bps; std. err. in parentheses): 47.0 ∗∗∗ (13.9), 63.9 ∗∗∗ (19.7), 75.1 ∗∗∗ (12.3), 86.0 ∗∗∗ (15.4) across windows.
  - Low-concentration half: 21.0 ∗∗ (7.8), 11.4 (11.5), 13.8 (13.1), 5.9 (17.1).
  - Top-1 split yields similar patterns (High-Top1: 47.2 ∗∗∗, 64.5 ∗∗∗, 75.3 ∗∗∗, 87.3 ∗∗∗).
  - Appendix E country decomposition: Germany dominates high-HHI subsample (Germany carries 33.6% of total EU revenue in high-HHI half and is top-1 for 83% of high-HHI firms), effect not driven by a single small country.
- Prior compliance experience (Section 7.4):
  - Compliance hiring growth measures constructed from Lightcast with 77 patterns across 11 categories.
  - Subsample (Panel A, Broader measure), top half: EU × AI Growth ≈ 40.9 ∗∗∗, 49.5 ∗∗∗, 59.0 ∗∗∗, 57.5 ∗∗∗ (N=376).
  - Bottom half: 23.6 ∗, 15.4, 24.6, 15.3 (N=379).
  - Horse race (Table 12): replacing EU revenue share with compliance (C) yields near-zero coefficients; EU × AI remains stable when both included.
  - Interpretation: jurisdictional capital is prominent among firms with prior IT/GDPR-era compliance scaling; compliance hiring does not mechanically replace EU revenue channel.

### Post-event within-firm reallocation (Section 8)
- Sample: HighAI ∩ HighEU subsample contains 156 firms.
- Peer set measure LowEUPeer_i defined as share of HighAI peers that are LowEU.
- Outcome: ∆realloc_i = change in logit(EU_Share) (2015–2020 → 2021–2025).
- Results (Table 13):
  - LowEUPeer coefficient on ∆realloc: +0.428 ∗∗ (t = 1.80; Wild-bs p = 0.047; Observations 156).
  - ∆ log(EU_Share): +0.386 ∗ (t = 1.76; Wild-bs p = 0.057; Observations 156).
  - ∆ log(EU_Rev): +0.210 (t = 0.58; Wild-bs p = 0.586; Observations 156).
- Magnitude interpretation:
  - +0.43 log-points implies approximately a 21% (= 0.428/2) larger expansion of EU revenue relative to non-EU revenue for median movement.
  - For median HighEU∩HighAI firm with pre-event EU share 13.2%: approximate absolute EU share gain ≈ 2.8 percentage points.
- Placebo in pooled non-HighEU∩HighAI sample (N=443): LowEUPeer coefficient = −0.001 (t = 0.00; Wild-bs p = 0.996) — null effect.
- Interpretation: evidence consistent with jurisdictional capital translating into real reallocation of within-firm EU market share when firms combine HighAI and HighEU.

### Implications and conclusion (Sections 9–10)
- For investors: markets price jurisdictional capital as an intangible asset; the EU×AI interaction produced approximately +40 bps announcement return around April 21, 2021.
- For regulators: streamlined requirements could make compliance capability more broadly accessible; text notes June 2026 updates include measures to simplify and delay certain regulatory requirements.
- For firms: longstanding presence in a regulated jurisdiction can become valuable when new rules arrive; benefits can accrue to smaller firms that invested in durable operating depth in the jurisdiction.
- Overall conclusion:
  - Using predetermined exposure measures across about 800 US-listed firms, firms with established EU market presence captured a positive announcement return on the EU×AI interaction of approximately +40 bps.
  - Three lines of evidence support interpretation as priced jurisdictional capital:
    - Effect operates through EU revenue share rather than firm size or international diversification.
    - Political access measures do not absorb the effect.
    - Effect amplified by high-risk AI capacity buildup, stable and geographically concentrated EU presence, and prior GDPR-era compliance experience.
  - Suggestive evidence that jurisdictional capital affects real activity: HighAI & HighEU firms facing LowEU peers subsequently increase EU revenue share.

_Italic source: Excerpt from IMF working paper sections 4.1, 6.1, 7.3, 7.4, 8, 9–10 and Appendices in wpiea2026180-source-pdf._

### 4.1  Estimation Method

### 4.1 Estimation Method

### Identification strategy and variables
- Two identification elements:
  - Firms’ ex ante exposure to the Act proxied by pre-2021 EU revenue share and AI hiring growth.
  - A narrow event window around the April 2021 proposal to isolate announcement-driven repricing.
- Baseline cross-sectional regression of cumulative abnormal returns on firms’ EU revenue share and AI hiring growth.
- EU (EU revenue share) and AI (AI hiring growth) are standardized to mean zero and unit variance within the analysis sample.
- X_i denotes financial controls described in Section 3; δ_s are Fama-French-48 industry fixed effects.
- Key regression (Equation 1):
  - CAR_i[τ1,τ2] = α + β1 (EU×AI)_i + β2 AI_i + β3 EU_i + Γ X_i + δ_s + ε_i
- Interaction interpretation:
  - EU revenue share measures historical importance of the EU market to the firm (jurisdictional capital).
  - AI hiring growth captures accumulation of AI-related growth options; markets should respond to firms actively building AI capabilities.
  - Key hypothesis: β1 > 0 (firms with deeper EU commitment and high AI growth better positioned to navigate the Act).
- Cumulative abnormal return:
  - CAR_i[τ1,τ2] = Σ_{τ=τ1}^{τ2} \^ε_{iτ}, where daily abnormal return \^ε_{iτ} is the residual from the Fama–French three-factor model:
    - r_it − r_ft = α_i + η_{1i} (Mkt-RF)_t + η_{2i} SMB_t + η_{3i} HML_t + ε_it
    - Estimated over trading days [−250,−11] relative to the event with a minimum of 120 valid days (median in-sample R^2 = 0.34).
- Four CAR windows reported: [0, +2] primary, [0, +1], [−1, +1], [−2, +2].
- Multi-event analysis: re-estimate CARs around subsequent legislative milestones holding exposure measures fixed at April 2021 values.

### Empirical sample visualization (descriptive)
- Sample of 771 firms plotted by EU revenue share (avg FY2015–2019, horizontal axis) and AI hiring growth (annualized change CY2015→2019, vertical axis).
- Correlation between the two measures: ρ = −0.038.
- FF48 industry-group and selected large technology firm plots shown (Figure 3).

### Event timing and leak consideration
- April 14, 2021 leak existed; leak-only short windows CAR[−5,−3] and CAR[−5,−1] show positive but insignificant EU×AI coefficients.
- Combined window CAR[−5, +2] gives a significant 54.3 bps EU×AI coefficient.
- Primary CAR[0, +2] interpreted as the official-release component; combined-window estimate larger, making official-window magnitude conservative.

---

### Main results (summary of empirical findings)
- EU×AI Growth interaction positive and significant at the 1% level in every window:
  - +40 bps on CAR[0, +2] (Panel A narrative).
- Detailed Panel A regression coefficients (CAR in bps):
  - EU × AI Growth:
    - CAR[0, +1]: +31.8 ∗∗∗ (SE 9.0)
    - CAR[0, +2]: +39.7 ∗∗∗ (SE 10.3)
    - CAR[−1, +1]: +43.4 ∗∗∗ (SE 7.2)
    - CAR[−2, +2]: +45.2 ∗∗∗ (SE 8.8)
  - EU Revenue Share:
    - CAR[0, +1]: +3.2 (SE 9.4)
    - CAR[0, +2]: +2.4 (SE 13.9)
    - CAR[−1, +1]: +15.1 ∗ (SE 8.5)
    - CAR[−2, +2]: +22.5 ∗∗ (SE 11.1)
  - AI Growth (standalone):
    - CAR[0, +1]: −21.9 ∗∗∗ (SE 6.0)
    - CAR[0, +2]: −16.7 ∗∗ (SE 6.4)
    - CAR[−1, +1]: −37.1 ∗∗∗ (SE 8.0)
    - CAR[−2, +2]: −32.2 ∗∗∗ (SE 7.6)
  - Industry FE (FF48): Yes; Firm controls: Yes; Observations: 769 (Panel A).
- Panel B: Marginal effect of EU exposure on CAR[0, +2] (∂CAR/∂EU = β3 + β1 AI):
  - 1 SD below mean: −37.3 bps (FF48-clustered p 0.092; Wild-cluster p 0.163; N 769)
  - At mean: +2.4 bps (FF48-clustered p 0.862; Wild-cluster p 0.897; N 769)
  - 1 SD above mean: +42.1 bps (FF48-clustered p 0.0005; Wild-cluster p 0.053; N 769)
- Interpretation:
  - A one-standard-deviation increase in pre-event AI hiring growth raises the marginal effect of EU revenue share by 39.7 basis points on CAR[0, +2].
  - Standalone AI hiring growth effect is significantly negative but not robust across some checks; standalone EU revenue share effect is economically small and not consistently significant.
- Economic magnitude (aggregation to market capitalization):
  - High EU × high AI firms (above median in both): lost 9 bps (−$16 billion loss on a $17.5 trillion base).
  - Low EU × high AI firms: lost 58 bps (−$47 billion on a $8.1 trillion base).
  - Differential: 49 bps.
  - Applying 49 bps differential to high EU group base implies jurisdictional commitment shielded roughly $86 billion (49bps×$17.5 trillion base) in shareholder value at announcement.
- Timing across legislative milestones:
  - EU×AI interaction large and significant at the 1% level only at the April 2021 Commission proposal; indistinguishable from zero at subsequent milestones (AI growth standalone effect follows same pattern).

### Pre-trend and event dynamics
- Pre-trend test:
  - Symmetric 81-day window [−40, +40]; factor loadings estimated over [−291,−41] so pre-trend window [−40,−1] disjoint from CAR estimation.
  - Table 2 pre-event EU×AI coefficients (bps): all economically small and statistically indistinguishable from zero:
    - CAR[−40,−1]: −25.2 (SE 59.5; t −0.42; p 0.674; N 769)
    - CAR[−40,−21]: −10.9 (SE 41.8; t −0.26; p 0.795; N 769)
    - CAR[−20,−1]: −14.3 (SE 23.4; t −0.61; p 0.545; N 769)
    - CAR[−10,−1]: 4.3 (SE 18.6; t 0.23; p 0.818; N 769)
    - CAR[−5,−1]: 14.6 (SE 11.3; t 1.29; p 0.205; N 769)
  - Event/post-event coefficients:
    - CAR[0, +2]: 39.7 ∗∗∗ (SE 10.3; t 3.87; p < 0.001; N 769)
    - CAR[0, +10]: 86.7 ∗∗∗ (SE 32.2; t 2.70; p 0.010; N 769)
    - CAR[0, +40]: 174.0 ∗∗∗ (SE 54.4; t 3.20; p 0.003; N 769)
  - Cumulative coefficient path near zero in pre-event period, rises sharply after day 0 and stabilizes at approximately +100 bps by day +40.

### Randomization inference
- Placebo distribution:
  - 483 trading dates between December 2020 and December 2022 used as placebo event dates, excluding ±10-day buffer around April 21, 2021.
  - For each placebo date, recompute CARs using [−250,−11] estimation window relative to placebo and re-estimate interaction model.
- Observed April 2021 coefficient on CAR[0, +2] lies in far right tail of placebo distribution for EU×AI interaction.
- Placebo rejection:
  - CAR[0, +1] actual = 31.8; RI p = 0.012
  - CAR[0, +2] actual = 39.7; RI p = 0.017
  - CAR[−1, +1] actual = 43.4; RI p = 0.002
  - CAR[−2, +2] actual = 45.2; RI p = 0.000
- EU revenue share only effect falls well within its placebo distribution.

### AI Growth vs. AI Level (horse race)
- Tests:
  - (i) Replace AI hiring growth with AI hiring level in baseline specification.
  - (ii) Enter both measures simultaneously.
- Table 3 key coefficients (CAR in bps):
  - EU×AI Level:
    - CAR[0, +1]: −9.1 (SE 7.7)
    - CAR[0, +2]: 0.2 (SE 14.4)
    - CAR[−1, +1]: −14.3 (SE 16.2)
    - CAR[−2, +2]: 1.2 (SE 12.8)
  - EU×AI Growth (with level included):
    - CAR[0, +1]: 37.2 ∗∗ (SE 14.7)
    - CAR[0, +2]: 39.6 ∗∗ (SE 17.8)
    - CAR[−1, +1]: 52.0 ∗∗∗ (SE 18.1)
    - CAR[−2, +2]: 44.5 ∗ (SE 23.7)
  - AI Level (main effect):
    - CAR[0, +1]: 11.4 (SE 13.2)
    - CAR[0, +2]: 4.9 (SE 15.9)
    - CAR[−1, +1]: 8.6 (SE 16.3)
    - CAR[−2, +2]: 3.1 (SE 21.3)
  - AI Growth (main effect, with level included): similar negative pattern:
    - CAR[0, +1]: −25.3 ∗∗ (SE 11.4)
    - CAR[0, +2]: −18.2 (SE 13.8)
    - CAR[−1, +1]: −39.6 ∗∗∗ (SE 14.1)
    - CAR[−2, +2]: −33.2 ∗ (SE 18.4)
  - Joint F-tests (null β(EU×AI Level) + β(EU×AI Growth) = 0):
    - CAR[0, +1]: F = 7.20; p = 0.002
    - CAR[0, +2]: F = 9.58; p ≈ 0.000
    - CAR[−1, +1]: F = 20.60; p ≈ 0.000
    - CAR[−2, +2]: F = 14.84; p ≈ 0.000
  - Industry FE: Yes; Firm controls: Yes; N = 769; Adj. R^2 reported per column.
- Findings:
  - AI level does not exhibit mitigating effects when interacted with EU share.
  - EU×AI Growth coefficient remains essentially unchanged and significant when both interactions included.
  - EU×AI Level coefficient insignificant in every window; point estimate of +0 bps on CAR[0, +2].
  - Correlation between interactions ρ = 0.60, indicating not a multicollinearity artifact.
- Interpretation:
  - Effect operates through AI growth (change) rather than AI level (stock).
  - Regulatory costs bind on future AI footprint; firms scaling AI rapidly pre-event have growth options repriced.

### Other robustness checks
- Longer-horizon CAR windows: jurisdictional capital effect maintained through several weeks (Appendix B.3 and B.4).
- Alternative specifications and checks (none overturn inference):
  - Replace three-factor with five-factor model (Appendix B.5).
  - Replace EU revenue share with other pre-2021 years or include 2020.
  - Use firms with fiscal year ends in December (Appendix B.6).
  - Add controls for firm equity duration and sensitivity to general macro news (Appendix B.7).
  - Exclude potential tax-haven countries in EU revenue share:
    - Excluding Ireland, Luxembourg, and the Netherlands yields 41.0 bp (p-value = 0.082).
  - Excluding April 22 (US capital gains tax proposal) and controlling for equity duration, yield and VIX news: remaining days positive and significant.
  - Remove potential outliers: deleting five and ten largest DFBETA observations yields 37.2 and 34.5 bp (p-values reported in Appendix).
- Evidence consistent with jurisdictional reach: proposal announcement only affects firms with EU revenue (Appendix B.8).

### Competing interpretations and tests
- Alternative explanations to the jurisdictional-capital story:
  - Firm size or international diversification rather than jurisdictional capital.
  - Lobby effects: firms deeply rooted in the EU may have lobbied to shape the Act in their favor.
- Section 6 outlines tests:
  - 6.1 distinguishes jurisdictional commitment from firm scale.
  - 6.2 addresses regulatory capture by testing political-access channels using EU and member-state level lobby proxies.
  - 6.3 confirms effect specificity to EU jurisdiction rather than general international diversification.

_Italic source: Excerpt from "4.1 Estimation Method" and related sections in the provided IMF PDF chapter._

### 6.1  Jurisdictional Commitment, Not Scale

### 6.1  Jurisdictional Commitment, Not Scale

### Core finding
- Jurisdictional capital—measured as EU revenue share interacting with AI growth—predicts positive abnormal returns around the 21 April 2021 EU AI Act proposal, and this effect is not explained by various scale measures of firm size or by broad international exposure.

### Tests replacing EU revenue share with scale proxies (Table 4, Panel A & B)
- Panel A: Alternative scale measures of EU exposure × AI Growth
  - EU Share × AI Growth: 31.8
    ∗∗∗
    (9.0) for [0, +1]; 39.7
    ∗∗∗
    (10.3) for [0, +2]; 43.4
    ∗∗∗
    (7.2) for [−1, +1]; 45.2
    ∗∗∗
    (8.8) for [−2, +2]
  - EU USD × AI Growth: -0.9 (4.0); -2.8 (4.6); 8.3 (7.0); 5.4 (7.0)
  - log(EU USD) × AI Growth: 1.4 (8.2); -4.4 (10.0); 14.2 (10.3); 4.6 (11.7)
- Panel B: Horse race of EU × AI vs. firm size × AI
  - EU × AI Growth: 33.5
    ∗∗∗
    (9.4); 41.5
    ∗∗∗
    (10.7); 45.4
    ∗∗∗
    (7.4); 47.6
    ∗∗∗
    (9.2)
  - log(Mcap) × AI Growth: -12.8 (7.5); -13.4 (9.1); -14.5
    ∗
    (8.3); -17.6
    ∗
    (11.4)
- Interpretation:
  - Interaction between absolute EU revenue (USD) and AI growth is economically small and statistically insignificant across windows.
  - Firm size×AI interaction is small in magnitude and low in significance, while EU×AI remains significant and close to baseline estimates.
  - Conclusion: the jurisdictional capital effect requires the relative importance of the EU market to the firm’s total business (EU revenue share), not absolute scale.

### Scale decomposition across alternative firm-level proxies (Table 5)
- Dependent: coefficient on ·× AI Growth (bps), CAR[0, +2]
  - Column (1) Baseline: EU × AI = 39.7
    ∗∗∗
    (10.3)
  - Column (2) + log(Mcap): EU × AI = 41.5
    ∗∗∗
    (10.7); log(Mcap) = -13.4 (9.1)
  - Column (3) + PP&E/AT: EU × AI = 34.0
    ∗∗∗
    (9.5); PP&E/AT = -48.8
    ∗∗∗
    (14.6)
  - Column (4) + log(Employ): EU × AI = 42.1
    ∗∗∗
    (11.3); log(Employ) = -12.8
    ∗
    (7.1)
  - Column (5) + R&D Intensity: EU × AI = 39.7
    ∗∗∗
    (10.3); R&D Intensity = 1.5 (8.1)
  - Column (6) Saturated (all interactions): EU × AI = 35.4
    ∗∗∗
    (9.4); log(Mcap) = -10.4 (8.7); PP&E/AT = -47.5
    ∗∗∗
    (15.5); R&D Intensity = -4.8 (5.2)
- Interpretation:
  - General size (log market capitalization, log employment) carries a negative interaction with AI growth (around −13 bps).
  - Physical capital intensity (PP&E/AT) loads strongly negative (−48.8 bps; −47.5 bps when saturated), suggesting firms with AI embedded in physical capital face heavier compliance requirements.
  - R&D intensity is essentially null.
  - Across all specifications, the EU×AI effect remains significant at the 1% level.
  - Conclusion: the EU jurisdictional capital effect operates through a dimension different from absolute EU operations, general firm size, physical-capital intensity, and R&D intensity.

### GDPR vs. AI Act — why scale alone is insufficient
- GDPR compliance has significant firm-level fixed costs that can be amortized across large user bases; scale benefits are plausible.
- The AI Act’s requirements often apply at the system or use-case level, creating more variable compliance costs: more AI products can mean more regulated systems and greater compliance burden.
- The AI Act creates a novel and evolving compliance regime (the world’s first comprehensive AI regulatory framework), with practical details to be clarified over time by regulators, certification bodies, and standard setters.
- Stable EU presence likely reflects local legal capacity, regulatory experience, and familiarity with implementation—factors beyond simple scale that lower adaptation costs to a new regime.

### Political engagement and lobby effect (Section 6.2; Table 6)
- Concern tested: whether firms with deeper EU roots shaped the AI Act in their favor (political access) rather than jurisdictional capital driving returns.
- Political engagement indicators constructed (pre-event):
  - Commissioner meetings (any meeting during 2015–2019): 73 of 771 firms
  - Lobby registration (pre-event appearance in EU Transparency Register): 74 firms
  - AI White Paper consultation (submitted comments): 17 firms
  - Country-level lobby (pre-event presence in mandatory member-state channels: France, Ireland, Germany): 86 firms
  - EU-level trade-association membership (pre-event roster of DigitalEurope, DOT Europe, or ITI): 45 firms
  - Any political engagement = one if any condition holds
- Tests run: (i) add engagement as a control; (ii) horse race include Engagement×AI alongside EU×AI; (iii) triple interaction Engagement×EU×AI.
- Key numeric evidence (EU × AI Growth on CAR[0, +2])
  - Commissioner meetings: EU × AI Growth = 38.6
    ∗∗∗
    (9.8) control; 40.0
    ∗∗∗
    (9.9) horse race; 43.2
    ∗∗∗
    (10.7) triple
  - Transparency Register: EU × AI Growth = 39.4
    ∗∗∗
    (9.7); 39.5
    ∗∗∗
    (10.1); 39.5
    ∗∗∗
    (10.3)
  - AI White Paper consultation: EU × AI Growth = 39.6
    ∗∗∗
    (10.2); 39.6
    ∗∗∗
    (10.0); 42.1
    ∗∗∗
    (9.7)
  - Country-level lobby: EU × AI Growth = 42.2
    ∗∗∗
    (10.4); 41.7
    ∗∗∗
    (9.9); 41.4
    ∗∗∗
    (8.9)
  - EU-level trade association: EU × AI Growth = 39.7
    ∗∗∗
    (10.4); 39.9
    ∗∗∗
    (10.9); 41.1
    ∗∗∗
    (10.2)
  - Any political engagement: EU × AI Growth = 39.5
    ∗∗∗
    (10.2); 39.5
    ∗∗∗
    (10.1); 38.5
    ∗∗∗
    (10.7)
- Interpretation:
  - The EU×AI coefficient is virtually unchanged from its baseline value of +40 bps on CAR[0, +2] when political engagement indicators are added as controls, horse race, or triple interaction.
  - None of the political engagement variables (dummy, interaction, or triple interaction) show consistent significance.
  - Post-April-2021 lobbying outcomes: EU×AI does not predict additional lobbying (all effects statistically insignificant), while AI hiring growth and firm size strongly do.
  - Conclusion: political engagement does not substitute for EU presence in generating the jurisdictional capital effect.

### Foreign revenue share vs. EU revenue share (Section 6.3; Table 7)
- Test: add Foreign × AI Growth (non-US, non-EU revenue share) to baseline.
- Key coefficients on CAR[0, +2]:
  - EU × AI Growth: 29.6
    ∗∗∗
    (10.2) for [0, +1]; 34.0
    ∗∗∗
    (12.0) for [0, +2]; 44.6
    ∗∗∗
    (7.8) for [−1, +1]; 43.0
    ∗∗∗
    (7.9) for [−2, +2]
  - Foreign Income × AI: 1.8 (7.8); 0.3 (10.6); 4.8 (8.7); 2.1 (13.7)
  - EU Revenue Share (main effect): 5.2 (6.5); 16.4
    ∗
    (9.5); 19.0
    ∗∗
    (9.4); 42.8
    ∗∗∗
    (10.1)
  - Foreign Income (main effect): -12.6 (9.2); -10.7 (13.5); -20.3
    ∗
    (10.8); -17.7 (13.0)
  - AI Growth (main effect): -22.4
    ∗∗∗
    (6.4); -22.4
    ∗∗∗
    (7.9); -33.4
    ∗∗∗
    (8.9); -32.8
    ∗∗∗
    (8.6)
- Interpretation:
  - The EU×AI interaction retains its magnitude and significance across windows; the Foreign×AI interaction is economically small and statistically insignificant.
  - Conclusion: general foreign exposure does not generate the jurisdictional capital effect; the effect is EU-specific.

### Amplifiers of the jurisdictional capital effect (preview)
- The paper tests amplification by: high-risk AI capability buildup (Section 7.1), stability of EU market presence (Section 7.2), geographic concentration of EU operations (Section 7.3), and pre-event compliance hiring (Section 7.4).

### High-risk AI amplifies the effect (Section 7.1; Table 8)
- Construction:
  - High-risk-AI (HRAI) postings are identified by dictionary mapping of Lightcast skills and titles to Annex III scope and Annex II Section A safety components; HRAI_i is the 2015–2019 annualised change in firm i’s share of postings so classified.
- Table 8, Panel A (triple interaction, full sample; all continuous variables in-sample z-scores):
  - EU × AI × HRAI: 10.18
    ∗∗∗
    (2.95); 15.02
    ∗∗∗
    (2.81); 15.82
    ∗∗∗
    (3.08); 20.58
    ∗∗∗
    (3.33) for windows [0, +1], [0, +2], [−1, +1], [−2, +2], respectively.
  - EU × AI (two-way from same regression): 25.53
    ∗
    (1.83); 31.69
    ∗∗
    (2.03); 26.57
    ∗∗
    (2.08); 23.56 (1.63)
- Table 8, Panel B (binary subsample split by HRAI hiring growth):
  - EU × AI, HRAI > 0 (N = 202): 30.76
    ∗∗∗
    (4.07); 34.86
    ∗∗∗
    (3.46); 72.17
    ∗∗∗
    (6.95); 83.65
    ∗∗∗
    (4.47)
  - EU × AI, HRAI = 0 (N = 554): 27.94 (1.35); 27.83 (1.49); 12.55 (0.42); 1.52 (0.05)
- Interpretation:
  - Each additional standard deviation of HRAI adds approximately +15 bps to the baseline effect on CAR[0, +2].
  - The effect is large and statistically significant in the high-risk-AI group, economically negligible in the complement.
  - Conclusion: jurisdictional capital is most valuable where regulatory requirements are greatest—firms scaling into high-risk AI exposure.

### Stability of EU commitment matters (Section 7.2)
- Measure: coefficient of variation (CV) of EU revenue share over time (standard deviation divided by mean) as a proxy for stability of EU revenue commitment.
- Test: split sample at median CV and re-estimate baseline regression.
- Result: the coefficient of EU×AI for the low-CV sample is significantly positive and close to the full-sample estimate; the high-CV sample counterpart is insignificant across all four event windows.
- Interpretation:
  - Stable EU presence (low CV) signals sustained strategic commitment and accumulated jurisdictional capital; episodic or volatile engagement (high CV) does not confer the same effect.

*Italic: Source: excerpt from IMF working paper chapter titled "6.1  Jurisdictional Commitment, Not Scale" (wpiea2026180-source-pdf).*

### 7.3  Geographic Concentration of EU Revenue

### 7.3 Geographic Concentration of EU Revenue

### Mechanism: jurisdictional capital from geographic concentration
- Jurisdictional capital depends on regulatory knowledge and compliance infrastructure buildup at the member-state level and on tacit, country-specific enforcement knowledge (how rules are applied in practice and what regulators expect).
- Holding total EU revenue share fixed, jurisdictional capital should be larger when revenue is concentrated in a few member states than when dispersed.
- Example logic from the text: "10% revenue concentrated in Germany builds depth with German enforcement institutions (e.g., Datenschutzkonferenz, BfDI), whereas 10% spread thinly across many member states is unlikely to generate comparable depth anywhere; under an EU-wide regime like the AI Act, the former converts this depth into benefit."

### Empirical test: within-EU concentration measures
- Two concentration measures used:
  - Within-EU Herfindahl index (HHI) across member states (Hf = Σc wf,c^2 where wf,c = sf,c / Sf).
  - Top-1 member-state revenue share.
- Samples split at the median of each concentration measure; baseline regressions re-estimated separately.

### Key findings (Table 10)
- Panel A (HHI split, high-concentration half):
  - EU×AI coefficient ranges from +47 bps to +86 bps across four event windows, all significant at the 1% level.
  - Reported estimates (High-concentration): 47.0
    ∗∗∗, 63.9
    ∗∗∗, 75.1
    ∗∗∗, 86.0
    ∗∗∗ (with standard errors (13.9), (19.7), (12.3), (15.4); N382).
- Panel A (HHI split, low-concentration half):
  - Low-concentration estimates are economically and statistically muted on three of four windows: 21.0
    ∗∗, 11.4, 13.8, 5.9 (with standard errors (7.8), (11.5), (13.1), (17.1); N375).
- Panel B (Top-1 split) shows a similar pattern:
  - High-concentration (above median) EU×AI: 47.2
    ∗∗∗, 64.5
    ∗∗∗, 75.3
    ∗∗∗, 87.3
    ∗∗∗ (std. err. (13.3), (18.7), (12.1), (14.6); N381).
  - Low-concentration (below median): 20.9
    ∗∗∗, 9.9, 13.0, 3.4 (std. err. (7.6), (11.3), (13.5), (17.6); N376).
- Interpretation: depth of engagement (concentration) — not breadth — converts EU market participation into regulatory know-how; two firms with identical total EU revenue share can face entirely different event returns depending on within-EU concentration.

### Robustness to single-country artifacts
- Appendix E decomposition shows major economies (Germany, France, Italy, Spain) dominate both halves; high-HHI firms concentrate more sharply within that same set, so the effect is not a single-country artifact.

---

### 7.4 Prior Compliance Experience

### Measurement of compliance hiring growth
- Compliance-related postings flagged via text analysis over Lightcast skills and titles dictionaries using a curated screen of 77 patterns across 11 functional categories.
- For each firm-year compute share of postings flagged as compliance-related.
- Firm-level compliance hiring growth = four-year-annualized change in this share over 2015–2019.
- Two variants:
  - Broader IT/data compliance: includes US-, UK-, and EU-statute hiring patterns.
  - EU IT/data compliance: excludes US- and UK-statute matches to isolate GDPR-driven hiring.

### Subsample results (Table 11)
- EU×AI effect concentrates among compliance-scaling firms:
  - Panel A (Broader measure), top half: EU × AI Growth ≈ 40.9
    ∗∗∗, 49.5
    ∗∗∗, 59.0
    ∗∗∗, 57.5
    ∗∗∗ (N=376; std. err. (11.5), (12.6), (11.4), (9.4)).
  - Panel A, bottom half: EU × AI Growth 23.6
    ∗, 15.4, 24.6, 15.3 (N=379; std. err. (12.4), (16.1), (26.7), (26.1)).
  - Panel B (EU-only measure), top half: EU × AI Growth 44.6
    ∗∗∗, 50.4
    ∗∗∗, 63.4
    ∗∗∗, 60.0
    ∗∗∗ (N=378; std. err. (13.6), (13.9), (12.5), (9.7)).
  - Panel B, bottom half: EU × AI Growth 19.1, 14.4, 21.2, 10.6 (N=379; std. err. (12.3), (15.0), (24.5), (24.4)).
- Interpretation: jurisdictional capital is prominent in firms with active interactions with regulators through prior IT compliance activities; effect largely absent among low compliance-growth firms.

### Horse race: EU×AI vs. Compliance×AI (Table 12)
- Baseline (Panel A, Broader C):
  - Baseline EU × AI Growth: 31.8
    ∗∗∗, 39.7
    ∗∗∗, 43.4
    ∗∗∗, 45.2
    ∗∗∗ (std. err. (9.0), (10.3), (7.2), (8.8)).
- Replacing EU revenue share with C (compliance) yields near-zero/insignificant coefficients:
  - Replace EU: −0.8, −2.5, 6.0, −0.1 (std. err. (5.0), (7.7), (7.0), (8.7)).
- Horse race (both interactions included):
  - EU × AI Growth remains essentially unchanged (31.6
    ∗∗∗, 39.5
    ∗∗∗, 43.4
    ∗∗∗, 44.9
    ∗∗∗) and C × AI remains indistinguishable from zero (−0.4, −2.0, 6.1, −0.2).
- Panel B (EU-only compliance) shows similar pattern: EU × AI stable; C × AI ~ −2.1, −3.7, 4.0, −3.3 (replacement) and ~ −2.1, −3.5, 3.8, −3.7 (horse race).
- Interpretation: compliance hiring growth does not mechanically replace the EU revenue channel; jurisdictional capital becomes valuable only when paired with EU-embedded revenue exposure.

---

### 8 EU Jurisdictional Capital and Post-Event Within-Firm Market-Share Reallocation

### Sample restriction and key variable construction (HighEU∩HighAI)
- Keep only HighAI firms (AI hiring growth 2015–2019 above sample median) and within them HighEU firms (2015–2019 average EU revenue share above sample median).
- Intersection HighEU ∩ HighAI contains 156 firms.
- Define peer set Pi using leave-one-out within-industry (4-digit SIC down to 2-digit as needed) with at least five total industry peers; LowEUPeer_i = #{p∈Pi: p is LowEU and HighAI} / #{p∈Pi: p is HighAI}, defined for 634 of N = 771 firms with peer sets containing at least three HighAI firms.
- Reallocation outcome:
  - realloc_i,t = log(EU_Rev_i,t) − log(nonEU_Rev_i,t)
  - ∆realloc_i = realloc_i,t∈2021–2025 − realloc_i,t∈2015–2020
  - Equivalently realloc_i,t = logit(EU_Share_i,t).
- Median EU share for HighEU∩HighAI subsample: 13% (compared with full sample 8% per Table A3).

### Empirical specification
- Estimate ∆realloc_i = α + β· LowEUPeer_i + γ′ X_i + δs + ε_i with X_i the usual firm controls and δs FF48 industry fixed effects.

### Results (Table 13)
- LowEUPeer coefficient estimates:
  - ∆realloc: +0.428
    ∗∗ (t = 1.80; Wild-bs p = 0.047; Observations 156).
  - ∆ log(EU_Share): +0.386
    ∗ (t = 1.76; Wild-bs p = 0.057; Observations 156).
  - ∆ log(EU_Rev): +0.210 (t = 0.58; Wild-bs p = 0.586; Observations 156).
- Magnitude interpretation:
  - +0.43 log-points implies moving from median to zero LowEUPeer maps into about a 21% (= 0.428/2) larger expansion of EU revenue relative to non-EU revenue.
  - Absolute EU revenue share gain for the median HighEU∩HighAI firm (13.2% pre-event EU share): approximately 2.8 percentage points (= e^(0.386*0.5+ln(13.2%)) − 13.2%).

### Placebo tests (Table 14)
- Placebo: estimate LowEUPeer effect on pooled non HighEU∩HighAI sample (N=443):
  - Coefficient = −0.001 (t = 0.00; Wild-bs p = 0.996) — essentially zero and not significant.
- Interpretation: the within-firm EU reallocation effect requires both HighAI and HighEU; absent either, LowEUPeer has no effect.

### Interpretation
- Evidence consistent with EU jurisdictional capital translating into real-economy outcomes: HighAI & HighEU firms facing weaker EU-embedded peers reallocate within-firm revenue toward EU markets post-announcement.
- The relevant margin is within-firm EU-versus-non-EU reallocation rather than absolute EU revenue growth.

---

### 9 Implications and 10 Conclusion

### Implications
- For investors: markets price jurisdictional capital as an intangible asset; firms’ ability to operate within a regulatory jurisdiction and adapt to rule changes is valued. Around the April 21, 2021 Commission proposal, the EU×AI interaction produced approximately +40 bps announcement return.
- For regulators: results highlight a design margin—streamlined requirements could make compliance capability more broadly accessible. The text notes recent updates (June 2026) include measures to simplify and delay certain regulatory requirements, which may have evolved market reactions.
- For firms: longstanding presence in a regulated jurisdiction can become valuable when new rules arrive; benefits can accrue to smaller firms that invested in durable operating depth in the jurisdiction.

### Conclusion
- Using predetermined exposure measures across about 800 US-listed firms around the April 21, 2021 Commission proposal, firms with established EU market presence captured a positive announcement return on the EU×AI interaction of approximately +40 bps.
- Three lines of evidence support interpretation as priced jurisdictional capital:
  - Effect operates through EU revenue share rather than firm size or international diversification.
  - Political access measures do not absorb the effect; lobbying unlikely to drive results.
  - Effect is amplified by high-risk AI capacity buildup, stable and geographically concentrated EU presence, and prior GDPR-era compliance experience.
- Suggestive evidence that jurisdictional capital affects real activity: HighAI & HighEU firms facing LowEU peers subsequently increase EU revenue share, consistent with jurisdictional capital shaping within-firm market allocation.

*Source: wpiea2026180-source-pdf - 7.3 Geographic Concentration of EU Revenue*

### References

### References and Appendices

### References
- Full bibliographic list of cited works includes:
  - Babina, Tania; Anastassia Fedyk; Alex He; James Hodson. 2024. “Artificial Intelligence, Firm Growth, and Product Innovation.” Journal of Financial Economics, 151: 103745. (cite on p. 4).
  - Bartov, Eli; Dan Givoly; Carla Hayn. 2002. “The rewards to meeting or beating earnings expectations.” Journal of Accounting and Economics, 33(2): 173–204. (cite on p. 51).
  - Bernard, Victor L; Jacob K Thomas. 1989. “Post-earnings-announcement drift: delayed price response or risk premium?” Journal of Accounting Research, 27: 1–36. (cite on p. 51).
  - Biguri, Kizkitza; Jörg R. Stahl. 2025. “Who Pays a Visit to Brussels? Firm Value Effects of Cross-Border Political Access to European Commissioners.” Journal of Financial and Quantitative Analysis, 60(2): 948–973. (cite on p. 39).
  - Canayaz, Mehmet; Zhe Wang. 2026. “Crafting an AI Compass: The Influence of Global AI Standards on Firms.” SSRN Working Paper. SSRN 4481608. (cite on p. 4).
  - Christensen, Hans B.; Luzi Hail; Christian Leuz. 2016. “Capital-Market Effects of Securities Regulation: Prior Conditions, Implementation, and Enforcement.” Review of Financial Studies, 29(11): 2885–2924. (cite on p. 3).
  - Comunale, Mariarosaria; Andrea Manera. 2024. “The Economic Impacts and the Regulation of AI: A Review of the Academic Literature and Policy Actions.” IMF Working Papers, 2024(065): 1. (cite on p. 6).
  - Coppola, Antonio; Matteo Maggiori; Brent Neiman; Jesse Schreger. 2021. “Redrawing the Map of Global Capital Flows: The Role of Cross-Border Financing and Tax Havens.” Quarterly Journal of Economics, 136(3): 1499–1556. (cite on p. 42).
  - Dechow, Patricia M; Richard G Sloan; Mark T Soliman. 2004. “Implied equity duration: A new measure of equity risk.” Review of Accounting Studies, 9(2): 197–228. (cite on p. 57).
  - Eisfeldt, Andrea L.; Gregor Schubert; Miao Ben Zhang. 2023. “Generative AI and Firm Values.” NBER Working Paper 31222. (cite on p. 4).
  - Faccio, Mara. 2006. “Politically Connected Firms.” American Economic Review, 96(1): 369–386. (cite on p. 3).
  - Fama, Eugene F.; Kenneth R. French. 1993. “Common Risk Factors in the Returns on Stocks and Bonds.” Journal of Financial Economics, 33(1): 3–56. (cite on p. 10).
  - Fama, Eugene F.; Kenneth R. French. 2015. “A Five-Factor Asset Pricing Model.” Journal of Financial Economics, 116(1): 1–22. (cite on p. 55).
  - Fisman, Raymond. 2001. “Estimating the Value of Political Connections.” American Economic Review, 91(4): 1095–1102. (cite on p. 3).
  - Goldberg, Samuel G.; Garrett A. Johnson; Scott K. Shriver. 2024. “Regulating Privacy Online: An Economic Evaluation of the GDPR.” American Economic Journal: Economic Policy, 16(1): 325–358. (cite on p. 3).
  - Johnson, Garrett A.; Scott K. Shriver; Samuel G. Goldberg. 2023. “Privacy and Market Concentration: Intended and Unintended Consequences of the GDPR.” Management Science, 69(10): 5695–5721. (cite on p. 3).
  - Klapper, Leora; Luc Laeven; Raghuram Rajan. 2006. “Entry Regulation as a Barrier to Entrepreneurship.” Journal of Financial Economics, 82(3): 591–629. (cite on p. 3).
  - Klein, Daniel; Christopher A. Ludwig; Christoph Spengel. 2022. “Taxing the Digital Economy: Investor Reaction to the European Commission’s Digital Tax Proposals.” National Tax Journal, 75(1): 61–92. (cite on p. 4).
  - Livnat, Joshua; Richard R Mendenhall. 2006. “Comparing the post-earnings announcement drift for surprises.” Journal of Accounting Research, 44(1): 177–205. (cite on p. 51).
  - Peukert, Christian; Stefan Bechtold; Michail Batikas; Tobias Kretschmer. 2022. “Regulatory Spillovers and Data Governance: Evidence from the GDPR.” Marketing Science, 41(4): 746–768. (cite on p. 3).
  - Smuha, Nathalie A. 2021. “From a ‘Race to AI’ to a ‘Race to AI Regulation’: Regulatory Competition for Artificial Intelligence.” Law, Innovation and Technology, 13(1): 57–84. (cite on p. 66).
  - Stigler, George J. 1971. “The Theory of Economic Regulation.” Bell Journal of Economics and Management Science, 2(1): 3–21. (cite on p. 3).
  - Voigt, Paul; Axel von dem Bussche. 2017. The EU General Data Protection Regulation (GDPR): A Practical Guide. Springer. (cite on p. 66).
  - Zeume, Stefan. 2017. “Bribes and Firm Value.” Review of Financial Studies, 30(5): 1457–1489. (cite on p. 4).

### A. Data Appendix — Data sources, variable construction, and sample procedures
A.1 Data Sources
- Job postings
  - Primary AI exposure measure constructed from Lightcast database covering 8,926 US-listed firms from 2010 to 2026.
  - For each firm-month: count total job postings and AI-related job postings (skills requiring AI ethics and governance postings, machine learning, natural language processing, and computer vision).
  - Firms identified by Compustat gvkey for linkage to financial data.
- Geographic revenue
  - EU revenue exposure from FactSet Geographic Revenue (GeoRev) for 10,382 US-listed firms.
  - Primary exposure: mean EU revenue share over FY2015–2019 (pre-COVID window).
  - Firms with disclosure in k ≤ 5 years enter with mean of their k disclosed annual shares.
- Stock returns
  - Daily stock prices and adjustment factors from CRSP/Compustat Merged (CCM).
  - Daily total returns computed as:
    r_it = prccd_it × trfd_it / ajexdi_it  /  prccd_i,t−1 × trfd_i,t−1 / ajexdi_i,t−1 − 1,
    where trfd is the total return factor for dividends.
  - For ~2,200 non-dividend-paying stocks where CRSP does not compute trfd, set trfd = 1.
  - Retain only primary securities (LINKPRIM ∈ {'P', 'C'}, LINKTYPE starting with 'L', iid = 1).
  - All 771 firms have continuous CRSP coverage through June 17, 2021 (day +40 relative to April 21, 2021); no silent truncation from CAR calculations due to mid-window delisting.
- Fama–French factors
  - Daily Fama–French three-factor and five-factor returns from Kenneth French’s data library.
- Compustat Annual
  - Firm-level financial controls drawn from Compustat Annual, averaged over FY2015–2019, applying standard filters (indfmt = 'INDL', datafmt = 'STD', consol = 'C').
- Foreign-income diversification control
  - Foreign income share (pifo/pi) constructed from Compustat; ratio computed only for firm-years with |pi| > 0.01.
- Political access
  - Meetings with European Commissioners from Integrity Watch platform: 78,029 total meetings, 25,258 involving companies; matched 73 sample firms to Commissioner meetings during 2015–2019.
  - EU Transparency Register: 17,061 organizations; matched 74 sample firms with registration dates before April 21, 2021.
  - AI White Paper consultation indicator: 17 sample firms matched to submissions among 1,216 submissions to the Commission’s “Have Your Say” portal (consultation ran Feb 19 to Jun 14, 2020; submissions: 150 companies, 117 NGOs, 112 academic institutions, 111 business associations, remainder individuals/other).
  - Member-state-level lobby registers checked for France, Ireland, and Germany (86 firms); collectively contain ≤ 7 firms with top-1 EU revenue in sample.
  - EU-level trade-association membership: indicator if firm appears on pre-event member rosters of DigitalEurope, DOT Europe, or ITI (45 firms); archived web snapshots from 2020-12-04, 2021-01-24, and 2020-04-24 used.
  - For member-state-level lobby and EU-level trade-association membership, include membership of any EU subsidiaries.

A.2 Financial Controls
- Controls from Compustat, averaged FY2015–2019:
  - log market capitalization: log(csho × prcc_f) (not winsorized).
  - book-to-market ratio: ceq/(csho × prcc_f).
  - leverage: (dltt + dlc)/at.
  - R&D intensity: xrd/sale.
  - profitability: oiadp/at.
- Winsorization and missing-value treatment:
  - Book-to-market, leverage, R&D intensity, and profitability winsorized at 1st and 99th percentiles within analysis sample.
  - Missing values of dltt, dlc, and xrd treated as zero.
  - Firm-years with non-positive sales set to missing for R&D intensity.
  - Strictly positive market capitalization required.
- FF48 industry fixed effects:
  - Firms classified by SIC code from most recent Compustat record; firms without valid match assigned to “Other” (FF48 industry 48; 7 firms, 0.9% of sample).

A.3 Sample-Construction Waterfall
- Crosswalk and matching:
  - Matched FactSet GeoRev firms to CRSP/Compustat using CUSIP and ISIN identifiers in a waterfall merge (ISIN-based CUSIP6, then FactSet CUSIP6): 5,224 of 10,382 FactSet firms match; 5,158 unmatched predominantly micro-cap and OTC-listed outside CRSP/Compustat.
  - Verified unmatched firms are genuinely absent from CRSP/Compustat rather than lost to identifier mismatches.
- Filters to form main analysis sample:
  - Require non-missing EU revenue data for ≥ 1 year in FY2015–2019: 1,739 matched firms.
  - From job postings, retain firms with at least 5 total postings in both CY2015 and CY2019 (defines AI hiring growth); AI share computed annually by summing AI and total postings within each year.
  - Require valid Fama–French three-factor CARs around April 21, 2021 (estimation window [−250,−11] with ≥ 120 trading days), non-missing Compustat controls averaged FY2015–2019, and pre-event share price (end of March 2021) ≥ $5.
  - Resulting analysis sample: 771 firms.
- Extensive-margin sample (for jurisdictional reach checks):
  - Include firms with no EU revenue disclosure; sample size increases to 1,264 firms.
  - EU presence indicator defined as dummy for firm ever having non-missing values of EU revenue share in FactSet GeoRev between 2015–2019.
  - The 493 additional firms satisfy other filters but have no EU revenue disclosed over pre-event window.

A.4 Summary Statistics — Main Sample (N = 771) (exact values preserved)
- Panel A: Key Exposure Variables
  - AI_Growth (ppts/yr): N = 771; Mean = 1.432; SD = 2.819; P25 = 0.000; Median = 0.532; P75 = 2.202; P99 = 12.035.
  - EU_Share: N = 771; Mean = 0.098; SD = 0.089; P25 = 0.034; Median = 0.081; P75 = 0.135; P99 = 0.442.
- Panel B: Cumulative Abnormal Returns (April 2021, bps)
  - CAR[0,+1]: N = 771; Mean = 16.301; SD = 259.957; P25 = -102.854; Median = 11.428; P75 = 130.436; P99 = 668.342.
  - CAR[0,+2]: N = 771; Mean = -18.861; SD = 326.153; P25 = -164.386; Median = -11.787; P75 = 130.964; P99 = 875.575.
  - CAR[-1,+1]: N = 771; Mean = 56.369; SD = 326.296; P25 = -107.603; Median = 72.002; P75 = 216.598; P99 = 949.879.
  - CAR[-2,+2]: N = 771; Mean = 10.668; SD = 432.706; P25 = -186.424; Median = 37.142; P75 = 208.123; P99 = 943.127.
- Panel C: Financial Controls
  - Market Cap (USD bn): N = 771; Mean = 21.368; SD = 57.590; P25 = 1.391; Median = 4.515; P75 = 14.069; P99 = 249.742.
  - Book-to-Market: N = 771; Mean = 0.384; SD = 0.303; P25 = 0.169; Median = 0.309; P75 = 0.535; P99 = 1.342.
  - Leverage: N = 771; Mean = 0.259; SD = 0.183; P25 = 0.111; Median = 0.248; P75 = 0.365; P99 = 0.841.
  - R&D Intensity: N = 771; Mean = 0.093; SD = 0.205; P25 = 0.000; Median = 0.020; P75 = 0.102; P99 = 1.405.
  - Profitability: N = 771; Mean = 0.063; SD = 0.120; P25 = 0.032; Median = 0.076; P75 = 0.119; P99 = 0.307.
- Panel D: Diversification
  - Foreign Income Share: N = 683; Mean = 0.305; SD = 2.351; P25 = 0.054; Median = 0.309; P75 = 0.698; P99 = 3.938.
- Notes
  - AI_Growth defined as annualized change in AI share of job postings CY2015 to CY2019 (percentage points per year).
  - EU_Share is firm mean FactSet GeoRev EU revenue share over FY2015–2019.
  - Market Cap (USD bn) is antilog of log market-cap control; regressions use log(market cap).
  - Book-to-market, leverage, R&D intensity, and profitability reported after 1%/99% within-sample winsorization.

- Extensive-margin sample (N = 1,264) — selected exact statistics:
  - AI_Growth (ppts/yr): N = 1,264; Mean = 1.145; SD = 2.542; P25 = 0.000; Median = 0.317; P75 = 1.534; P99 = 11.184.
  - EU Share (0 if non-discloser): N = 1,264; Mean = 0.060; SD = 0.084; P25 = 0.000; Median = 0.024; P75 = 0.098; P99 = 0.383.
  - EU Indicator (1=disclosed): N = 1,264; Mean = 0.610; SD = 0.488; P25 = 0.000; Median = 1.000; P75 = 1.000; P99 = 1.000.
  - CAR[0,+1]: N = 1,264; Mean = 13.961; SD = 260.543; P25 = -108.495; Median = 10.930; P75 = 129.627; P99 = 707.486.
  - Market Cap (USD bn): N = 1,264; Mean = 16.331; SD = 47.838; P25 = 1.073; Median = 3.130; P75 = 11.078; P99 = 222.056.
  - Foreign Income Share: N = 814; Mean = 0.281; SD = 2.169; P25 = 0.011; Median = 0.222; P75 = 0.663; P99 = 3.785.
  - Notes: EU Share set to zero for 493 non-disclosers; EU Indicator equals one if firm discloses positive EU segment revenue in at least one fiscal year during 2015–2019. CARs are market-model cumulative abnormal returns in basis points; day 0 = April 21, 2021.

### B. Robustness Appendix — Additional tests and auxiliary results
B.1 Pre-Trend Trajectory: Cumulative Coefficient
- Figure A.1 plots cumulative trajectory of EU × AI coefficient from day −40 through day +40.
- Notes: Each point plots coefficient on EU×AI (standardized) from regression of CAR[−40,t] based on Equation 1, t ranges day −40 to +40. Cumulation anchor fixed at day −40. CAR estimation window [−291,−41] is disjoint from event window.

B.2 Randomization Inference: Summary of reported robustness tables and timing tests
- Table A5: Interaction Effect Across Legislative Events (Dependent variable: CAR (bps), event window [0, +2])
  - Apr 2021 Proposal: EU × AI Growth = 39.7 ∗∗∗ (standard error 10.3).
  - Jun 2023 EP Vote: EU × AI Growth = 7.5 (standard error 14.4).
  - Dec 2023 Trilogue: EU × AI Growth = 6.8 (standard error 20.4).
  - Mar 2024 EP Final: EU × AI Growth = -16.0 (standard error 13.3).
  - EU Revenue Share coefficients (Apr 2021 to Mar 2024): 2.4 (13.9), -7.9 (16.8), -18.4 (13.6), -0.9 (10.1).
  - AI Growth coefficients (Apr 2021 to Mar 2024): -16.7 ∗∗ (6.4), 24.2 (16.8), -5.1 (10.2), -3.5 (13.9).
  - Industry FE (FF48): Yes for all events. Firm controls: Yes for all events.
  - Observations: Apr 2021 Proposal: 769; Jun 2023 EP Vote: 767; Dec 2023 Trilogue: 766; Mar 2024 EP Final: 765.
  - Notes: Each event is a separate FF3 market-model re-estimation with pre-event [−250,−11] window. Standard errors FF48-clustered. ∗, ∗∗, ∗∗∗ denote 10%, 5%, 1% significance.
- Table A6: Treatment Timing and Information Leakage (Dependent variable: CAR (bps))
  - Columns and windows:
    - [−5,−3]: EU × AI Growth = 9.1 (8.6).
    - [−5,−1]: EU × AI Growth = 14.6 (11.3).
    - [0, +2]: EU × AI Growth = 39.7 ∗∗∗ (10.3).
    - [−5, +2]: EU × AI Growth = 54.3 ∗∗∗ (13.5).
  - EU Revenue Share in same columns: -3.1 (7.3), 17.0 (10.7), 2.4 (13.9), 19.4 ∗ (11.0).
  - AI Growth coefficients: -8.3 (8.8), -23.8 ∗ (12.9), -16.7 ∗∗ (6.4), -40.5 ∗∗∗ (12.0).
  - Industry FE (FF48): Yes for all columns. Firm controls: Yes for all columns.
  - Observations: 769 in each column.
  - RI p-value, EU×AI t for column (4): 0.002 (two-sided randomization-inference p-value from placebo dates outside ±10-trading-day buffer around April 21, 2021).
  - Notes: CARs are Fama–French three-factor abnormal returns estimated over trading days [−250,−11] relative to April 21. Standard errors clustered by FF48 industry. ∗, ∗∗, ∗∗∗ denote 10%, 5%, and 1% levels.
- Table A7: Randomization inference per-window results reported (coefficient-based RI p-values for EU×AI interaction noted to range within reported procedure).

*Italic: Content summarized directly from the PDF section titled "References" and Appendices A and B.*

### 0.039 to 0.062 across the four windows. The AI Growth main effect yields t-based RI p-values

### wpiea2026180-source-pdf - 0.039 to 0.062 across the four windows. The AI Growth main effect yields t-based RI p-values

### Randomization Inference: April 21, 2021 coefficients vs. placebo distribution
- Placebo distribution constructed from 483 non-event trading dates (Dec 2020 to Dec 2022, excluding ±10 trading days around April 21, 2021).
- EU × AI Growth (actual coefficients and inference):
  - CAR[0, +1]: Actual coeff. +31.8 bps; Actual t-statistic +3.54; Placebo mean +0.0 bps; Placebo SD 15.2 bps; Placebo [5, 95] [−23.6, +23.2]; RI p-value (coeff.) 0.039; RI p-value (t-stat) 0.012.
  - CAR[0, +2]: Actual coeff. +39.7 bps; Actual t-statistic +3.87; Placebo mean −0.0 bps; Placebo SD 18.5 bps; Placebo [5, 95] [−28.5, +28.4]; RI p-value (coeff.) 0.041; RI p-value (t-stat) 0.017.
  - CAR[−1, +1]: Actual coeff. +43.4 bps; Actual t-statistic +6.06; Placebo mean +0.0 bps; Placebo SD 18.5 bps; Placebo [5, 95] [−28.5, +28.4]; RI p-value (coeff.) 0.031; RI p-value (t-stat) 0.002.
  - CAR[−2, +2]: Actual coeff. +45.2 bps; Actual t-statistic +5.15; Placebo mean +0.1 bps; Placebo SD 24.3 bps; Placebo [5, 95] [−37.8, +38.7]; RI p-value (coeff.) 0.062; RI p-value (t-stat) 0.000.
- EU Revenue Share (actual and placebo):
  - CAR[0, +1]: Actual coeff. +3.2 bps; Actual t-statistic +0.35; Placebo mean −0.1 bps; Placebo SD 18.5 bps; Placebo [5, 95] [−26.5, +28.0]; RI p-value (coeff.) 0.834; RI p-value (t-stat) 0.791.
  - CAR[0, +2]: Actual coeff. +2.4 bps; Actual t-statistic +0.18; Placebo mean −0.2 bps; Placebo SD 22.6 bps; Placebo [5, 95] [−32.5, +32.5]; RI p-value (coeff.) 0.882; RI p-value (t-stat) 0.886.
  - CAR[−1, +1]: Actual coeff. +15.1 bps; Actual t-statistic +1.78; Placebo mean −0.2 bps; Placebo SD 22.6 bps; Placebo [5, 95] [−32.6, +32.6]; RI p-value (coeff.) 0.447; RI p-value (t-stat) 0.197.
  - CAR[−2, +2]: Actual coeff. +22.5 bps; Actual t-statistic +2.03; Placebo mean −0.4 bps; Placebo SD 29.9 bps; Placebo [5, 95] [−43.7, +47.2]; RI p-value (coeff.) 0.364; RI p-value (t-stat) 0.135.
- AI Growth (actual and placebo):
  - CAR[0, +1]: Actual coeff. −21.9 bps; Actual t-statistic −3.66; Placebo mean −0.1 bps; Placebo SD 12.6 bps; Placebo [5, 95] [−19.6, +22.2]; RI p-value (coeff.) 0.091; RI p-value (t-stat) 0.002.
  - CAR[0, +2]: Actual coeff. −16.7 bps; Actual t-statistic −2.60; Placebo mean −0.1 bps; Placebo SD 15.2 bps; Placebo [5, 95] [−24.6, +25.9]; RI p-value (coeff.) 0.267; RI p-value (t-stat) 0.035.
  - CAR[−1, +1]: Actual coeff. −37.1 bps; Actual t-statistic −4.65; Placebo mean −0.2 bps; Placebo SD 15.2 bps; Placebo [5, 95] [−24.4, +25.8]; RI p-value (coeff.) 0.014; RI p-value (t-stat) 0.000.
  - CAR[−2, +2]: Actual coeff. −32.2 bps; Actual t-statistic −4.25; Placebo mean −0.2 bps; Placebo SD 20.2 bps; Placebo [5, 95] [−31.2, +35.5]; RI p-value (coeff.) 0.116; RI p-value (t-stat) 0.002.
- Placebo means for all three coefficients are centered near zero with no systematic bias. Figures A.2 and A.3 show full 2×2 placebo distributions.

### Long-run Effect: persistence across horizons (Table A8 and local projections)
- EU × AI Growth coefficients by horizon (OLS of CAR):
  - CAR[0, +1]: 31.8 bps, significant at 1% (standard error 9.0).
  - CAR[0, +2]: 39.7 bps, significant at 1% (standard error 10.3).
  - CAR[0, +5]: 43.3 bps, significant at 1% (standard error 10.5).
  - CAR[0, +10]: 86.7 bps, significant at 1% (standard error 32.2).
  - CAR[0, +20]: 113.0 bps, significant at 1% (standard error 35.1).
  - CAR[0, +30]: 126.3 bps, significant at 1% (standard error 33.3).
  - CAR[0, +40]: 174.0 bps, significant at 1% (standard error 54.4).
- AI Growth standalone coefficients by horizon (selected):
  - CAR[0, +1]: −21.9 bps, significant at 1% (standard error 6.0).
  - CAR[0, +2]: −16.7 bps, significant at 5% (standard error 6.4).
  - CAR[0, +40]: −1.6 bps (standard error 49.5).
- Observations: 769 for each horizon; Industry FE (FF48) and firm controls included.
- Interpretation: EU×AI interaction grows monotonically from ~+30 bps on CAR[0, +1] to +174 bps on CAR[0, +40]; significant at the 1% level at every horizon; absence of reversal suggests permanent repricing of jurisdictional capital.
- Local projection regressions (Equation 3) for CAR[+3, +h] on CAR[0, +2]:
  - θ_h estimates:
    - h = 5: θ = −0.0191 (SE 0.0561).
    - h = 10: θ = −0.1553, significant at 5% (SE 0.0753).
    - h = 20: θ = −0.1811, significant at 10% (SE 0.1054).
    - h = 30: θ = −0.0762 (SE 0.1391).
    - h = 40: θ = +0.1049 (SE 0.1592).
  - N = 769 for each column; Adj R^2 values: 0.111, 0.099, 0.055, 0.055, 0.065.
  - Interpretation: θ_h is statistically indistinguishable from zero at h = 5, mildly negative at h = 10 and h = 20 (temporary moderate reversal < 20%), and noisy thereafter; overall no long-run reversal predicting continued positive post-event returns for EU×AI-treated firms.

### Long-run Effect: Earnings-announcement contamination robustness
- Potential contamination channels: (i) in-window reporting D_t; (ii) surprise content SUE_z and |SUE|_z and beat indicator; (iii) post-earnings announcement drift (PEAD) measured via days_post and interactions.
- Correlation diagnostics (Table A10): Pearson correlations of earnings-control proxies with EU, AI, EU×AI at three horizons:
  - Largest correlation with EU×AI reported is +0.114 (days_post at [0, +2]).
  - At [0, +40], 5-day FY1 revision correlation with EU: −0.260*** and with EU×AI: +0.168***.
- Regression robustness ladder (Table A11): eight nested specifications (A–H) adding earnings controls and PEAD controls.
  - EU×AI coefficient at headline windows (ranges across specs A–G):
    - [0, +2]: range [+37, +40] bps; all seven restricted coefficients significant at 1%.
    - [0, +1]: range [+30, +34] bps.
    - [0, 0]: range [+11, +13] bps.
  - Spec (H) (unrestricted D_t interactions, identifying EU×AI from out-of-window firms) yields coefficients:
    - [0, 0]: +10.46 bps (significant at 10%).
    - [0, +1]: +21.82 bps (significant at 5%).
    - [0, +2]: +32.13 bps (significant at 1%).
  - Homogeneity F-test (Spec H):
    - [0, 0]: F = 1.07 (p = 0.37).
    - [0, +1]: F = 3.15 (p = 0.03) — marginal rejection driven by D_t × AI rather than D_t × EU×AI.
    - [0, +2]: F = 1.35 (p = 0.27).
- Falsification drop-test:
  - Excluding 73 firms that reported during April 19–23 yields 698-firm subsample; headline [0, +2] coefficient = +34 bps (t = 3.2), a 6-bp reduction, still significant at 1%.
  - Mean SUE: jurisdictional capital (High-EU×High-AI) cell = +0.20% of pre-event price vs Low-EU×Low-AI control = +0.33%.
- Bounding residual bias:
  - PEAD literature typical magnitudes ~30–60 bps over 60 trading days for a 1-SD surprise.
  - Observed |r| ≤ 0.11 correlation between PEAD-time and EU×AI implies an upper bound on PEAD-driven bias to long-horizon coefficient of roughly 20–40 bps.
  - Observed gap between baseline and full-PEAD-corrected coefficients at [0, +40] is 9 bps, within the bound.

### Fama–French Five-Factor robustness (Table A12)
- Abnormal returns recomputed using Fama–French five-factor model; full interaction model repeated.
- Key FF5 results (CAR in bps, standard errors in parentheses):
  - CAR[0, +1]: EU × AI Growth 32.2*** (8.6); EU Revenue Share 1.9 (9.4); AI Growth −20.7*** (6.0).
  - CAR[0, +2]: EU × AI Growth 40.8*** (10.5); EU Revenue Share 1.3 (14.2); AI Growth −17.2*** (6.3).
  - CAR[−1, +1]: EU × AI Growth 42.6*** (6.9); EU Revenue Share 14.1* (8.1); AI Growth −34.5*** (7.3).
  - CAR[−2, +2]: EU × AI Growth 44.7*** (8.5); EU Revenue Share 21.8* (11.1); AI Growth −30.3*** (6.7).
- Observations: 769 for each window; Adjusted R^2: 0.020, 0.091, 0.049, 0.080.
- Conclusion: EU×AI coefficient differs by at most 1.5 bps between FF3 and FF5 on any window; findings robust to including RMW and CMA factors.

### Historical EU Revenue Share robustness (Table A13)
- Replaces EU revenue-share measure with alternative historical EU-exposure measures; re-estimates CAR[0, +2] specification.
- Panel A: Average EU revenue shares over different historical windows (coefficients for EU×AI interaction, CAR[0, +2]):
  - Avg 2015–2019 (baseline): 39.7*** (10.3), N 769.
  - Avg 2014–2018: 40.7*** (10.3), N 758.
  - Avg 2012–2016: 32.4*** (8.0), N 711.
  - Avg 2010–2014: 14.7** (6.9), N 666.
  - Avg 2008–2012: 13.8* (8.0), N 620.
- Panel B: Fiscal-year single-year measures (selected):
  - FY2004: 11.2*** (3.7), N 430.
  - FY2007: 1.9 (5.3), N 485.
  - FY2012: 17.2** (8.1), N 606.
  - FY2015: 35.1** (13.0), N 674.
- Industry FE (FF48) and firm controls included in all regressions.
- Conclusion: Interaction effect persists across alternative historical EU-exposure definitions.

### Equity duration and sensitivity to general news (Table A14)
- Constructed implied equity-duration measure per Dechow, Sloan and Soliman (2004) forecasting cash flows 15 years with 12% discount rate.
- Estimated firm betas to daily changes in 10-year Treasury yield and VIX over 500 pre-event trading days ending 30 trading days before the proposal.
- CAR[0, +2] FF5 baseline and added controls (coefficients in bps; standard errors in parentheses; wild-bootstrap p-values reported):
  - FF5 base: EU × AI Growth +40.8* (10.8); Wild-bootstrap p 0.086.
  - Common base (complete-case): +33.3* (9.4); Wild-bootstrap p 0.067.
  - + Duration: +33.3* (9.4); Wild-bootstrap p 0.067.
  - + Risk betas: +33.6* (9.4); Wild-bootstrap p 0.064.
  - + All (Duration and Risk betas): +33.6* (9.4); Wild-bootstrap p 0.065.
  - Risk-adjusted CAR (abnormal returns additionally purged of pre-event yield/VIX component): +30.0* (8.7); Wild-bootstrap p 0.062.
- AI Growth coefficients and wild-bootstrap p-values reported; observations range 769 to 722 depending on sample.
- Conclusion: Adding duration and pre-event risk-betas produces essentially no attenuation of EU×AI effect; EU×AI remains positive and significant.

### Jurisdiction reach of the Act (Table A15)
- Firms split by EU revenue disclosure indicator (EU_Ind = 1 disclosers; EU_Ind = 0 non-disclosers).
- Panel A: Firm characteristic differences (mean values and difference p-values):
  - log(Market cap): EU_Ind = 0: 7.669; EU_Ind = 1: 8.421; Difference 0.751*** (test statistic −7.40).
  - AI hiring growth (ppts/yr): 0.007 vs 0.014; Difference 0.007*** (−5.08).
  - Foreign income share: 0.159 vs 0.305; Difference 0.145 (−0.70).
  - Book-to-market: 0.497 vs 0.379; Difference −0.117*** (5.70).
  - R&D intensity: 1.424 vs 0.600; Difference −0.824 (1.03).
  - Profitability: 0.030 vs 0.059; Difference 0.028*** (−3.17).
  - Tech industry (FF48 business/computers): 0.073 vs 0.331; Difference 0.258*** (−11.11).
- Panel B: AI Growth effect within each EU_Indicator arm (AI Growth coefficient on CAR):
  - EU_Ind = 0 (non-disclosers), CAR windows: coefficients 9.85 (23.0), 1.78 (53.4), 8.9 (24.9), 30.4 (56.8); N 488.
  - EU_Ind = 1 (disclosers), CAR windows: −20.4*** (6.6), −16.4** (6.7), −34.5*** (9.1), −30.8*** (8.9); N 769.
- Interpretation: AI standalone negative effect appears among firms with EU revenue disclosure and disappears for firms without material EU revenue; supports EU revenue share as measure of jurisdictional reach.

### Post-Proposal AI-Act lobbying (Table A16)
- Constructed four post-2021 firm-level lobbying indicators (all dated after April 2021):
  - EU-level firm lobbying: 74 firms flagged if engaged EU institutions via formal feedback, EU Transparency Register entry, Commissioner meeting, or signing EU AI Pact / Code.
  - Country-level lobbying: 18 firms flagged if appearing in member-state lobby register on AI.
  - EU-level trade-association membership: 84 firms flagged if member of major EU digital-industry association that lobbied the Act.
  - Any post-2021 AI lobbying: union of the three categories (115 firms).
- Linear probability model coefficients (percentage-point change per one-standard-deviation increase; FF48-clustered SEs in parentheses; wild-bootstrap p for EU×AI):
  - EU × AI Growth:
    - EU-level firm lobbying: +0.91 (1.22); Wild-bootstrap p = 0.44.
    - Country-level lobbying: +0.98 (0.58); Wild-bootstrap p = 0.73.
    - EU-level trade association: +0.80 (0.83); Wild-bootstrap p = 0.25.
    - Any post-2021 AI lobbying: +0.84 (1.46); Wild-bootstrap p = 0.59.
  - EU Revenue Share: +0.45 (0.77); +0.35 (0.36); −0.12 (0.64); −0.17 (0.84) across columns (not significant).
  - AI Growth: +4.26*** (1.20); +0.91* (0.52); +3.28*** (1.10); +3.90*** (0.90).
  - log Market Cap: +5.85*** (1.49); +2.10*** (0.72); +6.81*** (1.03); +8.36*** (1.04).
  - Wild-bootstrap p-values for EU×AI: 0.44, 0.73, 0.25, 0.59.
- Interpretation: EU×AI interaction does not predict post-proposal lobbying in any category; AI hiring growth and firm size strongly predict lobbying. Evidence consistent with effect reflecting compliance/implementation capability tied to embedded EU operations rather than political advocacy.

### High-Risk-AI Hiring Dictionary
- Classification rule: a Lightcast job posting is classified as high-risk AI if at least one listed skill matches the high-risk-AI skill dictionary AND its job title matches the high-risk-AI title dictionary.
- Included scope maps five high-risk areas from the April 2021 proposal COM(2021) 206 final:
  - Annex III(1) biometric identification (remote identification and verification systems).
  - Annex III(3) education and vocational training (access determination and student assessment).
  - Annex III(4) employment and worker management (recruitment, evaluation, performance monitoring) subject to HR/recruiting exclusions.
  - Annex III(5)(b) creditworthiness assessment (credit risk modeling and mortgage origination).
  - Article 6(1) Annex II Section A safety components (robotics and medical-device AI under Machinery Regulation (2006/42/EC) and Medical Devices Regulation (2017/745)).
- Excluded scope:
  - Annex III categories requiring deployer restrictions (public-benefit eligibility, law enforcement, migration, administration of justice) omitted.
  - Annex III(2) critical infrastructure excluded due to narrow 2021-proposal scope.
  - Annex II Section B safety components (motor vehicles, civil aviation, rail) excluded due to sectoral type-approval regimes.
- Excluded HR/recruiting patterns within Annex III(4): generic HR/recruiting patterns (talent acquisition, people analytics, automated recruitment, resume parsing) excluded; technical performance management variants and applicant tracking system skills retained.
- Dictionary size after exclusions: 128 entries (56 skills, 72 titles) across the five Annex categories. 

*Italic source attribution: content drawn from the supplied IMF PDF excerpt of the working paper appendix material.*

### Annex   III   category

### Annex III category

### Annex III (2021 proposal): Skills, titles, and sample dictionary entries
- Annex III(1) Biometric (103)
  - Skills: Biometrics; Fingerprint Identification And Classification; Fingerprint Recognition; Biometric Passport; Certified Biometrics Professional.
  - Titles: Biometric Health Screeners; Biometrics Technicians; Biometric Screeners.
- Annex III(3) Education and vocational training (21)
  - Skills: Educational Assessment; Admissions Operations.
  - Titles: Directors of Admissions Operations.
- Annex III(4) Employment, workers’ management (post-HR cleanup) (125)
  - Skills (ATS): Applicant Tracking Systems; CATS Applicant Tracking System.
  - Skills (technical PM): Operational/Network/Sales/Enterprise/Supplier Performance Management; Human Performance Evaluation; Oracle EPM; PeopleSoft EPM.
  - Titles: Performance Management Analysts/Directors/Consultants/Specialists/Managers.
- Annex III(5)(b) Credit-worthiness assessment (1014)
  - Skills: Credit Risk Modeling; Credit Risk Analysis; Credit Risk Management; Mortgage Loan Origination; Commercial Mortgage Loan Origination.
  - Titles: Credit Risk Managers; Credit Risk Analytics Managers; Credit Risk Underwriters; Credit Risk Associates; Credit Risk Analysts.
- Art. 6(1) + Annex II Section A safety components (2249)
  - Skills: Robotic Systems; Advanced Robotics; Industrial Robotics; Robotic Surgery; Computer Aided Diagnosis; Medical Imaging Physics.
  - Titles: Biomedical Imaging Technicians; Medical Imaging Technologists; Robotics Coaches; Biomedical Imaging Specialists; Medical Device Territory Managers.
- Total (headline) 5672 — 128 entries

Notes:
- A posting is classified as high-risk AI if at least one skill matches the skill dictionary and its title matches the title dictionary (AND-logic).
- Counts are after the HR/recruiting cleanup described in the appendix text.
- Excluded Annex III categories (2 critical infrastructure, (5)(a) public-benefit eligibility, (5)(c) emergency dispatch, 6 law enforcement, 7 migration, 8 justice) and Annex II Section B safety components under the Article 2(2) carve-out are documented above.

---

### E. Country Decomposition of the HHI-Based Subsamples — key findings
- Purpose: Assess whether the concentration-based jurisdictional capital effect reflects broad EU-level regulatory capital or jurisdiction-specific effects.
- Main pattern: Both HHI-based subsamples are Germany-dominant.
  - Germany carries 33.6% of total EU revenue in the high-HHI half and 21.5% in the low-HHI half.
  - Germany is the top-1 EU country for 83% of high-HHI firms (320 of 386) and 98% of low-HHI firms (379 of 385).
- Geographic concentration detail:
  - High-HHI subsample concentrates EU revenue more sharply in Germany, France (20.8% vs. 15.5%), and Italy (10.3% vs. 13.8%).
  - Low-HHI spreads revenue more evenly across countries.
  - Ireland is slightly over-represented in the high-HHI subsample (11 firms top-1 vs. 2 firms) but accounts for only 4.2% of high-HHI EU revenue—insufficient to drive the +47 to +86 bps effect reported in Section 7.3.
- Interpretation: The concentration effect operates through depth of engagement with major continental EU jurisdictions whose national competent authorities will implement and enforce the AI Act—consistent with a regulatory-capital interpretation at the country-authority level rather than at the Brussels-lobbying level.

Table A18: EU Member Country Decomposition of the HHI-Based Subsamples (selected figures)
- Sample split: median split at HHI = 0.1376, N High = 386, N Low = 385 (analysis sample restricted to firms with strictly positive bottom-up EU revenue).
- High-HHI subsample (concentrated) — Low-HHI subsample (diversified)
  - Germany: Agg. share 33.6% | Mean share 30.4% | # top-1 320  —  Agg. share 21.5% | Mean share 21.7% | # top-1 379
  - France: Agg. share 20.8% | Mean share 18.9% | # top-1 24  —  Agg. share 15.5% | Mean share 15.7% | # top-1 4
  - Italy: Agg. share 10.3% | Mean share 12.5% | # top-1 6  —  Agg. share 13.8% | Mean share 13.9% | # top-1 0
  - Spain: Agg. share 8.1% | Mean share 9.3% | # top-1 8  —  Agg. share 9.7% | Mean share 9.7% | # top-1 0
  - Ireland: Agg. share 4.2% | Mean share 4.2% | # top-1 11  —  Agg. share 2.7% | Mean share 2.7% | # top-1 2
  - Belgium: Agg. share 3.1% | Mean share 3.2% | # top-1 2  —  Agg. share 3.8% | Mean share 3.8% | # top-1 0
  - Sweden: Agg. share 3.1% | Mean share 3.1% | # top-1 1  —  Agg. share 4.0% | Mean share 4.0% | # top-1 0
  - Poland: Agg. share 3.0% | Mean share 3.9% | # top-1 5  —  Agg. share 4.0% | Mean share 4.0% | # top-1 0
  - Austria: Agg. share 2.0% | Mean share 2.3% | # top-1 0  —  Agg. share 3.3% | Mean share 3.2% | # top-1 0
  - Denmark: Agg. share 1.6% | Mean share 2.0% | # top-1 2  —  Agg. share 2.6% | Mean share 2.6% | # top-1 0
  - Czech Republic: Agg. share 1.3% | Mean share 1.3% | # top-1 1  —  Agg. share 2.3% | Mean share 2.2% | # top-1 0
  - Romania: Agg. share 1.2% | Mean share 1.5% | # top-1 2  —  Agg. share 2.1% | Mean share 2.1% | # top-1 0
  - Other (15 countries): Agg. share 7.9% | Mean share 7.2% | # top-1 4  —  Agg. share 14.7% | Mean share 14.3% | # top-1 0
  - Total: Agg. share 100.0% — 386  —  Agg. share 100.0% — 385

Notes on Table A18:
- Agg. share is the subsample’s total revenue in country c divided by its total bottom-up EU revenue (weighted average of within-EU shares, weighted by each firm’s total EU revenue).
- Mean share is the simple cross-firm average of each firm’s within-EU share in country c.
- # top-1 counts firms whose mean 2015–2019 share is highest in country c.
- Rows sorted by the high-HHI aggregate share, with the remaining countries collapsed into “Other.”

---

### F. Compliance Hiring Pre EU AI Act: Detail and Robustness

F.1 Variable construction
- Data sources: curated regex screen over the Lightcast distinct-skills (31,542) and distinct-titles (72,194) dictionaries.
- Two firm-year compliance hiring measures:
  - EU IT/data compliance measure: restricts to postings whose core matches are not exclusively in US-statute or UK-statute categories.
  - Broader IT/data compliance measure: includes US- and UK-statute matches as well.
- Counting rule: For each firm-year in 2015–2019, count flagged postings, divide by total firm-year postings, require at least 5 total postings per firm-year.
- Firm-level metrics: level = mean over valid years; growth = four-year-annualized change.
- Headline: Section 7.4 uses the EU IT/data compliance measure.
- Empirical note: The two measures yield quantitatively similar patterns; the EU measure is marginally sharper.

F.2 Compliance categories and EU regulatory drivers (six functional areas)
- Privacy and data protection
  - Roles: Data Protection Officers, Chief Privacy Officers, Privacy Counsels and Engineers, privacy-by-design practitioners, DPIA analysts, breach response staff, consent-management implementers, privacy professional certifications (CIPP/CIPM/CIPT from IAPP, CDPSE from ISACA).
  - Principal driver: General Data Protection Regulation (applicable from 25 May 2018; entered into force 24 May 2016), together with Directive 95/46/EC and the ePrivacy Directive.
- Cybersecurity
  - Skills: Security operations (SIEM, SOAR, EDR, DLP), incident response, vulnerability and threat management, network/cloud/endpoint security, cryptography, certifications including ISO 27001 and ISO 27701.
  - Principal drivers: GDPR Article 32 (security of processing), the NIS Directive, and the EU Cybersecurity Act.
- Data governance and lifecycle
  - Skills: Data lineage and mapping, data catalogs and metadata management, data classification, master data management, records management and retention, e-discovery.
  - Principal driver: GDPR Article 30 (records of processing activities) together with Article 5(1).
- IT governance, risk, audit, and vendor management
  - Skills: IT general controls, GRC platforms, enterprise and operational risk management, internal IT audit, vendor and third-party risk management.
  - Principal drivers: GDPR Articles 24, 28, and 39 (controller demonstrability, controller–processor contracting and sub-processor due diligence, DPO compliance monitoring).
- AI and algorithmic governance
  - Skills: Explainable AI, algorithmic accountability and fairness, AI bias testing, AI risk and safety management, model validation and monitoring.
  - Principal driver: GDPR Article 22 (right against solely automated decision-making) together with Article 35(3)(a) and the transparency obligations of Articles 13(2)(f) and 14(2)(g). The EC’s High-Level Expert Group Ethics Guidelines for Trustworthy AI (April 2019) and the AI White Paper (February 2020) shaped this category in the pre-event window.
- Anti-money laundering and financial-crime compliance
  - Skills: AML programs, KYC and customer due diligence, sanctions screening, transaction monitoring.
  - Principal driver: AML Directive series (4MLD/5MLD/6MLD); for US-listed firms, AML hiring is overwhelmingly driven by US and global frameworks (Bank Secrecy Act, FinCEN, OFAC, FATF) rather than AMLD specifically.

Additional instruments captured indirectly: eIDAS, PSD2, MiFID II, and Council of Europe Convention 108.
- Broader compliance measure additionally captures US privacy and financial statutes (Sarbanes-Oxley, Gramm-Leach-Bliley, FCRA), US bank IT examination guidance (FFIEC), US export controls (ITAR, EAR), and UK transpositions and UK-specific regulation (UK GDPR, Data Protection Act 2018, PECR, UK NIS Regs 2018, ICO, FCA/PRA cyber-resilience, BS 7799, SMCR, UK Open Banking, Cyber Essentials).
- Caveats:
  - Non-EU cybersecurity frameworks (NIST CSF, FedRAMP, FISMA, HITRUST, CMMC; COBIT and ITIL) are categorized under cybersecurity and therefore enter the EU measure.
  - GDPR’s cross-border-transfer regime (Articles 44–50, including Standard Contractual Clauses, Binding Corporate Rules, and Schrems II) is substantively important but not captured by the Lightcast vocabulary.

F.3 GDPR-to-AI-Act article mapping — functional carry-over
- GDPR Article 30 (records of processing) → AI Act Article 10 (data-quality and data-management obligations).
- GDPR Article 32 (security of processing) → AI Act Article 15 (cybersecurity obligations).
- GDPR Article 24 (demonstrability) → AI Act Article 17 (quality-management-system requirement) and Article 72 (post-market monitoring).
- GDPR Articles 37–39 (DPO function) → organizational structures relevant to AI Act risk-management and governance obligations.

Interpretation:
- Firms that staffed up for GDPR plausibly face a lower marginal cost of complying with the mapped AI Act articles.
- Compliance hiring on its own is not priced: the AI Act is principles-based and risk-tiered with broader scope and interpretive ambiguity relative to GDPR.
- Unique AI Act features without direct GDPR analogues: conformity assessments under Article 43; producer/deployer asymmetry under Articles 16 and 26; high-risk classification regime under Article 6 and Annex III.
- What matters for these features is both explicit compliance staffing and jurisdictional embedding (capacity to anticipate, interpret, and shape implementation-stage decisions by national competent authorities, notified bodies, and standards organizations).

F.4 Compliance category taxonomy (Table A19) — examples
- Privacy and data protection
  - Example Lightcast skills/titles: Data Protection Officer; Privacy Engineer; DPIA analyst; CIPP/CIPM/CIPT certifications.
  - Principal EU regulation: GDPR.
  - Key provisions: Articles 37–39 (DPO function); Article 25 (privacy by design); Article 35 (DPIA); Articles 12–22 (data subject rights); Articles 33–34 (breach notification).
- Cybersecurity
  - Example Lightcast skills/titles: SIEM/SOAR/EDR/DLP operations; incident response; ISO 27001/27701; CISSP, CISA, CISM.
  - Principal EU regulation: GDPR; NIS Directive; EU Cybersecurity Act.
  - Key provisions: Article 32 (security of processing); Article 5(1)(f) (integrity and confidentiality); Articles 33–34 (breach notification).
- Data governance and lifecycle
  - Example Lightcast skills/titles: Data lineage and catalogs; metadata management; records retention; e-discovery.
  - Principal EU regulation: GDPR.
  - Key provisions: Article 30 (records of processing); Article 5(1)(b)–(e) (purpose limitation, data minimization, accuracy, storage limitation).
- IT governance, risk, audit, and vendor management
  - Example Lightcast skills/titles: IT general controls; GRC platforms; internal IT audit; third-party risk management.
  - Principal EU regulation: GDPR.
  - Key provisions: Article 24 (demonstrability); Article 28 (processor contracting and sub-processor due diligence); Article 39 (DPO compliance monitoring).
- AI and algorithmic governance
  - Example Lightcast skills/titles: Explainable AI; algorithmic accountability; AI bias testing; model validation.
  - Principal EU regulation: GDPR; HLEG Ethics Guidelines (2019); AI White Paper (2020).
  - Key provisions: Article 22 (automated decision-making safeguards); Article 35(3)(a) (mandatory DPIA for automated profiling); Articles 13(2)(f), 14(2)(g) (transparency on automated-decision logic).
- AML and financial-crime compliance
  - Example Lightcast skills/titles: KYC and customer due diligence; sanctions screening; transaction monitoring; CAMS, CAFP.
  - Principal EU regulation: AML Directive series (4MLD - 6MLD).
  - Key provisions: For US-listed firms, AML hiring is overwhelmingly driven by US frameworks (Bank Secrecy Act, OFAC); GDPR Article 6(1)(c) provides lawful basis for AML processing.

Notes:
- Functional areas are captured by the core-tier patterns in the merged Lightcast taxonomy (77 patterns across 11 categories).
- The example skills/titles column lists representative entries.

*Jurisdictional Capital and AI Regulation: Evidence from the EU AI Act — Working Paper No. WP/2026/180*

---


_Source: https://www.imf.org/-/media/files/publications/wp/2026/english/wpiea2026180-source-pdf.pdf_