POLISY research memo · Political Ideology × AI × Innovation

POLISY: What Is New, What Is Not, and Where the Frontier Is

A critical assessment of the first POLISY run on politics, AI and innovation

28 September 2026 · figures and tables built from the lab run of 2026-09-28 13:38 · polisy_lab 0.2.0

In brief

  • POLISY’s first run linked workforce partisanship (VRscores) to ten AI-exposure measures, IRS county migration and the Correlates of State Policy, and graded 28 findings. Checked against the 2024–2026 literature and put through three stress tests, none of its headline findings is new in the form it was first reported. Most replicate documented patterns; several are artefacts of how VRscores is built.
  • The “AI exodus” is not about AI. Households did leave AI-exposed counties, and faster every year, but 88% of the 2020–22 gradient disappears once density, education, income, housing costs, climate, the 2016 vote and telework are held equal. What grew after 2016 is the partisan and telework gradient of migration (Figure 3).
  • AI exposure’s Democratic lean is composition. Generative-AI exposure leans Democratic only because of who holds the exposed jobs: once education, pay, telework, gender and race are held equal, it leans neither way. Routine-automation risk is different: it leans Republican net of all of these and of occupation group (Figure 4).
  • Most of what looks like organisational ideology is occupational structure. The occupations an industry employs account for 72% of the variance in its partisanship and 54% of the variance in its political balance (Figure 5).
  • The frontier is at the level of the firm, which the first run never reached. Four seams look open and fit your projects: structural versus elective partisanship in organisations; the political distance between those who decide on AI and those whose work it touches; whether leaders or inventors set the direction of invention when their politics differ; and technology as a force that reshapes the political composition of firms (Section 5).

1. What POLISY is, and where it sits

Your workbook maps 211 items. It is strongest on leader ideology and upper echelons, on workforce ideology and fit, on partisanship in finance and on cross-border knowledge flows. Two projects anchor it: P1 asks whether political incongruence between a CEO and the workforce shapes innovation; P2 asks whether CEO ideology shapes cross-border innovation networks, and deliberately makes its claim about the direction and structure of innovation rather than its volume, because the evidence on volume is mixed (Kiss et al., 2026; Fehder et al., 2024).

POLISY’s first run worked one level up from that seam. Its units were 825 occupations, 246 industries, 359 metro areas, 51 states and 3,130 counties. It linked VRscores (Kagan, Frake & Hurst, 2026) to AI exposure from Felten, Raj & Seamans (2021) and DAIOE 2010–2024 (Engberg et al., 2026), to comparison measures from Eloundou et al. (2024), Webb (2020) and Frey & Osborne (2017), to IRS county flows from 2012–13 to 2021–22, and to state policy. It never reached the firm: the public VRscores employer file has no identifiers, DIPI (Mannor & Busenbark, 2025) and Compustat sit in the POLISY_DA pipeline but not in the lab, and patents were deferred.

That has two consequences. First, the first run speaks mainly to the political economy of AI, a literature in economics and political science, rather than to politics inside organisations. Second, your workbook holds almost none of that literature: no AI-exposure economics, no politics of automation, no recent work on AI and partisanship, and not the two 2025–26 papers closest to P1 (Frake, Hurst & Kagan, 2026; Red, blue, and purple employee populations, 2025). Appendix B lists what to add. Figure 1 places the first run’s findings and the proposed seams on one map.

first run: occupations, industries, counties and states POLITICS Workforce Partisanship at work (VRscores) Leaders CEO, top team, board (DIPI) Environment State policy, local vote ORGANISATIONS Composition Lean, balance, sorting Incongruence Leaders against workforce Decisions What to adopt and invent INNOVATION & TECHNOLOGY Exposure Which tasks AI touches Adoption Who uses AI, how fast Direction What gets invented 1 4 2 3 the four research seams (Section 5), all at the level of the firm the first POLISY run, which linked politics to AI directly and never reached the firm
Figure 1. The research seam: from politics, through organisations, to innovation and technology.The dashed arc is what the first run linked: workforce partisanship and the political environment against AI exposure and migration, with occupations, industries, counties and states standing in for firms. The numbered arrows are the four seams of Section 5: (1) structural and elective partisanship; (2) who decides and who is exposed; (3) whether leaders or inventors direct invention; (4) technology reshaping the political composition of firms. Grey arrows are the links the seams build on.

2. What the first run found, and what is new

Table 1 sets every finding against the closest published or working-paper evidence, checked in September 2026. The verdicts are blunt on purpose. A pattern the literature already documents is established; one that the stress tests explain by who holds the jobs or where they live is composition; one produced by how the data were built is an artefact.

Table 1. Every finding of the first run, against the literature.
FindingLab gradeVerdictClosest literatureWhy
The 'AI exodus' is mostly the partisan and telework geography of migrationstress-mig · countyRobustTestRamani & Bloom 2021; Economic Innovation Group 2023Shows that mig-ai-exodus is the partisan and telework geography of migration.
AI exposure's Democratic lean is composition; automation risk's Republican lean is notstress-occ · occupationRobustExtensionBloom & Makridis 2026; Muro et al. 2026Moves the composition argument from AI use (surveys) to AI exposure (voter files), and shows that it holds for generative AI but not for routine automation.
Most of an industry's partisanship, and half of its political balance, is the occupations it employsstress-structure · industryRobustExtensionFrake, Hurst & Kagan 2026; Castiglia 2025; Gupta, Briscoe & Hambrick 2017Consistent with segregation net of occupation; new is the consequence for organisation-level ideology and balance measures.
AI exposure measures disagree about who is exposed: benchmark and language-model measures lean Democratic, patent and automation measures Republicanocc-waves · occupationFragileExtensionWebb 2020; Yin & Ogut 2026That exposure measures disagree is known; that their partisan sign depends on the measure and on composition is a caution for studies of AI and politics.
Occupational mix explains most of an industry's partisanship; the rest points to place and missionind-composition · industryRobustExtensionFrake, Hurst & Kagan 2026See stress-structure.
Interstate movers go to more Republican states, and the gap widened after 2019mig-partisan-state · stateRobustEstablishedMcKee, Smith & Zhang 2025; Gimpel, Newton & Reeves 2025Movers to Republican states, rising after 2019, is well documented.
AI exposure looks Democratic across occupations, but the link runs through educationocc-ai-education · occupationRobustEstablishedMuro et al. 2026; Bloom & Makridis 2026; Eloundou et al. 2024Exposure concentrated among educated, Democratic-leaning workers.
AI-exposed counties lose richer households than they gainmig-income · countyDescriptiveEstablishedEconomic Innovation Group 2023High earners leaving large cities is documented in the same IRS data.
Education pulls occupations Democratic and pay pulls them Republican; the pay effect runs through who does the jobocc-education-vs-pay · occupationRobustEstablishedGethin, Martínez-Toledano & Piketty 2022Education left, income right: the Brahmin left and merchant right, by occupation.
Partisans are sorting into different employers while sorting across industries and occupations declinessorting · employerSuggestiveEstablishedFrake, Hurst & Kagan 2026; Fos, Kempf & Tsoutsoura 2022; Colonnelli, Pinho Neto & Teso 2025Also in the VRscores report (8.4% to 8.7%).
County movers go to more Republican counties than they leavemig-partisan · countySuggestiveEstablishedMcKee, Smith & Zhang 2025; Gimpel, Newton & Reeves 2025The county-level version of the same pattern.
Households are leaving AI-exposed counties, but the gradient is mostly density, politics and teleworkmig-ai-exodus · countyFragileCompositionRamani & Bloom 2021; Economic Innovation Group 2023AIGE stands for dense, Democratic, telework-intensive counties (stress-mig).
Households left states with liberal economic policy and AI-exposed jobs; at the state level the two are hard to separatestate-environment · stateSuggestiveCompositionEconomic Innovation Group 2023Policy liberalism, costs and AI exposure cannot be separated across 51 states.
AI-exposed industries have more Democratic workforces, mostly because their workers are highly educated; language-model exposure keeps a modest link within sectorsind-ai-partisanship · industryFragileCompositionBloom & Makridis 2026Runs through education (r = 0.86 with exposure).
Two kinds of AI exposure: language-AI jobs lean Democratic, perception-AI jobs Republican, until you control for educationocc-two-ais · occupationFragileComposition–Disappears with education and demographics.
Movers trade down in AI exposure, and more so since 2020mig-exposure-gap · countyDescriptiveCompositionRamani & Bloom 2021Moves from dense cores to suburbs and smaller places.
Where party is inferred rather than registered, metro workforces drift Democratic much fastermetro-regime · metroRobustArtefactKagan, Frake & Hurst 2026Party is inferred rather than registered in these states; flagged in the VRscores report.
Partisan drift clusters in space, along state linesmetro-space · metroRobustArtefactKagan, Frake & Hurst 2026The clusters follow states' party-data regimes.
AI-exposed states have more Democratic workforces, and VRscores covers them more denselystate-aige · stateSuggestiveArtefactKagan, Frake & Hurst 2026VRscores covers AI-exposed labour markets more densely.
States where party is inferred rather than registered drift Democratic fasterstate-regime · stateSuggestiveArtefactKagan, Frake & Hurst 2026As metro-regime, at the state level.
Nearly every industry drifted Democratic, and AI exposure does not predict by how muchind-drift · industryRobustArtefactKagan, Frake & Hurst 2026Party is fixed at 2024, so drift is cohort replacement.
Large employers drift Democratic eight times as often as Republican; the most Republican and the largest drift mostemp-drift · employerRobustArtefactKagan, Frake & Hurst 2026As ind-drift, for employers.
State maps built from VRscores metros misplace workers of multi-state metrosstate-coverage · stateArtefactArtefact–Multi-state metros credited to one state.
The largest employer drifts look like artefacts of restructured firmsemp-artifacts · employerArtefactArtefact–Restructured firms.
AI and big-tech workforces barely joined the Democratic drift, and several moved Republicanemp-bigtech · employerSuggestiveDescriptive–Intriguing given the tech industry's rightward turn in 2024, but 34 firms, and voter files miss non-citizen workers.
Dynamic exposure mostly rises for everyone; where it rose most, workforces are more Democraticocc-daioe · occupationDescriptiveDescriptiveEngberg et al. 2026A cumulative index rises for everyone; relative standing moves little.
Which state policy variables move with AI outcomes within states over timepolicy-twfe · stateSuggestiveDescriptive–One of 23 pairs passes a 10% false-discovery threshold; underpowered.
The US migration network in the latest yearmig-network · stateDescriptiveDescriptive–The corridors are well known.
State archetypes across politics, AI and innovationlandscape · stateDescriptiveDescriptive–Red and blue state archetypes.
Occupations far more partisan than their exposure, pay, education and demographics predictocc-anomalies · occupationDescriptiveDescriptive–Residual outliers; several look like ethnic occupational niches (e.g. manicurists).
Actual AI adoption by firms (BTOS)adoption · sector/stateNeeds dataNot run–Needs the Census BTOS files.
AI patenting: growth, geography and technologypatents · patent/county/stateNeeds dataNot run–Needs PatentsView (deferred).

Lab grade: how well the pattern holds in POLISY's own data (after the stress tests). Verdict: how it stands against published and working-paper evidence checked in September 2026. Counts: test 1, extension 4, established 6, composition 5, artefact 8, descriptive 6, not run 2.

The balance is clear. Only a handful of findings extend a conversation; the rest replicate documented patterns, dissolve into composition, describe, or are artefacts. About a quarter are artefacts of VRscores: each worker’s party is fixed at the November 2024 voter file, so every “drift” is cohort replacement, and in states where party is inferred rather than registered the drift is larger again (Kagan, Frake & Hurst, 2026). Your own VRscores report found the same.

This is not a failure. A lab that rediscovers known patterns from new data has validated its plumbing. But it means POLISY’s contribution has to come from what it can link next, not from what it found first.

3. Three stress tests

Each test takes a headline finding, names the first rival explanation a referee would raise, and re-estimates the finding with that explanation held equal. The tests run inside the lab (analyses/stress.py) and re-grade the findings they test, so the interactive lab and the tables above already reflect them.

3.1 AI exposure and county migration

The claim was that households are leaving AI-exposed counties, faster every year and within the same state: one SD more AI exposure (AIGE) meant −0.16 points of households a year in 2013–16 and −0.86 in 2020–22. The rivals are well known. Remote work emptied dense cores into suburbs and smaller places (Ramani & Bloom, 2021); high earners left large expensive cities (Economic Innovation Group, 2023); movers went to Republican places (McKee, Smith & Zhang, 2025; Gimpel, Newton & Reeves, 2025). AIGE is high exactly where density, telework and Democratic votes are high.

−1.0 −0.8 −0.6 −0.4 −0.2 0.0 Change in net domestic migration (percentage points of households a year) per SD of AI exposure (AIGE) 2013-16 2017-19 2020-22 AIGE, state × year effects + population density + education and income + housing costs, January temperature, 2016 vote + telework (industry mix)
Figure 2. The county AI-exposure gradient of migration disappears once the rival explanations are held equal.Each dot is the coefficient on AIGE (z-scored) in a regression of net domestic migration on AIGE and the controls named in the legend, with state-by-year fixed effects, weighted by households (IRS returns), for county-years in each period; lines are 95% intervals from errors clustered by county. Darker dots add more controls. 2020–22: −0.86 with state−by−year effects only, −0.11 with every control.

With state-by-year fixed effects only, the 2020–22 coefficient is −0.86 (t = −13.6). Adding density, education, income, housing costs, January temperature, the 2016 Republican vote share and the share of the county’s jobs that can be done at home (Dingel & Neiman, 2020) brings it to −0.11 (t = −2.2). Table 2 shows what does the work: the 2016 vote and telework.

Table 2. The full migration model: which county traits predict net migration.
Per SD of2013-162017-192020-22
AI exposure (AIGE)−0.13[t = −4.2]−0.17[t = −3.7]−0.11[t = −2.2]
Population density (log)+0.26[t = 4.5]+0.28[t = 4.1]+0.02[t = 0.3]
Bachelor's degree share+0.31[t = 5.8]+0.31[t = 4.5]+0.43[t = 6.2]
Average income (log)+0.01[t = 0.2]+0.01[t = 0.1]−0.12[t = −1.9]
Housing costs (log)+0.03[t = 0.8]+0.01[t = 0.1]−0.05[t = −1.0]
January temperature−0.13[t = −1.1]−0.28[t = −1.9]−0.07[t = −0.6]
Republican vote share, 2016+0.33[t = 7.1]+0.61[t = 11.2]+0.74[t = 15.4]
Jobs that can be done at home−0.21[t = −5.9]−0.27[t = −5.7]−0.30[t = −6.4]

Same models as the last row of the previous table, every term shown (z-scored). The 2016 vote comes from the county returns; telework is the share of the county's jobs (by sector) that can be done at home, after Dingel & Neiman (2020).

The year-by-year estimates in Figure 3 tell the same story over time. The controlled AI coefficient is small and flat from 2013 to 2022. What grows is the partisan gradient: the coefficient on the 2016 vote rises from +0.33 in 2013–16 to +0.74 in 2020–22, net of every other control.

2013 2014 2015 2016 2017 2018 2019 2020 2021 2022 −1.0 −0.5 0.0 0.5 1.0 Points of households a year per SD AI exposure, state effects only Telework, all controls AI exposure, all controls 2016 Republican vote, all controls
Figure 3. What grew after 2016 is the partisan and telework gradient of migration, not the AI gradient.One regression per year (the second year of each IRS filing pair) with state fixed effects, weighted by households; shaded bands are 95% intervals clustered by county. The dashed line is AIGE with state effects only; the solid lines come from the model with every control (density, education, income, housing costs, January temperature, the 2016 Republican vote share and the share of the county's jobs that can be done at home).

Verdict. The “AI exodus” should be dropped as a finding. The partisan gradient that survives is interesting, but it is a crowded political-geography question, and the IRS files cannot see movers’ party, occupation or employer. They cannot tell political sorting apart from taxes, housing supply or pandemic-era policy.

3.2 The partisan face of AI exposure

Across occupations, the correlation between exposure and the Republican share of a workforce ran from −0.18 for generative-AI exposure (DAIOE) to +0.23 for Webb’s AI-patent measure. The rival explanation is composition. Generative-AI exposure falls on educated, disproportionately female, teleworkable occupations (Eloundou et al., 2024; Gmyrek, Berg & Bescond, 2023), and each of those traits leans Democratic.

−4 −2 0 2 4 Republican share of an occupation's workers (percentage points) per SD of the measure GPT automation potential (Eloundou et al.) GPT exposure with tools (Eloundou et al.) GPT exposure, direct (Eloundou et al.) AIOE (Felten, Raj & Seamans) DAIOE, all applications DAIOE, generative AI AI patents (Webb) Robot patents (Webb) Software patents (Webb) Computerisation risk (Frey & Osborne) ← more Democratic workforce more Republican workforce → Measure alone + education, wage + telework + gender, race + occupation group
Figure 4. Generative-AI exposure leans Democratic only through who holds the jobs; routine-automation risk leans Republican under every set of controls.Each dot is the coefficient on one exposure measure (z-scored) in a regression of the Republican share of an occupation's VRscores workers on the measure and the controls in the legend, weighted by matched workers; lines are 95% intervals (HC1). Darker dots add more controls. Gender and race shares exist for about 300 occupations, so the last two models use that smaller sample. Measures are grouped by how they were built: language-model exposure, AI benchmark exposure, patent text, and routine automation.

The test adds telework, then gender and race shares, then occupation-group fixed effects. Generative-AI exposure leans Democratic on its own (−2.0 points per SD, t = −3.3). The lean disappears once gender and race are held equal (−0.1) and stays near zero with occupation groups (+0.1). Computerisation risk goes the other way and strengthens with controls, to +2.4 points per SD (t = 4.0). Table 3 puts the two waves in one model.

Table 3. Two exposure measures side by side: generative AI against routine automation and AI patents.
Measures in the same model+ education, wage+ telework+ gender, race+ occupation group
DAIOE, generative AIDAIOE, generative AI vs Computerisation risk (Frey & Osborne)−3.0[−3.9]−2.8[−2.8]−1.4[−1.3]−0.9[−0.8]
Computerisation risk (Frey & Osborne)DAIOE, generative AI vs Computerisation risk (Frey & Osborne)+2.9[3.6]+2.9[3.5]+3.5[4.8]+2.6[3.9]
DAIOE, generative AIDAIOE, generative AI vs AI patents (Webb)−2.0[−2.9]−2.0[−2.2]−0.3[−0.2]+0.1[0.1]
AI patents (Webb)DAIOE, generative AI vs AI patents (Webb)+1.5[2.0]+1.5[2.2]+1.4[1.7]+1.3[1.8]
GPT exposure with tools (Eloundou et al.)GPT exposure with tools (Eloundou et al.) vs Computerisation risk (Frey & Osborne)−1.4[−1.5]−0.9[−1.0]−0.7[−0.9]−0.7[−0.6]
Computerisation risk (Frey & Osborne)GPT exposure with tools (Eloundou et al.) vs Computerisation risk (Frey & Osborne)+2.4[2.8]+2.5[2.9]+3.2[4.9]+2.4[4.0]

Republican share (points) per SD of each measure when both enter the same model; t statistics in brackets (HC1). Controls are cumulative from left to right, as in the figure above.

Three things follow.

  1. This extends a known result rather than overturning one. Bloom & Makridis (2026) find that the partisan gap in AI use disappears once education, occupation and industry are held equal, and Muro et al. (2026) map exposure onto Democratic counties. The stress test moves the argument from use to exposure, and from surveys and county votes to voter-file partisanship.
  2. The two waves are not symmetric. Automation risk keeps a Republican lean among otherwise similar occupations. That fits the politics of automation: routine workers facing status decline move right (Kurer, 2020; Anelli, Colantone & Stanig, 2021; Frey, Berger & Chen, 2018; Gallego & Kurer, 2022). Generative AI’s partisan face, by contrast, is carried entirely by who holds the jobs. A generative-AI backlash therefore need not follow the partisan script of the automation backlash. It is more likely to run through professional, gender and status groups, and through organisations, than through a partisan realignment.
  3. Exposure measures are not politically neutral. Felten, Raj & Seamans’s AIOE turns from a slight Democratic lean to a clear Republican one once composition is held equal. The choice of measure decides whether AI’s exposed workforce looks blue or red. A related selection problem affects usage-based measures (Yin & Ogut, 2026). Any study that uses exposure to explain political outcomes (Grant, Green & Evans, 2026; Antoniades et al., 2025) should report more than one measure.

Caveats. Occupation partisanship is national, so occupations concentrated in Republican regions look Republican for that reason alone. Demographic shares cover about 300 occupations.

3.3 How much of workforce partisanship is structural

The first run predicted each industry’s partisanship from the occupations it employs, using national occupation partisanship and the industry’s staffing pattern. The stress test extends this to political balance: how evenly a workforce is split, the construct Castiglia (2025) links to lower innovation.

0.3 0.4 0.5 0.6 0.7 Predicted from the occupations employed 0.3 0.4 0.5 0.6 0.7 Actual R² = 0.72 Republican share 0.5 0.6 0.7 0.8 0.9 1.0 Predicted from the occupations employed 0.5 0.6 0.7 0.8 0.9 1.0 Actual R² = 0.54 Political balance (1 = evenly split)
Figure 5. Industries' partisanship, and half of their political balance, follow from the occupations they employ.Each circle is a four-digit industry with at least 70% of its jobs linked to VRscores occupations, sized by matched workers. Predicted = national Republican share of each occupation, weighted by the industry's staffing pattern (OES). Balance = 1 − |2p − 1|, where p is the Republican share. R² from a regression weighted by workers; the dashed line is equality.

Occupations account for 72% of the variance in industry partisanship and 54% of the variance in balance, across 178 industries. The result is consistent with Frake, Hurst & Kagan (2026), who find that the partisan gap between Democrats’ and Republicans’ coworkers falls from 11.7 to 2.9 points once geography, industry and occupation are held equal.

The consequence for management research is direct. Organisation-level political constructs are measured as raw aggregates of donations or registrations: organisational ideology (Gupta, Briscoe & Hambrick, 2017), workforce balance (Castiglia, 2025), and P1’s leader–workforce incongruence. They therefore mix two things. One is the occupational structure the work requires. The other is whatever the organisation attracts, selects and keeps. Only the second is plausibly organisational. This cannot be tested at the industry level, but it can be tested at the firm level. That is Seam 1.

3.4 What survives, and why it matters

Three results survive the stress tests and are unusual enough to build on. None is a paper on its own; each is the foundation for one.

  1. The asymmetry between technological waves. Routine automation leans Republican among otherwise similar occupations; generative AI’s partisan lean is carried entirely by who holds the jobs. This predicts that conflict over generative AI will run through professional, gender and status groups inside organisations rather than through a partisan realignment. Seams 2 and 4 test that prediction.
  2. Exposure measures are not politically neutral. The sign of AI’s partisan incidence depends on the measure and on what is held equal. Any study of AI and politics should report more than one measure; POLISY already does.
  3. Organisational ideology is mostly occupational structure. The field’s core constructs mix what the work requires with what the organisation selects. Seam 1 separates them, and P1 needs the separation anyway.

4. What is not new, and why

Established, or explained by composition. Five patterns are already documented or are explained by composition:

Artefacts. The Democratic drift of employers, industries, metros and states reflects cohort replacement, because party is fixed at 2024, and inferred rather than registered party. Treat any VRscores trend as composition, never as political change.

Intriguing but fragile. Large AI and big-tech workforces barely joined the Democratic drift: −1.1 points against −4.1 for other large employers. The comparison covers 34 firms, and voter files miss non-citizen workers, a large share of technology workforces. It is not a finding until it holds on firm-matched data restricted to registration states.

5. The research frontier: four seams

A seam qualifies if it meets four tests:

All four seams below are at the level of the firm. Each continues something POLISY found.

5.1 Seam 1: Structural and elective partisanship in organisations

Question
How much of an organisation’s political composition follows from the work it does and where it does it? That includes its lean, its balance, and the distance between leaders and workforce. And does only the remainder shape what the organisation does, such as innovating, taking stands and keeping people? The remainder is what the organisation attracts, selects and retains.
Builds on
Organisational political ideology (Gupta, Briscoe & Hambrick, 2017); workforce balance and innovation (Castiglia, 2025); workplace segregation net of occupation (Frake, Hurst & Kagan, 2026); ideological skew through attraction, selection and attrition (Red, blue, and purple employee populations, 2025); ideological misfit and exit (Bermiss & McDonald, 2018); co-partisan hiring (Colonnelli, Pinho Neto & Teso, 2025); executive sorting (Fos, Kempf & Tsoutsoura, 2022).
What is missing
A counterfactual. Organisation-level measures are raw aggregates, and none asks what an organisation’s occupations and locations would imply. Frake et al. show that structure removes three quarters of coworker similarity. Management research still treats the raw aggregate as the organisation’s ideology, and no study tests which component drives outcomes.
What POLISY adds
The machinery for the counterfactual: occupation partisanship, staffing patterns and exposure on one crosswalk. It also adds first evidence that structure explains most of industry partisanship and half of balance (Figure 5).
Rival explanations
Location: a firm in a Republican county hires Republican workers, which is structural too, but by place. Also founding conditions and founder ideology; unionisation and public ownership; the party-data regime; and who keeps an online profile.
Data and design
Firm × occupation × metro headcounts (Revelio positions, the data behind VRscores) give expected partisanship E[p] = Σ so,m po,m. Elective partisanship is p − E[p]. Apply the same split to balance and to CEO–workforce distance. Then re-estimate balance → innovation and P1’s incongruence → innovation with both components. If the theory holds, only the elective component predicts innovation and turnover.
Mechanisms
Attraction, selection, attrition and co-partisan hiring networks create elective partisanship. Value congruence (Edwards & Cable, 2009) and affective polarisation (Iyengar et al., 2019) explain why only that part shapes cooperation, voice and exit. Structural partisanship is a by-product of technology and location.
Towards causality
Shocks that change elective composition with structure held fixed: acquisitions, exogenous CEO departures, headquarters relocations (Barber & Blake, 2024). Shift-share predictions of structural change from national occupation growth. Election periods raise the salience of partisanship without changing composition (Reinwald et al., 2026), giving a within-firm test of mechanism.
First paper
Structural and elective partisanship: decomposing the political composition of organisations. A measurement paper with a reanalysis of balance and incongruence. It becomes P1’s measurement backbone. Outlets: Organization Science, Strategic Management Journal, Organizational Research Methods.

5.2 Seam 2: Who decides and who is exposed

Question
When the leaders who decide how to adopt AI differ politically from the workers whose tasks AI touches, what changes? Is adoption faster or slower, and more often labour-replacing than labour-augmenting? Is it followed by more voice or exit?
Builds on
CEO partisanship and AI hiring (Chu, Li & Zhu, 2026); partisanship and AI use at work (Bloom & Makridis, 2026); exposure and support for AI regulation (Grant, Green & Evans, 2026); CEO ideology and inventiveness (Kiss et al., 2026); employees’ reactions to their CEO’s politics (Wowak, Busenbark & Hambrick, 2022).
What is missing
The two sides have only been studied apart. Chu, Li & Zhu (2026) look at the CEO: Republican-leaning CEOs devote a smaller share of hiring to AI. Bloom & Makridis (2026) look at individual workers. No study measures the political distance between those who decide and those who are exposed. Section 3.2 shows that exposure is carried by identifiable occupational groups, which makes that distance measurable inside a firm.
What POLISY adds
Occupation partisanship and ten exposure measures on one crosswalk. It also adds the finding that exposure’s partisan face depends on the measure and on composition, which tells a design which measures to report and what to hold equal.
Rival explanations
CEO ideology standing in for risk preference (Christensen et al., 2015); technological opportunity by industry; tech-hub location; workforce partisanship standing in for education and occupation (Seam 1). Also reverse causality: AI-intensive firms hiring Democratic engineers.
Data and design
DIPI for leaders, and the VRscores employer panel matched to Compustat. Firm occupation mix × exposure for the exposed. AI adoption from job postings, AI patents, and AI mentions in 10-Ks and earnings calls. Voice and exit from NLRB petitions, WARN notices and Revelio turnover. The design is a triple difference around the release of ChatGPT on 30 November 2022: the workforce’s pre-period exposure × leader–exposed distance × post.
Mechanisms
Motivated reasoning that discounts costs borne by out-partisans (Iyengar et al., 2019); trust and legitimacy (Edwards & Cable, 2009; Wowak, Busenbark & Hambrick, 2022); hiring AI talent through co-partisan networks (Colonnelli, Pinho Neto & Teso, 2025); moral framing of automation versus augmentation.
Towards causality
Exogenous CEO turnover varies leader politics with the workforce held fixed. The partisan composition of the local CEO labour market can serve as an instrument (Chu, Li & Zhu, 2026). State and city AI employment laws (New York City Local Law 144, 2023; Colorado’s AI Act, 2024) shift the cost of adoption. The 2016–2022 period provides a check on pre-trends.
First paper
Who decides, who is exposed: political distance and the adoption of AI. The AI-era extension of P1. Outlets: Strategic Management Journal, Organization Science, Management Science.

5.3 Seam 3: Who directs the direction of invention?

Question
When a firm’s leaders and its inventors differ politically, whose views set the direction of the firm’s invention? And does the direction of AI invention follow politics even though AI carries no party label, through whose work it automates?
Builds on
Inventors’ party and the direction of invention (Dossi & Morando, 2026; Fehder et al., 2024); political sentiment and patenting (Engelberg et al., 2025); partisan lines in inventor collaboration (Chen & Gong, 2025); CEO ideology and inventiveness (Kiss et al., 2026); ideological misfit and exit (Bermiss & McDonald, 2018); co-partisan hiring (Colonnelli, Pinho Neto & Teso, 2025); the direction of technical change (Acemoglu & Restrepo, 2020; Autor et al., 2024). P2’s positioning already argues for direction over volume.
What is already known
Dossi & Morando (2026) settle the question at the level of the inventor. They link about 95,600 inventors in four closed-primary states to voter files. Democrats are 31% more likely than Republicans to patent green technologies and 35% more likely to patent women’s-health technologies; they are 39% less likely to patent weapons. The gaps are not explained by ability or skills, they widened through 2015 and have stayed wide since, and they carry into citations. Asking whether partisan inventors pursue partisan-coded AI (climate AI, health AI, defence AI) would repeat their design on a new technology class.
What is missing
Two things their design leaves open. The first is organisations. Adding organisation fixed effects removes 38–42% of the green and weapons gaps, and they treat that share as a robustness check rather than a question. Nobody asks whether leaders’ politics steer the firm’s portfolio, or how a contest between leaders and inventors over direction is resolved: by reassignment, exit or voice. The second is technologies without a party label. They restrict attention to issues with a clear party position and a clear mapping to patents, and note that other politically relevant technologies are harder to identify. AI is the leading case. It has no stable party position, but it carries partisan weight through incidence, meaning whose work it replaces, which is exactly what POLISY measures (Section 3.2).
What POLISY adds
Occupation partisanship and exposure on one crosswalk, which turns any patent mapped to occupations into a measure of whose work it targets. Also VRscores employer composition, DIPI for leaders, and PatentsView adapters written but not yet run. Their data recipe can be replicated: closed-primary voter files for Florida, New Jersey, New York and Pennsylvania, matched to PatentsView on name and city (53% of patents).
Rival explanations
Organisational specialisation: when firms specialise, organisation effects absorb technology, as Dossi & Morando (2026) note. Also customers and procurement, especially defence contracts; location; and reverse selection, where weapons firms attract Republican leaders.
Data and design
Firm portfolios from DISCERN, inventor party from the four-state voter files, and leader ideology from DIPI. The core design uses CEO successions with the inventor team held fixed. Does the firm’s share of green, weapons and contested-AI patents move toward the new leader’s politics? Do misaligned inventors leave or switch domains? For AI, map each patent to the occupations whose tasks it performs (Webb, 2020; Autor et al., 2024) and weight by their partisanship. Do inventors and firms automate out-partisans’ work more readily?
Mechanisms
Managerial discretion over the R&D portfolio set against bottom-up idea generation; misfit exit; and, for incidence, out-group discounting (Iyengar et al., 2019).
Towards causality
Stacked difference-in-differences around exogenous CEO departures. Inventors who move between firms whose leaders differ politically. For AI subdomains whose politicisation can be dated, event studies of alignment as the coding emerges: facial recognition after the 2020 moratoria, military AI after the 2026 disputes over defence contracts.
First paper
Who directs the direction? Leader and inventor politics in the firm’s technological portfolio. This is the upper-echelons contribution that Dossi & Morando (2026) leave open, and a natural home for P1’s incongruence argument. Outlets: Strategic Management Journal, Academy of Management Journal, Research Policy. The AI-incidence question is a second, riskier paper.

5.4 Seam 4: Technology and the political composition of firms

Question
Does adopting automation or AI change who works in a firm, and so its political composition? Does that change feed back into the firm’s politics: its public stands, its PAC giving, its lobbying on AI and its internal incongruence?
Builds on
Trade shocks and polarisation (Autor et al., 2020); robots and the radical right (Anelli, Colantone & Stanig, 2021; Frey, Berger & Chen, 2018); the declining middle (Kurer, 2020); theories of ideological skew (Red, blue, and purple employee populations, 2025); the political transformation of corporate America (Steel, 2025); executive sorting (Fos, Kempf & Tsoutsoura, 2022).
What is missing
In political science, technology is a shock to voters and regions. In organisation research, it plays no part at all as a driver of political composition. The theory of how organisations become red, blue or purple (Red, blue, and purple employee populations, 2025) runs through founding conditions and attraction, selection and attrition, with no role for the technology of production.
What POLISY adds
A flaw of VRscores for measuring political change becomes an asset here. Party is fixed at 2024, so a change in an employer’s score is a change in who works there, which is exactly the channel of interest. The analysis should be restricted to registration states, where drift is not an artefact.
Rival explanations
Growth and relocation; cohort replacement unrelated to technology; mergers and restructuring; the party-data regime.
Data and design
The VRscores employer-year panel, 2012–2024. Occupation flows from Revelio. Adoption timing from first AI postings and first AI patents. FEC PAC data, LobbyView and CEO statements for the firm’s politics. The design compares a shift-share prediction of composition change (initial occupation shares × exposure × post) with the actual change.
Mechanisms
Occupational substitution: clerical and administrative work, which is disproportionately female and Democratic-leaning, against engineering hires. Also the relocation of work, and attrition of workers who object.
Towards causality
An exposure-based shift-share after 2022, with 2016–2022 as a placebo window. Earlier automation waves (robots, software) should shift composition the other way, which gives a sign test.
First paper
Automation and the political composition of the firm. The most novel of the four, and the longest data build. Outlets: Administrative Science Quarterly, Academy of Management Journal, Organization Science.
Table 4. The four seams compared (as of September 2026)
Seam How new Closest work Feasible with data in hand Main risk
1. Structural vs elective partisanship Extends segregation research into management constructs Frake, Hurst & Kagan (2026); Castiglia (2025) High, once firm occupation mix is added Read as a methods note unless the reanalysis changes a published result
2. Who decides, who is exposed Open Chu, Li & Zhu (2026); Bloom & Makridis (2026) Medium: needs AI-adoption data Others extend CEO-level work to the workforce first
3. Who directs the direction Answered for inventors by Dossi & Morando (2026); open for leaders and for AI coded by incidence Dossi & Morando (2026); Fehder et al. (2024); Engelberg et al. (2025) Medium: replicate the four-state voter-file match; DIPI for leaders Read as Dossi & Morando plus CEOs unless leader turnover identifies the contest
4. Technology → political composition Open; new theory Red, blue, and purple employee populations (2025); Autor et al. (2020) Medium: VRscores plus occupation flows Short window since 2022

What not to pursue. Migration and AI is not a POLISY paper: the field is crowded, the exposure measure is a proxy, and the IRS files cannot see who moves. Use migration only as a source of variation. For example, remote-work inflows of Democratic-leaning professionals into Sun Belt metros shift the partisanship of local labour pools, a possible shock to firm composition in Seam 1. Inflows also bring education, so the design must hold it equal.

6. Evidence still needed

The first run was limited by what it could link, not by method. The data below move POLISY to the firm level, roughly in order of payoff.

Table 5. Data that would move POLISY to the level of the firm
Data Unlocks Access Priority
VRscores employer file matched to Compustat (GVKEY) Seams 1, 2, 4; P1 In hand; validated name matching, LobbyView names, or the authors’ crosswalk (routes in your VRscores report) 1
Firm × occupation × location headcounts (Revelio positions) Seam 1 baseline; Seam 2 exposure; Seam 4 flows WRDS (check the Groningen licence) or a Revelio academic licence 1
DIPI: CEO, top-team and board ideology Seams 2, 3; P1 In hand 1
Firm AI adoption: AI job-posting shares; USPTO AI Patent Dataset; AI mentions in 10-Ks (EDGAR full text) and earnings calls Seams 2, 4 Postings via WRDS or Lightcast; EDGAR and USPTO free; transcripts via WRDS 1
PatentsView, with patents mapped to the occupations whose tasks they perform (Webb, 2020; Autor et al., 2024) and firms linked through DISCERN Seam 3 Free; compute and validation 2
Inventor partisanship: closed-primary voter files (Florida, New Jersey, New York, Pennsylvania) matched on name and city, as in Dossi & Morando (2026); donations (DIME) for other states Seam 3; P1; P2 State voter files are public or low-cost; DIME is free 2
Occupation × metro partisanship Holds geography equal in Seam 1 and in Section 3.2 Ask the VRscores authors; or L2 with Revelio 2
Worker-level party, occupation and AI use (Cooperative Election Study; Gallup Workforce Panel) Mechanisms for Seam 2 CES free; Gallup restricted 3
Voice and exit: NLRB petitions, WARN notices, Glassdoor reviews Outcomes for Seams 2 and 4 Free, except Glassdoor 3
State and city AI employment laws (NCSL database) Shocks and moderators for Seam 2 Free 3
Census BTOS AI use by sector and state Exposure against actual use Free 3
Official county controls (Census Gazetteer, ACS, USDA codes) Replace the compiled county file used in Section 3.1 Free 3

7. From findings to papers

A sequence that builds each step on the last:

  1. Now: a short research note on the partisan face of AI exposure (Section 3.2), for Research & Politics, Socius or PNAS Nexus. It needs no new data, and its contribution is the asymmetry between waves and the measure-dependence. Check its overlap with Bloom & Makridis (2026) before writing.
  2. Next six to nine months: Seam 1. It is the foundation for P1 and for Seams 2 and 4, and it needs one data addition, firm occupation mix.
  3. Then Seam 2 or Seam 4. Both use the same firm panel; Seam 2 extends P1 into AI, and Seam 4 has the stronger claim to a new theory.
  4. In parallel with P2: Seam 3. Start by replicating the four-state inventor–voter match of Dossi & Morando (2026), which P1 and P2 need anyway. Then ask the question they leave open: whether leaders or inventors set the firm’s direction. It shares P2’s patent infrastructure and its claim about direction.

8. Data and methods

Sources. VRscores occupation, industry, metro and employer summaries from the VRscores report (Kagan, Frake & Hurst, 2026); AIOE, AIIE and AIGE (Felten, Raj & Seamans, 2021); DAIOE v1.0.0 and its SOC 2018 comparison panel (Engberg et al., 2026), which carries Eloundou et al. (2024), Webb (2020) and Frey & Osborne (2017); IRS SOI county migration files, 2012–13 to 2021–22; the Correlates of State Policy; telework shares from Dingel & Neiman (2020); county context from a public compilation of ACS 2019, BEA, NOAA and County Business Patterns data (JsonOfCounties, 2023); and county presidential returns for 2016–2024 compiled by McGovern (2024). No synthetic data were used anywhere.

Estimation. All models are weighted least squares, with every continuous regressor z-scored, so coefficients read as effects per standard deviation. Migration models use state-by-year fixed effects with errors clustered by county. Occupation models use HC1 errors and, where stated, occupation-group fixed effects. Link rates for every join are in the interactive lab’s data tab.

Limits. County controls come from a compiled source; official files should replace them before publication. Telework is measured through each county’s sector mix, not its occupations. Occupation partisanship is national. VRscores covers workers with online profiles who are registered to vote.

Reproducing. The Colab notebook in the repository runs every step. Figures and tables here are regenerated from results.json at every build; the text is not.

Appendix A. All findings

Table 6. All findings of the first run, in the lab's reading order.
FindingWhat the data showQuestion it raises
The 'AI exodus' is mostly the partisan and telework geography of migrationstress-mig · county · RobustIn 2020-22, one SD more AI exposure meant -0.86 points of households a year within the same state (t = -13.6). With density, education, income, housing costs, January temperature, the 2016 vote and telework held equal it is -0.11 (t = -2.2): 88% of the gradient goes. Year by year the controlled AI coefficient stays small and flat while the unadjusted one grows. The strongest predictor in the full model is the county's 2016 Republican vote share (+0.74, t = 15.4); the coefficient on the 2016 vote grew from +0.33 (2013-16) to +0.74 (2020-22).
  • County context is a 2019 cross-section (ACS 2019, CBP); telework comes from the county's industry mix, not its occupations.
  • The 2016 vote is fixed before the flows it predicts, but it stands for everything that differs between red and blue counties.
Once AI exposure is read as a proxy for dense, Democratic, telework-intensive counties, the open question is the political one: do movers carry their workplace politics into destination firms?
AI exposure's Democratic lean is composition; automation risk's Republican lean is notstress-occ · occupation · RobustGenerative-AI exposure (DAIOE) leans Democratic on its own (-2.0 points per SD, t = -3.3, 825 occupations), but the lean disappears once gender and race are held equal (-0.1, t = -0.1) and stays near zero with occupation-group effects (+0.1, t = 0.1). Computerisation risk (Frey & Osborne) goes the other way: +0.9 (t = 1.2) alone, +2.4 (t = 4.0) with education, pay, telework, gender, race and occupation group held equal. Measures whose sign flips and becomes clear (|t| of 2 or more) with every control: AIOE (Felten, Raj & Seamans).
  • Gender and race shares exist for about 300 occupations, so the last two models use a smaller sample.
  • Occupation partisanship is national; occupations concentrated in Republican regions look Republican for that reason alone.
If exposure's partisan face is who holds the jobs (educated women, teleworkers), does political conflict over AI inside firms run along occupational and gender lines rather than party lines?
Most of an industry's partisanship, and half of its political balance, is the occupations it employsstress-structure · industry · RobustAcross 178 industries, the Republican share predicted from national occupation shares and each industry's staffing pattern explains 72% of the variance in the actual share (r = 0.85, weighted by workers) and 54% of the variance in political balance (how evenly split the workforce is; r = 0.74). Workforce 'ideology' and 'balance' measured at the organisation level therefore mix the occupational structure of the work with any organisational sorting.
  • National occupation partisanship is applied to every industry, so the prediction ignores where industries are located.
Does the part of workforce partisanship and balance that occupations do not explain (the elective part) predict innovation, where the structural part does not?
Households are leaving AI-exposed counties, but the gradient is mostly density, politics and teleworkmig-ai-exodus · county · FragileAcross 3,130 counties, the correlation between AI exposure (AIGE) and net domestic migration went from -0.19 (2012-13) to -0.51 (2020-21), weighted by households. The most exposed fifth of counties lost 0.16% of households a year at the start and 0.80% at the worst. Comparing counties within the same state, one SD more exposure means -0.16 (t = -3.9) points a year in 2013-16, -0.56 (t = -11.1) in 2017-19 and -0.86 (t = -13.6) in 2020-22. Biggest losers 2020-22 among large counties: San Francisco County, CA (-3.9%); Suffolk County, MA (-3.7%); New York County, NY (-3.7%); Bronx County, NY (-3.4%); Kings County, NY (-2.7%). Exposed counties still gaining: Collin County, TX (+2.4%); Chester County, PA (+0.6%); Polk County, IA (+0.3%).
  • AIGE describes the county's 2019 jobs, not the movers themselves.
  • IRS counts tax returns (households); year = the second year of each filing pair.
  • Stress test (stress-mig): with county controls the 2020-22 gradient falls from -0.86 to -0.11 points per SD; AIGE mostly stands for dense, Democratic, telework-intensive counties.
Is this remote work and housing costs pulling knowledge workers out of dense cores, and does it move AI-exposed work, and its politics, into less exposed and more Republican places?
AI exposure measures disagree about who is exposed: benchmark and language-model measures lean Democratic, patent and automation measures Republicanocc-waves · occupation · FragileAcross 825 occupations (weighted by workers), the correlation with the Republican share of workers runs from -0.18 (DAIOE, generative AI) to +0.23 (AI patents (Webb)). Weighting occupations equally, the benchmark and GPT measures sit near -0.32, patent and automation measures near +0.25. Holding education and wage equal, measures with a clear Republican lean (|t| of 2 or more): Computerisation risk (Frey & Osborne); clear Democratic lean: DAIOE, generative AI.
  • VRscores occupations are linked to SOC 2018 by title, then code; 12% of workers are not linked.
  • Webb's and Frey-Osborne's scores come from older SOC versions mapped in the DAIOE panel.
  • Stress test (stress-occ): the Democratic lean of generative-AI and language-model measures disappears once gender and race are held equal; the Republican lean of computerisation risk survives every control.
Which measure predicts what actually happens to workers (employment, wages, hours) since 2022, and do exposed occupations shift politically, as robot exposure did after 2016?
Interstate movers go to more Republican states, and the gap widened after 2019mig-partisan-state · state · RobustWeighted by households, the average interstate move ends in a state whose last presidential vote was +0.8 points more Republican than the origin in 2013, +2.5 at the peak (2021) and +2.2 in 2022; weighted by the income moved, +1.4 to +4.2. Moves from 2017 on hold the map at the 2016 election, so the widening comes from where people moved, not from how states voted.
  • CSPP's presidential share is the state's result in the last election before each year (through the 2016 election).
  • Flows are county-to-county pairs summed to states (about half of interstate moves).
Is partisan re-sorting by migration large enough to move state electorates, and is it driven by policy, housing or jobs?
AI exposure looks Democratic across occupations, but the link runs through educationocc-ai-education · occupation · RobustAcross 698 occupations, more AI-exposed ones have more Democratic workforces (r = -0.30); weighted by workers the link nearly vanishes (r = -0.03, 95% CI -0.15 to +0.08). With wage and education held equal, exposure's coefficient is -0.2 points per SD (t = -0.3); adding gender, race and occupation group gives +3.3 (t = +2.8).
  • Demographic controls exist for fewer occupations, so the full model has a smaller sample.
  • VRscores partisanship is party identity from voter files, not ideology.
  • Stress test (stress-occ): gender and race account for the rest of the lean.
Will generative AI's labor-market effects land on educated, Democratic-leaning workers, and does exposure change political attitudes or sorting?
Households left states with liberal economic policy and AI-exposed jobs; at the state level the two are hard to separatestate-environment · state · SuggestiveAcross states, net interstate migration (IRS, 2012-13 to 2021-22, weighted by households) correlates most negatively with Economic policy liberalism (Caughey and Warshaw) (-0.75); Income per capita ($) (-0.64); Public economic liberalism (Caughey and Warshaw) (-0.63) and most positively with Right-to-work law (1 = yes) (+0.62); Republican share of state legislators (%) (+0.61). Entered together, economic policy liberalism gives -0.42 (t = -5.5) points a year per SD and AIGE -0.05 (t = -0.5): the two move together across states. Within states, where policy is constant, county exposure still predicts out-migration (see the county finding).
  • CSPP political variables mostly end in 2014-2017; the profile averages 2012-2016.
  • 51 states: correlations, not causal estimates.
Which part is policy (taxes, housing regulation, labour law) and which is the kind of work a state does? Border-county comparisons across state lines can separate them.
AI-exposed counties lose richer households than they gainmig-income · county · DescriptiveIn the most exposed fifth of counties, households moving out report more income than those moving in every year: -3.6 thousand dollars of AGI per return in 2013 and -7.5 in 2022. In the least exposed fifth the gap is +0.6 and +16.1. The correlation of exposure with the income gap is -0.36 in 2022.
  • AIGE stands for dense, Democratic, telework-intensive counties (see stress-mig); read as a description, not an AI effect.
Do high earners in AI-exposed jobs carry their (remote) work with them, spreading AI-exposed employment to new places?
Education pulls occupations Democratic and pay pulls them Republican; the pay effect runs through who does the jobocc-education-vs-pay · occupation · RobustWith both in the model, one SD more required education goes with -8.1 points Republican share (t = -9.3) and one SD higher pay with +6.1 (t = +7.6), 614 occupations. Adding gender and race shares and occupation group leaves education at -4.6 (t = -3.0) and pay at +0.6 (t = +0.5); the female share is the strongest single predictor (-7.0 points per SD, t = -7.0).Is AI's partisan footprint just the education realignment, or does exposure add something within education levels?
Partisans are sorting into different employers while sorting across industries and occupations declinessorting · employer · SuggestiveSame-party over-exposure (colleagues share one's party more often than by chance), first to last year: employer: 8.43% -> 8.71%; industry: 2.74% -> 2.54%; metro: 7.36% -> 7.46%; occupation: 3.34% -> 3.08%. Rising: employer, metro; falling: industry, occupation. Employer sorting rose +0.28 points, metro sorting +0.10: most of the growth in workplace segregation happens between firms within the same kinds of jobs, not across occupations or industries.Which organisational practices (hiring networks, location choice, mission statements, leadership politics) produce between-firm sorting?
AI-exposed industries have more Democratic workforces, mostly because their workers are highly educated; language-model exposure keeps a modest link within sectorsind-ai-partisanship · industry · FragileAcross 246 four-digit industries, AI exposure correlates -0.34 with the Republican share of the matched workforce (weighted; 95% CI -0.53 to -0.16). Language-model exposure is stronger (r = -0.43). Exposure tracks education closely, and with education and wages held equal the coefficient falls to -0.9 points per SD (t = -1.2); within sectors it is -1.7 (t = -1.9) and -2.5 for language-model exposure (t = -2.2).Is there an industry-level (firm culture, geography) component of AI exposure's partisan footprint beyond the education of the workforce?
AI and big-tech workforces barely joined the Democratic drift, and several moved Republicanemp-bigtech · employer · SuggestiveThe 34 large AI and big-tech employers (identified by name) changed -1.1 points on average, against -4.1 for all others. Moving Republican: Snowflake, Inc. (+11.9); Broadcom Inc. (+7.0); Uber Technologies, Inc. (+4.7); Amazon Web Services, Inc. (+4.4); Tesla, Inc. (+3.8); Palantir Technologies, Inc. (+3.3).Did AI and cloud firms' expansion (data centres, new hubs in Texas, Virginia, Utah) or the 2022-24 layoffs shift their workforces' politics?
Where party is inferred rather than registered, metro workforces drift Democratic much fastermetro-regime · metro · RobustAcross 359 metros, 87% moved towards the Democrats between 2012 and 2024. Metros in states where party comes from primaries or L2's model drifted 1.8 points further than metros in registration states, holding the starting level and size equal (t = -5.6). The drift is a trait of the measurement regime, not only of places.How much of VRscores' Democratic drift is cohort replacement, and how much is the inference model aging with voters?
Occupational mix explains most of an industry's partisanship; the rest points to place and missionind-composition · industry · RobustPredicting each industry's Republican share from national occupation shares and its staffing pattern gives r = +0.85 with the actual share (178 industries with 70%+ of jobs linked). The remaining gap has an SD of 4.2 points (actual: 7.3). Most Democratic relative to their occupations: Footwear Manufacturing, Local Messengers and Local Delivery, Museums, Historical Sites, and Similar Institutions; most Republican: Merchant Wholesalers, Durable Goods, Other Wood Product Manufacturing, Rental and Leasing Services. The gap is unrelated to AI exposure once education is held equal (t = -0.1).How much of the residual is geography (where the industry sits) versus organisational culture or mission?
Two kinds of AI exposure: language-AI jobs lean Democratic, perception-AI jobs Republican, until you control for educationocc-two-ais · occupation · FragileBeyond overall exposure, occupations differ in which AI touches them: the second principal component contrasts image recognition and games with language modelling, translation and reading. It correlates +0.22 with Republican share (weighted), but with wage, education, gender, race and occupation group it is -0.6 points per SD (t = -0.7).Does the generative-AI wave (language) shift the politics of exposure towards Democratic-leaning professions compared with automation and vision AI?
Partisan drift clusters in space, along state linesmetro-space · metro · RobustMetro drift is spatially autocorrelated (Moran's I = 0.24, permutation p = 0.001; level: I = 0.37). Clusters moving least towards the Democrats (or towards the Republicans): Rome GA MSA (+7.0); Dalton GA MSA (+5.2); Gainesville GA MSA (+4.8); Cleveland TN MSA (+2.4); Brunswick GA MSA (+5.8). Clusters moving most: Columbia MO MSA (-13.2); Springfield MO MSA (-10.1); Jefferson City MO MSA (-7.2); Grand Rapids-Wyoming MI MSA (-8.0); Lansing-East Lansing MI MSA (-6.8). Clusters line up with states' party-data regimes.Are the Georgia and Missouri clusters real political change or artefacts of primary-based and modelled party?
AI-exposed states have more Democratic workforces, and VRscores covers them more denselystate-aige · state · SuggestiveState AI exposure (AIGE) correlates -0.63 with the Republican share of the matched metro workforce in 2024 (51 states, weighted by jobs). The same states' Republican vote share correlates -0.72 with AIGE. Coverage (matched workers per job) correlates +0.60 with AIGE: the data lean towards AI-exposed labour markets.Is the partisan gap in AI exposure a feature of places or of the people the data happen to cover?
Dynamic exposure mostly rises for everyone; where it rose most, workforces are more Democraticocc-daioe · occupation · DescriptiveDAIOE grows about 51-fold between 2012 and 2024 for every occupation alike (levels in 2010 and 2024 correlate 0.98), so the change is mostly a shared trend. Relative standing moved a little: occupations that gained standing since 2012 have more Democratic workforces (r = -0.26, 95% CI -0.44 to -0.09, weighted by workers), led by Telemarketers, Technical Writers, Historians, Door-to-Door Sales Workers, Sales and Related Workers. The correlation of standing with the Republican share drifted from -0.12 (2010) to -0.17 (2024).
  • Partisanship is measured once (2024), so this is a cross-section of changes, not a panel.
Does a rise in an occupation's exposure precede changes in who works in it (entry by younger, more Democratic cohorts)?
Movers trade down in AI exposure, and more so since 2020mig-exposure-gap · county · DescriptiveWeighted by households, the average county-to-county move ends in a county 0.020 SD of AIGE less exposed than its origin in 2013 and 0.073 SD in 2022 (largest: 0.096 in 2021); weighted by income moved it is 0.204 at its largest.
  • County pairs with fewer than 20 returns are suppressed by the IRS, so small flows are missing.
  • AIGE stands for dense, Democratic, telework-intensive counties (see stress-mig); read as a description, not an AI effect.
Does the down-trading in exposure change the local AI adoption and politics of destination counties?
States where party is inferred rather than registered drift Democratic fasterstate-regime · state · SuggestiveHolding the starting level equal, states whose party is inferred (primary participation or L2's model) drifted 1.9 points further towards the Democrats between 2012 and 2024 (t = -3.9, 51 states).Validate inferred party against registration in states that switched regime, or against survey self-reports.
Nearly every industry drifted Democratic, and AI exposure does not predict by how muchind-drift · industry · Robust98% of 246 industries moved towards the Democrats between the first and last year (weighted mean -3.6 points). Exposure does not predict the size of the move (t = -0.6); industries that started more Republican moved more (t = -2.0), consistent with party being fixed at 2024 while younger cohorts enter.Separate cohort replacement from real change: does drift vanish once workforce age is held constant?
Large employers drift Democratic eight times as often as Republican; the most Republican and the largest drift mostemp-drift · employer · RobustOf 6,000 large employers present every year, 2,127 moved 5+ points towards the Democrats and 269 towards the Republicans (ratio 7.9 to 1). Drift is larger for more Republican employers (t = -13.4) and for larger ones (t = -11.1).Is drift hiring (new cohorts), attrition, or relocation? Which organisational events precede it?
Which state policy variables move with AI outcomes within states over timepolicy-twfe · state · Suggestive1 of 23 policy-outcome pairs pass a 10% false-discovery threshold. Strongest: Median household income ($) -> net interstate migration (t = -4.2).Use policy adoptions as events (staggered difference-in-differences) for the variables that pass.
The US migration network in the latest yearmig-network · state · Descriptive2021-22: the largest net gainers of households were FL, TX, NC, SC, the largest net losers CA, NY, IL, MA. Biggest corridors: CA -> TX (44,415 returns); NY -> NJ (44,325 returns); NY -> FL (43,518 returns); CA -> AZ (30,784 returns); CA -> NV (28,869 returns). The network splits into 5 communities: AK, AR, AZ, CA, CO, HI, ID, LA, MT, NM, NV, OK, OR, TX, UT, WA, WY | IA, IL, IN, KS, KY, MI, MN, MO, ND, NE, OH, SD, WI | CT, DE, FL, GA, MA, ME, NH, NJ, NY, PA, RI, VT | DC, MD, NC, SC, VA, WV | AL, MS, TN.
  • Corridors come from county-to-county pairs, which show about half of interstate moves; net totals use all moves.
Do migration communities follow political regions, and do AI hubs sit at their centre or their edge?
State archetypes across politics, AI and innovationlandscape · state · Descriptive22 state measures (Workforce Republican share, Republican vote share, AI exposure, Net migration, Gross migration, Movers' income gap, Republican share of the last presidential vote (%), Republican share of state legislators (%)...) reduce to components explaining 50% and 12% of the variation. 3 archetypes: type 1: AL, AK, AZ, AR, FL, GA, ID, IN, IA, KS, KY, LA, MI, MS, MO, MT, NE, NV, NM, NC, ND, OH, OK, SC, SD, TN, TX, UT, WV, WI, WY; type 2: CA, CO, CT, DE, HI, IL, ME, MD, MA, MN, NH, NJ, NY, OR, PA, RI, VT, VA, WA; type 3: DC. Most unusual profiles (isolation forest): DC, WY, MA, VT, HI.Do archetypes predict how states respond to AI (adoption, invention, policy)?
State maps built from VRscores metros misplace workers of multi-state metrosstate-coverage · state · ArtefactMatched workers per job range from 1.3% to 77.5%. The lowest: NJ 1.3%, DE 1.9%, NH 3.6%, WY 4.7%, where large metros are named after a neighbouring state first (New York-Newark, Philadelphia-Wilmington, Boston-Nashua) and are credited to that state; the District of Columbia absorbs its suburbs the same way.Rebuild state figures by splitting metro workers across states in proportion to county jobs.
Occupations far more partisan than their exposure, pay, education and demographics predictocc-anomalies · occupation · DescriptiveMore Republican than predicted: Manicurists and Pedicurists (+16); Clergy (+14); Private Detectives and Investigators (+14); Interior Designers (+13); Merchandise Displayers and Window Trimmers (+12). More Democratic: Ushers, Lobby Attendants, and Ticket Takers (-17); Landscape Architects (-15); Natural Sciences Managers (-13); Pharmacy Technicians (-13); Financial Quantitative Analysts (-13).What do the outliers share (mission, religion, public-facing work, organisational setting)?
The largest employer drifts look like artefacts of restructured firmsemp-artifacts · employer · ArtefactThe twelve largest absolute drifts: Qualtrics International, Inc. (-46); MONAT Global Corp. (-42); Newfold Digital, Inc. (-38); Hopebridge LLC (-34); Anthology, Inc. (-31); Paparazzi Accessories (-30); Project 1972, Inc. (-29); Rock Central LLC (-28); Haleon Plc (-28); BAMTech LLC (-27); Piper Sandler Cos. (-27); M Health Fairview Center For Women (-26). Many are spin-offs, mergers or rebrands after 2020, whose early years are rebuilt from today's profiles.Flag restructured employers before using drift as an outcome.
County movers go to more Republican counties than they leavemig-partisan · county · SuggestiveWeighted by households, the average destination county is +2.05 points more Republican than the origin (years 2016-2022, latest +2.27); weighted by income moved, +3.48.Is this political sorting, housing costs, or jobs? Do AI-exposed workers move differently?
Actual AI adoption by firms (BTOS)adoption · sector/state · Needs dataNot run: needs Census BTOS data downloads.Where is AI actually used, compared with where it could be (exposure)?
AI patenting: growth, geography and technologypatents · patent/county/state · Needs dataNot run: needs PatentsView g_patent, g_cpc_current, g_inventor_disambiguated, g_location_disambiguated.Where is AI invented, how fast is it spreading, and does it follow the political map?

Text as generated by the lab from the data.

Appendix B. Literature to add to Gyaanpatrika

These items are not in the workbook and bear directly on the seams above. Each reference was checked online in September 2026; items marked in the reference list still need a detail verified.

Table 7. Additions to the Literature tab, by stream
Reference Stream Project Why it matters
Frake, Hurst & Kagan (2026) Workplace politics & ideological fit P1 Workplace political segregation net of occupation; the benchmark for Seam 1
Red, blue, and purple employee populations (2025) Workplace politics & ideological fit P1 Theory of how organisations become ideologically skewed; Seam 4 adds technology to it
Colonnelli, Pinho Neto & Teso (2025) Workplace politics & ideological fit P1 Owners hire co-partisans; causal evidence on sorting
Chu, Li & Zhu (2026) Leader & firm political ideology P1 CEO partisanship and AI hiring; closest competitor for Seam 2
Bloom & Makridis (2026) Leader & firm political ideology P1 Partisan gap in AI use is composition; Section 3.2 extends it to exposure
Muro et al. (2026) Polarisation & party systems General AI exposure mapped on county votes
Grant, Green & Evans (2026) Polarisation & party systems General Occupational AI exposure raises support for AI regulation
Haslberger, Gingrich & Bhatia (2025) Polarisation & party systems General Experimental exposure to generative AI and policy preferences
Antoniades et al. (2025) Polarisation & party systems General Electoral effects of local AI adoption
Magistro et al. (2026) Political psychology & neuroscience of ideology Micro leg Common attitudes toward AI and globalisation
Chueri & Törnberg (2026) Polarisation & party systems General How parties frame AI and work in 33 parliaments
Dossi & Morando (2026) Innovation measurement & science of science P2 Inventors’ party shapes what they invent and cite; the benchmark Seam 3 must go beyond
Chen & Gong (2025) Innovation search, recombination & networks P2 Collaboration between Democratic and Republican inventors fell after 2016
Engelberg et al. (2025) Innovation measurement & science of science P2 Political sentiment and patenting; partisans cluster by technology
Felten, Raj & Seamans (2021) Innovation measurement & science of science General AIOE, AIIE and AIGE; in SMJ, so the exposure measure management readers know
Webb (2020) Innovation measurement & science of science P2 Patent-based exposure; AI targets high-skill tasks
Eloundou et al. (2024) Innovation measurement & science of science General Language-model exposure
Autor et al. (2024) Innovation search, recombination & networks P2 Links patents to the occupations they affect; the incidence measure for Seam 3
Acemoglu & Restrepo (2020) Comparative political economy of innovation P2 The direction of AI as a choice
Kurer (2020) Polarisation & party systems General Routine workers, status decline and the populist right
Anelli, Colantone & Stanig (2021) Polarisation & party systems General Individual robot exposure and radical-right support
Gallego & Kurer (2022) Polarisation & party systems General Review: automation, AI and political behaviour
Autor et al. (2020) Comparative political economy of innovation General Trade shocks and polarisation; the template for Seam 4
Gethin, Martínez-Toledano & Piketty (2022) Polarisation & party systems General Education left, income right
Dingel & Neiman (2020) Methods & identification General Telework shares, a standard control for AI exposure

References

  1. Acemoglu, D., & Restrepo, P. (2020). The wrong kind of AI? Artificial intelligence and the future of labour demand. Cambridge Journal of Regions, Economy and Society, 13(1), 25–35. doi.org/10.1093/cjres/rsz022
  2. Anelli, M., Colantone, I., & Stanig, P. (2021). Individual vulnerability to industrial robot adoption increases support for the radical right. Proceedings of the National Academy of Sciences, 118(47), e2111611118. doi.org/10.1073/pnas.2111611118
  3. Antoniades, A., Balcazar, C. F., Chatzikonstantinou, M., & Kern, A. (2025). The electoral consequences of AI adoption: Evidence from US elections (SSRN Working Paper 5975634). papers.ssrn.com/sol3/papers.cfm?abstract_id=5975634
  4. Autor, D., Chin, C., Salomons, A., & Seegmiller, B. (2024). New frontiers: The origins and content of new work, 1940–2018. Quarterly Journal of Economics, 139(3), 1399–1465. doi.org/10.1093/qje/qjae008
  5. Autor, D., Dorn, D., Hanson, G., & Majlesi, K. (2020). Importing political polarization? The electoral consequences of rising trade exposure. American Economic Review, 110(10), 3139–3183. doi.org/10.1257/aer.20170011
  6. Barber, B., IV, & Blake, D. J. (2024). My kind of people: Political polarization, ideology, and firm location. Strategic Management Journal, 45, 849–874. sms.onlinelibrary.wiley.com/doi/full/10.1002/smj.3572
  7. Bermiss, Y. S., & McDonald, R. (2018). Ideological misfit? Political affiliation and employee departure in the private-equity industry. Academy of Management Journal, 61(6), 2182–2209. journals.aom.org/doi/10.5465/amj.2016.0817
  8. Bloom, N., & Makridis, C. (2026). The politics of AI (NBER Working Paper 34813). www.nber.org/papers/w34813
  9. Castiglia, L. (2025). Striking a political balance? Political polarization and firm innovation. Journal of Management. journals.sagepub.com/doi/10.1177/01492063251389882
  10. Chen, T., & Gong, R. K. (2025). Drawing partisan lines in innovation: Political polarization and U.S. inventor collaboration (SSRN Working Paper 5666755). papers.ssrn.com/sol3/papers.cfm?abstract_id=5666755 [year not checked]
  11. Christensen, D. M., Dhaliwal, D. S., Boivie, S., & Graffin, S. D. (2015). Top management conservatism and corporate risk strategies: Evidence from managers' personal political orientation and corporate tax avoidance. Strategic Management Journal, 36(12), 1918–1938.
  12. Chu, Y., Li, A., & Zhu, R. (2026). Politically coded: CEO partisanship and firm AI adoption (SSRN Working Paper 6427178). papers.ssrn.com/sol3/papers.cfm?abstract_id=6427178
  13. Chueri, J., & Törnberg, P. (2026). Meeting the coming wave: The emerging politics of AI and work across 33 parliaments (arXiv:2609.02296). arxiv.org/abs/2609.02296
  14. Colonnelli, E., Pinho Neto, V., & Teso, E. (2025). Politics at work. American Economic Review, 115(10), 3367–3414. www.aeaweb.org/articles?id=10.1257/aer.20240151
  15. Dingel, J. I., & Neiman, B. (2020). How many jobs can be done at home? Journal of Public Economics, 189, 104235. doi.org/10.1016/j.jpubeco.2020.104235
  16. Dossi, G., & Morando, M. (2026). Polarized technologies (Working paper, version of 13 September 2026; earlier versions as CEP Discussion Paper 1969, 2023, and RFBerlin Discussion Paper 064/26, March 2026). www.rfberlin.com/wp-content/uploads/2026/03/26064.pdf
  17. Economic Innovation Group (2023). Tax data reveals large flight of high earners from major cities during the pandemic. eig.org/high-earners-migration/
  18. Edwards, J. R., & Cable, D. M. (2009). The value of value congruence. Journal of Applied Psychology, 94(3), 654–677. doi.org/10.1037/a0014891
  19. Eloundou, T., Manning, S., Mishkin, P., & Rock, D. (2024). GPTs are GPTs: Labor market impact potential of LLMs. Science, 384(6702), 1306–1308. doi.org/10.1126/science.adj0998
  20. Engberg, Görg, Hellsten, Javed, Lodefalk, Längkvist, Monteiro, Kyvik Nordås, Pulito, Schroeder, & Tang (2026). AI unboxed: Capability arrival and the clerical decline. Data: DAIOE v1.0.0, Zenodo record 21873968. zenodo.org/records/21873968 [initials not checked]
  21. Engelberg, J., Lu, R., Mullins, W., & Townsend, R. (2025). Political sentiment and innovation: Evidence from patenters. Review of Financial Studies, 38(9), 2718–. academic.oup.com/rfs/article/38/9/2718/8121164
  22. Fehder, D. C., Teodoridis, F., Raffiee, J., & Lu, J. (2024). The partisanship of American inventors. Research Policy, 53(7), 105034. doi.org/10.1016/j.respol.2024.105034
  23. Felten, E., Raj, M., & Seamans, R. (2021). Occupational, industry, and geographic exposure to artificial intelligence: A novel dataset and its potential uses. Strategic Management Journal, 42(12), 2195–2217. doi.org/10.1002/smj.3286
  24. Fos, V., Kempf, E., & Tsoutsoura, M. (2022). The political polarization of corporate America (NBER Working Paper 30183; revised 2026). www.nber.org/papers/w30183
  25. Frake, J., Hurst, R., & Kagan, M. (2026). Political segregation in the US workplace. Nature Human Behaviour. www.nature.com/articles/s41562-026-02501-9
  26. Frey, C. B., & Osborne, M. A. (2017). The future of employment: How susceptible are jobs to computerisation? Technological Forecasting and Social Change, 114, 254–280. doi.org/10.1016/j.techfore.2016.08.019
  27. Frey, C. B., Berger, T., & Chen, C. (2018). Political machinery: Did robots swing the 2016 US presidential election? Oxford Review of Economic Policy, 34(3), 418–442. doi.org/10.1093/oxrep/gry007
  28. Gallego, A., & Kurer, T. (2022). Automation, digitalization, and artificial intelligence in the workplace: Implications for political behavior. Annual Review of Political Science, 25, 463–484. doi.org/10.1146/annurev-polisci-051120-104535
  29. Gambit, E. JsonOfCounties: county-level data compiled from the Census ACS, BEA, NOAA, County Business Patterns and the MIT Living Wage Calculator [Data set]. GitHub. github.com/evangambit/JsonOfCounties [compilation; year of last update not checked]
  30. Gethin, A., Martínez-Toledano, C., & Piketty, T. (2022). Brahmin left versus merchant right: Changing political cleavages in 21 Western democracies, 1948–2020. Quarterly Journal of Economics, 137(1), 1–48. doi.org/10.1093/qje/qjab036
  31. Gimpel, J. G., Newton, J., & Reeves, A. (2025). Short-haul moves and the political geography of partisanship: Intrametropolitan migration as a force for change in U.S. politics. Urban Affairs Review. doi.org/10.1177/10780874251380677
  32. Gmyrek, P., Berg, J., & Bescond, D. (2023). Generative AI and jobs: A global analysis of potential effects on job quantity and quality (ILO Working Paper 96). International Labour Organization.
  33. Grant, Z., Green, J., & Evans, G. (2026). New tech, new threat? Occupational exposure to artificial intelligence increases support for AI regulation. Journal of European Public Policy. www.tandfonline.com/doi/full/10.1080/13501763.2026.2710714
  34. Gupta, A., Briscoe, F., & Hambrick, D. C. (2017). Red, blue, and purple firms: Organizational political ideology and corporate social responsibility. Strategic Management Journal, 38(5), 1018–1040. doi.org/10.1002/smj.2550
  35. Haslberger, M., Gingrich, J., & Bhatia, J. (2025). Rage against the machine? Generative AI exposure, subjective risk, and policy preferences. Journal of European Public Policy. www.tandfonline.com/doi/full/10.1080/13501763.2025.2554903
  36. Iyengar, S., Lelkes, Y., Levendusky, M., Malhotra, N., & Westwood, S. J. (2019). The origins and consequences of affective polarization in the United States. Annual Review of Political Science, 22, 129–146. doi.org/10.1146/annurev-polisci-051117-073034
  37. Kagan, M., Frake, J., & Hurst, R. (2026). VRscores: A new measure and data set of workforce politics using voter registrations. Organization Science, 37(2), 444–465. ideas.repec.org/a/inm/ororsc/v37y2026i2p444-465.html
  38. Kiss, A. N., Yu, Q., Neville, F., & Ward, A. (2026). Breaking through? The divergent consequences of CEO political ideology on firm inventiveness. Journal of Management, 52(3), 1010–1037.
  39. Kurer, T. (2020). The declining middle: Occupational change, social status, and the populist right. Comparative Political Studies, 53(10–11), 1798–1835. doi.org/10.1177/0010414020912283
  40. Magistro, B., et al. (2026). Attitudes toward artificial intelligence (AI) and globalization: Common microfoundations and political implications. American Journal of Political Science. onlinelibrary.wiley.com/doi/10.1111/ajps.12959
  41. Mannor, M. J., & Busenbark, J. R. (2025). A donation-based indicator of political ideology (DIPI): An open dataset for studying the political ideologies of employees, top management teams, CEOs, boards, and industries. Organizational Behavior and Human Decision Processes, 188. ideas.repec.org/a/eee/jobhdp/v188y2025ics0749597825000317.html
  42. McGovern, T. US county-level presidential election results, 2008–2024 [Data set]. GitHub. github.com/tonmcg/US_County_Level_Election_Results_08-24
  43. McKee, S., Smith, D. A., & Zhang, O. (2025). “Welcome to the Free State of Florida”: In-migration and rising Republicanism in the Sunshine State. Political Science Quarterly. academic.oup.com/psq/advance-article/doi/10.1093/psquar/qqaf107/8341509
  44. Muro, M., et al. (2026). The political geography of AI exposure. Brookings Institution. www.brookings.edu/articles/ai-political-geography-worker-exposure/ [author list not verified]
  45. Ramani, A., & Bloom, N. (2021). The donut effect of COVID-19 on cities (NBER Working Paper 28876). www.nber.org/papers/w28876
  46. Reinwald, M., Kanitz, R., Bamberger, P., Backmann, J., & Hoegl, M. (2026). The elephant and donkey in the room: Time-varying effects of political dissimilarity on social interactions at work during US elections. Organization Science.
  47. Steel, R. S. (2025). The political transformation of corporate America, 2001–2022. American Political Science Review.
  48. The origins and evolution of red, blue, and purple employee populations: A theory of how organizations become ideologically skewed (2025). Academy of Management Review. doi.org/10.5465/amr.2025.0229 [authors not verified]
  49. Webb, M. (2020). The impact of artificial intelligence on the labor market (SSRN Working Paper 3482150). papers.ssrn.com/sol3/papers.cfm?abstract_id=3482150
  50. Wowak, A. J., Busenbark, J. R., & Hambrick, D. C. (2022). How do employees react when their CEO speaks out? Intra- and extra-firm implications of CEO sociopolitical activism. Administrative Science Quarterly, 67(2), 553–593.
  51. Yin, M., & Ogut, B. (2026). Who uses AI? Platforms, workforce, and AI exposure (arXiv:2605.21743). arxiv.org/abs/2605.21743