India AI Impact Summit Insights
All sessions

Global Dialogue on AI and Labour Market Resilience

Contents

Executive Summary

This AI summit convenes leading experts, policymakers, and industry leaders to address the critical intersection of AI advancement and labor market stability. The discussion reveals consensus around two urgent needs: (1) empirical data on AI adoption and employment impacts, and (2) proactive international coordination to manage transitions, particularly in the Global South, where formal labor market infrastructure is weaker but AI adoption is accelerating.

Key Takeaways

  1. The Measurement Crisis is Urgent: Without firm-level, task-level, and country-specific data on AI adoption and employment impact, policymakers are flying blind. Early data gaps mean responses may arrive too late.

  2. Global South Needs Tailored Approaches, Not US-Centric Models: India's DPI framework and Kenya's sector-focused strategy (food systems, SMEs) show that labor market resilience strategies must account for formal vs. informal economy splits, median ages, colonial labor legacies, and existing infrastructure gaps—not imported wholesale from wealthy economies.

  3. Resilience ≠ Predicting the Future; It = Planning for Scenarios: Policymakers should build social safety nets, education capacity, and institutional coordination before job displacement peaks, not wait for definitive forecasts. The conversation should shift from "Will 50% of jobs be automated?" to "What do we do if they are?"

  4. Data Sharing & International Coordination Are Not Optional: The countries/regions with the fastest adoption (Israel, Japan, India, Kenya) and largest AI companies must share usage and employment data in privacy-preserving formats. Without coordination, Global South countries face a compounding disadvantage.

  5. Power Dynamics Matter as Much as Technology: The concentration of AI capabilities in two countries, coupled with differential access and labor market strength, creates geopolitical and economic inequality risks that dwarf traditional automation concerns. Policy must be global and aligned with shared values.

Key Topics Covered

  • Employment Impact Evidence: Early empirical findings on AI exposure and actual job displacement, particularly for entry-level workers in the US
  • Global Inequality & Geopolitics: Concentration of advanced AI capabilities in two countries and resulting competitive advantages/disadvantages
  • Data Infrastructure Gaps: Need for better firm-level AI adoption data, usage metrics, and employment outcome tracking across countries
  • Regional Responses: India's digital public infrastructure (DPI) model, Kenya's labor market challenges and KAISA initiative, UK policy frameworks
  • Measurement Challenges: Distinction between AI exposure (technical feasibility) vs. adoption (actual implementation), and reasons for conflicting research results
  • Skills, Education & Transitions: Role of AI in education, reskilling program effectiveness, and social safety nets
  • Multilingual & Cultural Considerations: Ensuring AI systems reflect local contexts, languages, and cultural values
  • International Coordination: Need for country alliances, data sharing agreements, and collaborative governance

Key Points & Insights

  1. Empirical Evidence of Impact (US-Specific)

    • Analysis of payroll records through December 2025 shows 16% relative decline in employment for young workers in AI-exposed jobs, beginning late 2024–early 2025, with no reversal observed
    • Impact concentrated on entry-level workers; experienced workers show far less displacement to date
    • Different from macroeconomic forces like interest rate changes, suggesting AI-specific mechanism

  2. Exposure vs. Adoption Paradox

    • Research on AI exposure (technical feasibility) consistently predicts job declines
    • Research on AI adoption (actual firm implementation) shows mixed-to-null employment effects—a critical discrepancy
    • Possible explanations: firms anticipate AI need and reduce hiring preemptively; personal account usage not captured in enterprise metrics; competition drives non-adopters out of market; measurement error

  3. Global Disparity in AI Access & Impact

    • Advanced AI capabilities concentrated in ~2 countries (US and China)
    • GenAI usage: 24% of working adults in Global North vs. 14% in Global South—gap widening
    • Countries lacking reliable AI access face "dual disadvantage": unable to generate wealth via AI and must support displaced workers—potential fiscal crisis

  4. India's Optimistic Positioning

    • 92% of Indian knowledge workers already use AI; high firm-level adoption
    • Sector growth in fintech, healthtech, edtech solving traditional problems = job creation potential despite high exposure
    • Digital Public Infrastructure (DPI) model (Aadhaar, UPI) attracting 60 countries' interest post-G20 2023; replicable governance approach

  5. Kenya's Context-Specific Challenges

    • 52% job automation risk applies only to 15% of formal economy; 85% informal sector not captured in standard metrics
    • Opportunity: food systems, SME scaling (fintech example: companies operating with 10 employees vs. 100 previously)
    • Broken labor market information systems in Global South prevent accurate policy targeting
    • Median age 20, but large aging farming population (60+)—different transition needs

  6. Data as Foundation for Policy

    • All panelists emphasize inadequate data collection on AI adoption outside US/UK
    • Need for: firm-level usage data, executive beliefs about future AI capabilities, productivity effects, task-level rather than occupation-level analysis
    • Companies beginning to share usage data (e.g., Microsoft's 2025 Copilot report: 37.5M conversations; announced India-company data-sharing commitment)

  7. Anticipation Effects & Timing Lags

    • Firms may reduce hiring in anticipation of AI capabilities before actual deployment
    • Three-layer delay: (1) AI achieves capability, (2) market/society absorbs it, (3) economists study and inform policy—creates lag for proactive intervention
    • Speed of capability advancement may outpace policy response capacity

  8. Task-Level > Job-Title-Level Analysis

    • Microsoft research shows copilot use differs by context: mobile = health, desktop = work/career—same tool, different labor impacts
    • Framework: augmentation (seeking advice) vs. automation (end-to-end task replacement) varies by role, sector, company
    • Policy requires granular task measurement, not broad occupational categories

  9. Education & Skills as Central Lever

    • India: acute shortages of teachers, doctors, nurses; AI tools could address these gaps but cannot replace foundational education/cognitive development
    • Need for "education sector AI-ready" infrastructure; younger populations (median age 28 India, 20 Kenya) more adaptable but face displacement first
    • Reskilling programs historically underperform; without clarity on what to reskill to, programs risk ineffectiveness

  10. Multilingual & Cultural Safety

    • 40% of pre-training data is English; safety benchmarks lose context when translated; must co-design evaluations with communities in local languages (Hindi, Tamil, etc.)
    • Risk example (sonogram sex-determination): culturally specific harms emerge when AI systems deployed without local value consideration
    • Need for local AI talent and model development to embed cultural context

Notable Quotes or Statements

Yoshua Bengio (Turing Award winner, AI safety):

"71% of respondents to a US survey said they were concerned that AI will be putting too many people out of work permanently... this is exactly why we need to study this these questions scientifically so that there'll be real incentives for governments to mitigate what can happen."

"Around 60% of jobs in advanced economies and 40% in emerging economies are exposed... to general purpose AI. The impacts will depend on how AI capabilities develop... and how institutions respond, which is why we're having those discussions because we can still steer the future."

"Access or control of these most advanced AIs is going to become a decisive competitive advantage. Countries without reliable access would be facing a dual disadvantage."

Bharat Chander (Stanford economist):

"The work that looks at exposure has overwhelmingly found that exposure predicts declines in employment. Whereas the work on adoption has found very mixed results... we need better data to tease through each of these four different possibilities."

"[Speculating on other countries] I do want to say that we have uncertainty about the future because historically there have always been new jobs that are created... the question is whether AI is going to be different from what we saw in the past."

Ambassador Philip Thiggo (Kenya):

"It's 52% of what? So it's 52% of 15%, because only 15% of our economy is formal. So that's really the number they're looking at."

"Currently in the global south we really have broken labor market information systems. That's why you can't find data, simply because I think there's been years of disinvestment in the labor market."

Shamika Ravi (India, Economic Advisory Council):

"We are going to see the government increasingly get into those aspects of the labor market. But I think there are far more fundamental problems such as low skills... technology has definitely a very important role to play there."

Pamela Mishkin (OpenAI researcher):

"Resilience really isn't about predicting the future... it's primarily about how do we plan for all possible scenarios... you don't wait for a flood forecast to build levies."

"It's not only about whether long-term equilibrium looks okay, are there more jobs in the end? It's really what does that period of transition look like? Because transitions are really hard."

Hector Derwa (Microsoft, Responsible AI):

"Measurement discipline" is essential; the 2013 Frey-Osborne paper erred by looking at "jobs and occupations and not at jobs as a basket of tasks."

"The right metric is really tasks, it's not job titles... you need a repeatable, aggregated, privacy-preserving indicators [framework]."

Speakers & Organizations Mentioned

Speaker/RoleOrganizationKey Role
Yoshua BengioTuring Award winner; Law Zero (scientific director); Mila Quebec AI Institute (founder); International AI Safety Report (chair)Opening keynote; framing labor market risk
Bharat ChanderStanford Digital Economy LabEmpirical researcher on employment impacts
Dr. Shamika RaviEconomic Advisory Council to PM of IndiaIndia's policy perspective
Ambassador Philip ThiggoSpecial Envoy for Technology, Republic of KenyaKenya's labor and tech strategy; KAISA chair
Hector DerwaMicrosoft, Office of Responsible AIData infrastructure and adoption measurement
Pamela MishkinOpenAI (independent researcher)Resilience planning and task-level impact
Robert TragerOxford Martin AI Governance Initiative; Center for Governance of AIModerator
—Oxford Martin AI Governance InitiativeEvent host
—Center for Governance of AIEvent co-host
—International AI Safety Report (29 countries, EU, OECD, UN)Policy framework referenced
—Mila Quebec AI InstituteResearch; Canadian AI hub
—World BankGenAI usage index project; social protection research
—LinkedIn (Microsoft subsidiary)"Economic Graph" labor data infrastructure since 2014

Technical Concepts & Resources

Key Datasets & Methodologies

  • Payroll Records Analysis (Chander et al.): Covered millions of workers through December 2025; identified 16% relative employment decline in AI-exposed jobs for young workers
  • Microsoft Copilot Usage Report (2025): 37.5M conversations analyzed via machine classifiers; task- and context-labeled (not human-read for privacy)
  • Firm-Level AI Adoption Index: Measured across 40+ countries using online job postings; Microsoft/Bala Klein Tisink research
  • LinkedIn Economic Graph: Privacy-preserving labor market signals since 2014; hiring trends, labor shortages; now expanded globally

Frameworks & Models

  • AI Exposure vs. Adoption Framework:

    • Exposure: technical feasibility of automation (driven by model capabilities)
    • Adoption: actual firm implementation (driven by ROI, regulation, culture, costs)

  • Task-Level Classification (Copilot):

    • Topics (what the task is about)
    • Intents (information-seeking vs. advice-seeking)
    • Contexts (mobile/health vs. desktop/work)
    • Outcome: augmentation vs. end-to-end automation assessment

  • Privacy-Preserving Telemetry Framework:

    • Aggregated, anonymized, taxonomized, repeatable indicators
    • Comparable over time; scalable across countries
    • No raw personal data exposed

Key Papers & Reports Cited

  • Frey & Osborne (2013, Oxford): Estimated 50% of US jobs at high automation risk (critiqued as analyzing occupations rather than task baskets)
  • International AI Safety Report (29 countries, EU, OECD, UN): 60% of jobs in advanced economies / 40% in emerging economies exposed to general-purpose AI risk
  • Microsoft AI Diffusion Report (2025): 1-in-6 global individuals using GenAI; 24% vs. 14% Global North/South usage split

Policy Mechanisms & Initiatives

  • India's Digital Public Infrastructure (DPI):

    • Aadhaar (identity), UPI (payments)
    • 60 countries expressed interest post-G20 2023
    • Model for labor market data infrastructure

  • Kenya's AI Skilling Alliance (KAISA):

    • Focus: skills development, innovation, opportunity creation, policy collaboration
    • Sectors: food systems (farm-to-bottle), SME scaling

  • World Bank GenAI Usage Index: Aggregates company usage signals for policy-level GDP-equivalent metrics

  • Indian Government–Company Data-Sharing Commitment: Announced commitment by companies to share model usage data (announced later in summit week)

Technical Concepts

  • Anticipation Effects: Firms reduce hiring preemptively based on expected AI capabilities, not current ones
  • Task Substitution vs. Complementarity: Whether AI augments worker productivity or replaces work entirely; varies by task, role, sector
  • Multilingual Pre-training Gap: ~40% of web-crawled training data is English; safety benchmarks lose cultural context in translation
  • Labor Market Information Systems (LMIS): Data infrastructure for job reporting, skills matching; severely underfunded in Global South

Context & Significance

This summit represents a critical turning point in AI governance discourse: moving from abstract risk assessment to empirical, data-driven policymaking informed by multiple geographies and stakeholder perspectives. The emphasis on data sharing, international coordination, and task-level (not job-title) analysis signals a maturation in how AI impact is studied and governed. The prominent voices from India and Kenya underscore that Global South contexts—with different labor market structures, median ages, and institutional capacities—require tailored rather than imported solutions. The convergence on "resilience planning" (preparing for multiple scenarios rather than betting on a single forecast) offers a pragmatic path forward even amid deep uncertainty about AI capability timelines.