Economic Scenarios for Transformative AI: An India Adaptation
Summary
An India version of Anthropic's economic scenarios model finds AI raising India's GDP by 0.5%, 2.7% and 9.2% above its no-AI path by 2030 in the modest, substantial and extreme scenarios: roughly a third of the US gains. The costs to India's cognitive workers are as large as in the US or larger, so the room to compensate them is much smaller.
The adaptation re-implements Korinek et al. (2026), reproduces its published US results to within rounding, and recalibrates it with Indian data. It adds four India-specific mechanisms: an IT and business-services export sector whose demand follows the US scenario, partial earnings portability for displaced export workers, foreign ownership of AI capital, and labour weighted by income rather than headcount. The companion explorer runs the same model in the browser so readers can enter their own expectations.
Gains are smaller because AI reaches less of the economy. Cognitive work is about 26% of India's workers and 40% of its labour income, against 62% in the US, and Indian firms adopt more slowly. Even in the extreme scenario AI performs under 12% of India's labour tasks by 2030, against 30% in the US.
Costs are concentrated and arrive through two doors. Domestic adoption delivers almost all of the GDP gain. Foreign clients' adoption barely moves GDP but cuts export employment by 1.5%, 11% and 55% below its no-AI path, about 0.1, 0.7 and 3.3 million jobs, and causes about a fifth of the cognitive wage-bill loss. Cognitive employment adjusts mostly through hiring, so graduates seeking a first job bear most of it.
The compensation arithmetic is tight. Indian residents gain 2.1, 1.6 and 1.0 times what cognitive workers lose; in the US the GDP gain is 14.6, 4.8 and 2.8 times the loss. Up to 2.5% of GDP in AI capital income flows abroad. The ordering of results survives every robustness check, but the ratio is sensitive: 0.6 to 2.1 in the extreme scenario, driven by how easily India adds capital. The most protective single lever for exports is winning the new AI-services work that automation creates.
The scenarios separate in Indian data by early 2028, mainly in export headcount against its pre-AI trend (about −1%, −3% and −10%) and in educated unemployment.
Why India needs its own version
The original model transfers to India, but seven structural facts change what the same AI technology implies. Three of them shrink AI's aggregate effect in India; four concentrate its costs on a narrow group.
| Feature | United States (original paper) | India (this adaptation) | Why it matters in the model |
|---|---|---|---|
| Workers in AI-exposed (cognitive) occupations | 62.4% of employees | about 25% of jobs have high AI exposure (IMF figure cited by NITI Aayog); 26% used here | Caps how much of the economy AI can reach by 2030 |
| Cognitive share of labour income | set equal to employment share | about 40% (assumption: cognitive workers earn ~1.9x others) | Hulten weights: the GDP gain scales with income share, not headcount |
| Labour share of GDP | 60% | 53% (Penn World Table 11.0) | Divides gains between wages and capital |
| Unemployment in normal times | 3.8%; cognitive 2.9%, other 5.4% | 5.0% overall, urban youth 19.0% (PLFS, Sept 2026) | Educated job-seekers already queue for cognitive jobs |
| Export-oriented cognitive work | not modelled (closed economy) | tech exports USD 246bn and a workforce near 6 million in FY26 (Nasscom) | Foreign AI adoption removes Indian work without raising Indian productivity |
| Location and ownership of AI capital | mostly domestic | about 1% of global data-centre capacity, vs 47% for the US (Goldman Sachs via Business Standard) | Part of the income on automated tasks accrues abroad |
| How AI is used today | broad professional use | 2nd in total Claude.ai use but 101st per capita; 45.2% of tasks are software-related, the highest share of any country (Anthropic India brief) | Deep use in the export sector, thin diffusion elsewhere |
The three dampeners are the small cognitive share, slower domestic diffusion and foreign ownership of AI capital. The four concentrators are export exposure, a pre-existing queue of educated unemployed, entry-level hiring as the main adjustment margin, and geographic clustering. Four states (Maharashtra, Tamil Nadu, Karnataka and Delhi) account for over half of India's Claude.ai use, mirroring the IT sector's geography.
Early evidence points to the export channel first. The World Development Report 2026 finds that multinationals and firms in global value chains in South Asia cut hiring more than domestic firms after ChatGPT. It also reports that outsourced jobs to developing countries on one freelance platform fell 39% in 2025, most steeply in AI-exposed occupations. Nasscom's FY26 figures show revenue up 6.1% while headcount grew 2.3%.
What we keep from the original
The India version keeps the original model's machinery unchanged, so every difference in results traces to calibration or to the four extensions in the next section. I re-implemented the model from the paper's equations and reproduced its Table 3 almost exactly (Appendix): GDP gaps of 1.61, 8.29 and 32.42% against the published 1.6, 8.3 and 32.4%, and cognitive unemployment of 2.92, 4.50 and 17.90% against 2.9, 4.5 and 17.9%.
Output is a CES aggregate of task instances with elasticity σ = 0.5, so tasks are complements. AI reaches a share m of the economy's labour tasks, all of them cognitive. It is used on a share d of their instances and cuts the unit cost of each AI-performed instance by a log points. A share ψ of AI-performed instances is automated and done by capital; the rest are augmented and stay with workers. New labour tasks arrive at ρ per automated task.
Capital is supplied along an upward-sloping schedule with elasticity ε = 3. A semi-endogenous ideas block adds a little growth, but by 2030 its effect is well under 1%. Workers sit in two groups, cognitive and all other. When AI lowers demand for cognitive labour, the cognitive wage adjusts slowly (half-life one year) and firms lay off the excess. Displaced workers search with a discount μ on applications outside their group, and expanding occupations post vacancies at monthly speed θ^H.
| Object | Meaning | US modest / substantial / extreme |
|---|---|---|
| m in 2030 | share of labour tasks AI affects | 0.20 / 0.30 / 0.50 (ceiling 0.624) |
| d in 2030 | share of affected instances done with AI | 0.20 / 0.40 / 0.60 |
| a in 2030 | log cost saving per AI-performed instance | 0.30 / 0.45 / 0.80 |
| ψ | automated share of AI use | 0.50 / 0.75 / 0.90 |
| ρ | new labour tasks per automated task | 0.50 / 0.25 / 0 |
| μ | search discount when switching groups | 0.17 / 0.08 / 0.04 |
| θ^H | share of hiring shortfall posted each month | 0.10 / 0.25 / 0.50 |
Two properties of the original matter most for India. First, the wage tracks productivity (m·d·a), while the labour share falls with net displacement ((1−ρ)·ψ·m·d). Second, the cognitive group's proportional decline equals the other group's proportional rise scaled by s_N/s_C. In the US that ratio is 0.6; in India's domestic economy it is about 1.7. A given expansion of non-cognitive work therefore implies a much larger proportional squeeze on India's cognitive workers.
Four India-specific extensions
The original is a closed economy with homogeneous workers and domestically owned capital. India is none of these, so the adaptation adds four mechanisms. Each can be switched off, which recovers the original model.
1. An export island driven by foreign AI adoption. Part of India's cognitive work is sold abroad: tech exports of USD 246bn, roughly 4.5% of GDP in value added and 5.5% of labour income. Demand for this work follows the matching US scenario. Foreign clients' AI penetration of offshored tasks is ζ times their average penetration of cognitive work, because offshored tasks were chosen for being codified, digital and separable. The export employment target is:
The bracket is the share of offshored work that survives foreign AI; the last term is the extra demand for remaining human work that the US model generates. With ζ = 1 the path equals the US cognitive target. I use ζ = 1.5 and test 1.0 to 2.0. Export value added falls with export employment, except that Indian providers keep 15% of the lost labour income as margins.
2. Absorption at home. Displaced export workers join India's cognitive unemployed and search like other cognitive workers. In the frictionless targets they raise domestic labour supply, which lowers domestic wages by about 0.23% per 1% of extra labour and raises domestic output. In the frictional block they press down on the domestic cognitive wage, because they count in the attached cognitive labour force.
3. Foreign ownership of AI capital. Automated instances are performed by capital, and in India much of that capital is foreign: frontier models, cloud platforms and most data centres. I assume 60% of income earned on domestically automated instances accrues abroad. GDP is unchanged by this assumption, but resident income is lower.
4. Income-weighted labour units. The original sets the cognitive share of labour payments equal to its employment share, which is harmless when pay gaps are modest. In India cognitive workers are about 26% of workers but earn about 40% of labour income. The production block uses income shares, which are the correct Hulten weights. Headcounts are converted back using group earnings ratios, assuming pay is uniform within each group.
Two smaller changes complete the adaptation. Ideas are imported: India's technology frontier moves with the US scenario's ideas path rather than India's own research budget. And the inflow into cognitive job search includes new graduates searching for a first job, not only separations, which is how India's educated unemployment arises.
One mechanism I considered and did not add as a default is a labour-cost cushion on automation. AI is priced globally while Indian wages are far lower, so automating a task saves less in India. But Indian Claude.ai users delegate more autonomy to AI than the global average (3.60 vs 3.38 on a five-point scale), so usage gives no support for a lower automation share. The cushion enters instead through slower domestic diffusion, and the robustness section tests a lower automation share.
Calibration for India
The calibration keeps every technology parameter from the original and replaces the economy's structure and labour market with Indian values. Numbers marked "assumption" have no direct Indian estimate; the robustness section varies the ones that matter. Base year 2024, anchor mid-2026, scenarios read at the start of 2030, as in the original.
Structure and factor markets (common to all scenarios)
| Parameter | India value | US value | Source or basis |
|---|---|---|---|
| Task elasticity σ | 0.5 | 0.5 | unchanged |
| Labour share of GDP | 0.53 (0.524 in the domestic economy) | 0.60 | Penn World Table 11.0, 2019–2023 |
| Cognitive share of workers | 0.26 | 0.624 | IMF high-exposure share of about 25% (NITI Aayog) |
| Cognitive share of labour income | 0.40 (0.365 of domestic labour income) | 0.624 | assumption: implies cognitive pay ~1.9x other pay; tested at 0.30 and 0.50 |
| Export island value added | 4.5% of GDP | none | tech exports USD 246bn in FY26 (Nasscom) at ~75% value added, over nominal GDP of ₹346.36 lakh crore (MoSPI) |
| Labour share of export value added | 0.65 | none | assumption from IT firms' employee-cost ratios; gives 5.5% of India's labour income |
| Export headcount | 6.0 million | none | Nasscom FY26 headcount of about 5.95 million |
| Relative AI penetration of offshored work ζ | 1.5 | none | assumption; tested at 1.0 and 2.0 |
| Margin retained on lost export work | 15% | none | assumption |
| Foreign share of income on automated instances | 60% | 0 | assumption; India holds ~1% of global data-centre capacity |
| Capital supply elasticity ε | 3 | 3 | unchanged; tested at 1, 6 and ∞ |
| Growth without AI | 6.5% a year | 2.0% | assumption below FY26's 7.7% real growth (MoSPI) |
| Ideas stock | imported from US path | own block | India adopts frontier ideas; effect under 0.7% |
AI in India's domestic economy and disruptiveness (by scenario)
| Object | Modest | Substantial | Extreme | Basis |
|---|---|---|---|---|
| Affected mass m, mid-2026 anchor | 0.073 | 0.073 | 0.073 | within-cognitive exposure 0.20 (US 0.22), as India's cognitive work has more teaching, clerical and shop-management tasks |
| Affected mass m, 2030 (ceiling 0.365) | 0.105 | 0.158 | 0.263 | US capability as a share of cognitive work (32/48/80%) times 0.9 for language and local-context gaps |
| Diffusion d, mid-2026 anchor | 0.06 | 0.06 | 0.06 | US 0.10; India ranks 101st of 116 in per-capita Claude.ai use |
| Diffusion d, 2030 | 0.12 | 0.25 | 0.45 | US 0.20/0.40/0.60; slower firm adoption outside IT |
| Log gain a, 2030 | 0.30 | 0.45 | 0.80 | unchanged; Indian users see a 15x speed-up vs 12x globally (Anthropic) |
| Automation share ψ | 0.50 | 0.75 | 0.90 | unchanged; Indian users delegate more autonomy than average |
| Reinstatement ρ (also for exports) | 0.50 | 0.25 | 0 | unchanged |
| Search discount μ | 0.12 | 0.056 | 0.028 | India's normal-times value cut in the original's proportions |
| Posting speed θ^H | 0.10 | 0.25 | 0.50 | unchanged |
Labour market in normal times
| Parameter | India value | US value | Basis |
|---|---|---|---|
| Normal unemployment | 5.6% of income-weighted labour force; 4.8% of heads | 3.8% | matches PLFS 5.0% (CWS, ages 15+) |
| Cognitive / other unemployment rate | 8.5% / 3.5% | 2.9% / 5.4% | assumption reflecting the educated-youth queue; urban youth rate is 19.0% |
| Aggregate monthly job-finding rate | 0.12 (mean spell ~8 months) | 0.23 | assumption; no published Indian flow estimate of comparable form |
| Inflow into search, cognitive / other | 15.3% / 4.1% a year | 7.6% / 16.8% | calibrated to the rates above; cognitive inflow includes graduates seeking a first job |
| Share of the inflow that responds to job prospects | 0.35 | 0.55 | graduates keep arriving when prospects worsen |
| Normal-times search discount μ̄ | 0.12 | 0.17 | assumption: less movement between white-collar and manual work |
| Wage rigidity ξ | 0.5 a year | 0.5 | unchanged; tested at 0, 0.75 and 0.9 |
| Mean vacancy filling rate; matching curvature | 0.65; 1.27 | same | unchanged (implies matching efficiency 0.70) |
Three scenarios for India
Each India scenario pairs one of the original US scenarios with an Indian domestic counterpart. The technology is shared: AI that can do half of US cognitive work by 2030 can do roughly the same share of Indian cognitive work. What differs is how fast Indian firms adopt it, how much of India's economy it can reach, and what foreign adoption does to India's exports.
| Modest | Substantial | Extreme | |
|---|---|---|---|
| Global technology | internet-like; gains fixed at 35% per task | bigger than the internet; gains rise to 57% per task | recursively improving AI; gains more than double output per task |
| Foreign clients (US scenario) | 4% of US tasks done with AI by 2030 | 12% of US tasks, 75% of that automated | 30% of US tasks, 90% automated |
| India's domestic adoption by 2030 | AI used on 12% of instances it can do | 25% | 45% |
| Share of India's labour tasks done with AI, 2030 | 1.3% | 3.9% | 11.8% |
| Share of US labour tasks done with AI, 2030 | 4% | 12% | 30% |
| Offshored Indian work done by AI abroad, 2030 | about 10% | about 29% | about 72% |
| New tasks per automated task | 0.5 | 0.25 | 0 |
| Ease of switching occupations | normal | half of normal | a quarter of normal |
The gap between the two "share of tasks" rows is the core of the India story. Even in the extreme scenario, AI performs under 12% of India's labour tasks by 2030, against 30% in the US. Offshored Indian work, by contrast, is more exposed than average US work, so the export channel bites early and hard.
The scenarios bracket existing India estimates. Goldman Sachs projects AI adding about 0.4 percentage points a year to India's labour productivity growth (range 0.1 to 0.8). Applied from mid-2026, that lands between the modest and substantial scenarios. NITI Aayog's downside case, tech services headcount falling from 7.5–8 million in 2023 to 6 million by 2031, lies between the substantial and extreme export paths here. Its upside case of up to 4 million new jobs resembles the modest scenario with India capturing a large share of new AI-services work.
Results: India in 2030
AI raises India's GDP in all three scenarios, by 0.5%, 2.7% and 9.2% above the no-AI path, roughly a third of the US gains. The costs to cognitive workers are as large as in the US, or larger. All figures are for the start of 2030, relative to the economy without AI; US columns are the original paper's Table 3.
| Outcome in 2030 | India modest | India substantial | India extreme | US modest | US substantial | US extreme |
|---|---|---|---|---|---|---|
| GDP, % above no-AI path | 0.5 | 2.7 | 9.2 | 1.6 | 8.3 | 32.4 |
| GDP growth in 2030, % a year (no-AI: India 6.5, US 2.0) | 6.6 | 7.7 | 10.7 | 2.4 | 5.4 | 15.4 |
| Domestic economy output, % | 0.6 | 3.3 | 12.2 | 1.6 | 8.3 | 32.4 |
| Export value added, % | −1.5 | −11.1 | −53.6 | — | — | — |
| Resident income, % above no-AI path | 0.3 | 1.9 | 6.8 | 1.6 | 8.3 | 32.4 |
| AI capital income paid abroad, % of GDP | 0.2 | 0.7 | 2.5 | 0 | 0 | 0 |
| Average wage, % | 0.2 | 0.4 | 2.8 | 0.7 | 2.1 | 9.7 |
| Cognitive wage, % | −0.2 | −2.1 | −14.8 | 0.4 | −0.3 | −11.5 |
| Other wage, % | 0.4 | 1.9 | 11.7 | 1.1 | 5.9 | 33.6 |
| Labour share, % of income (no-AI: India 53.0, US 60.0) | 52.8 | 51.8 | 47.6 | 59.4 | 56.1 | 45.2 |
| Total labour income, % | 0.2 | 0.3 | −1.9 | 0.6 | 1.4 | 0.5 |
| Cognitive wage bill, % | −0.8 | −5.7 | −32.1 | −0.3 | −4.6 | −31.0 |
| Capital income, % | 0.9 | 5.4 | 21.8 | 3.1 | 18.9 | 81.4 |
| Net return to capital, % (no-AI: India 9.0, US 6.5) | 9.0 | 9.2 | 9.8 | 6.6 | 7.0 | 8.3 |
| Cognitive employment incl. exports, % since mid-2026 | −0.4 | −3.3 | −19.9 | −0.5 | −3.9 | −21.5 |
| Export employment, % vs no-AI path | −1.5 | −11.0 | −54.5 | — | — | — |
| Cognitive unemployment rate, % (normal: India 8.5, US 2.9) | 8.6 | 9.8 | 20.1 | 2.9 | 4.5 | 17.9 |
| Overall unemployment, % of workers (normal: India 4.8, US 3.8) | 4.8 | 4.4 | 6.4 | 3.9 | 4.6 | 11.9 |
| Measured TFP, % | 0.2 | 0.9 | 4.6 | 0.7 | 3.1 | 13.4 |
Modest. AI is a marginal technology for India's economy by 2030. GDP is 0.5% higher, growth reaches 6.6% instead of 6.5%, and no labour-market indicator moves by more than a few tenths of a point. Export employment is 1.5% below its no-AI path, about 90,000 jobs, which is within the noise of normal IT hiring cycles.
Substantial. India gains about a third of what the US gains, but cognitive workers lose almost as much. GDP is 2.7% higher and growth in 2030 is 7.7%, while the cognitive wage falls 2.1% below its no-AI path; in the US it is essentially flat. Export employment is 11% below its no-AI path, about 660,000 fewer jobs, and cognitive unemployment rises from 8.5% to 9.8%. Overall unemployment edges down, because expanding non-cognitive work absorbs people from India's larger pool of non-cognitive job-seekers.
Extreme. India grows at 10.7% in 2030, but the gains split very unevenly. Cognitive wages fall 14.8%, more than the US's 11.5%, and cognitive unemployment reaches 20.1%. Export employment is 55% below its no-AI path, about 3.3 million jobs. Total labour income ends 1.9% below its no-AI level even though GDP is 9.2% higher, so every rupee of the gain goes to capital, and a quarter of the extra capital income goes abroad.
Three features distinguish India from the US across all scenarios. Gains are smaller because AI reaches less of the economy: under 12% of India's labour tasks even in the extreme scenario. Cognitive losses are proportionally larger because India's cognitive group is small relative to everyone else, and because foreign adoption removes export work without raising Indian productivity. And a growing wedge opens between GDP and resident income, reaching 2.5% of GDP in the extreme scenario.
Who gains and who loses
In India's extreme scenario, what Indian residents gain is roughly equal to what cognitive workers lose; in the US the gain is almost three times the loss. That one ratio is the sharpest difference between the two economies, and it comes from three places at once: smaller gains, a leak of capital income abroad, and export losses that land entirely on cognitive workers.
Income changes in 2030 by group, % of no-AI GDP
| Group | Modest | Substantial | Extreme |
|---|---|---|---|
| Cognitive workers (wage bill, unemployed counted at zero) | −0.16 | −1.20 | −6.81 |
| All other workers | +0.25 | +1.34 | +5.79 |
| Domestic owners of capital | +0.26 | +1.80 | +7.78 |
| Foreign owners of AI capital | +0.16 | +0.74 | +2.47 |
| GDP | +0.51 | +2.68 | +9.23 |
Non-cognitive workers gain in every scenario, because their work complements AI-cheapened cognitive tasks. In India these are mostly lower-paid workers, so AI by itself narrows the gap between India's cognitive and non-cognitive earnings.
The compensation arithmetic, India vs US
| India modest | India substantial | India extreme | US modest | US substantial | US extreme | |
|---|---|---|---|---|---|---|
| Cognitive workers' loss, % of no-AI GDP | 0.16 | 1.20 | 6.81 | 0.11 | 1.73 | 11.59 |
| Transfer to hold them at no-AI income, % of 2030 GDP | 0.16 | 1.17 | 6.24 | 0.11 | 1.60 | 8.75 |
| Gain available ÷ cognitive loss | 2.1 | 1.6 | 1.0 | 14.6 | 4.8 | 2.8 |
For India the available gain is resident income; for the US it is GDP, since all capital is domestic. The US extreme transfer of 8.75% matches the paper's "about 9 percent of GDP". India's transfer is smaller as a share of GDP, but the gain it must come from is smaller still.
Two doors: domestic adoption versus foreign adoption
| Channel switched on | GDP, % (M / S / E) | Cognitive wage bill, % (M / S / E) | Export employment, % (M / S / E) |
|---|---|---|---|
| Domestic AI only | 0.53 / 2.86 / 10.53 | −0.57 / −4.19 / −24.93 | 0 / 0 / 0 |
| Foreign AI only | 0.03 / −0.02 / −0.25 | −0.15 / −1.24 / −6.51 | −1.5 / −11.0 / −54.5 |
| Both | 0.51 / 2.68 / 9.23 | −0.77 / −5.65 / −32.14 | −1.5 / −11.0 / −54.5 |
Foreign adoption on its own barely moves GDP, because displaced export workers find other work and produce something, at half their former earning power. But it moves about a fifth of the cognitive wage-bill loss. Domestic adoption delivers almost all of the gain. The combination is worse than the sum for cognitive workers, because export-displaced workers compete for the same domestic cognitive jobs that domestic AI is shrinking.
The adjustment runs through hiring, so first-job seekers bear it. Layoffs explain none of the fall in cognitive employment in the modest and substantial scenarios and 19% of it in the extreme scenario. The rest comes from departures that are not replaced and graduates who are not hired: hiring into cognitive jobs runs 2%, 12% and 52% below normal between mid-2026 and 2030. That lands on a group already under strain, with urban youth unemployment at 19.0% and 26.6% for young urban women in September 2026 (PLFS via The Tribune).
In people. Export jobs end 0.09, 0.66 and 3.3 million below their no-AI path. The number of unemployed cognitive workers rises by about 0.2, 2.0 and 18 million from a normal level of about 13 million. These head counts assume uniform pay within each group, so they slightly overstate the extreme-scenario rise, because displaced export workers earn more than the cognitive average.
In places. Software exports are concentrated in four states: Karnataka 35.0%, Maharashtra 19.8%, Telangana 15.3% and Tamil Nadu 12.8% of the ₹19.36 lakh crore total in 2024–25 (MeitY data via OfficeChai). Claude.ai use shows the same geography, with Maharashtra, Tamil Nadu, Karnataka and Delhi accounting for over half (Anthropic). The export channel's losses would fall mostly on Bengaluru, Hyderabad, Pune, Mumbai and Chennai, and on the local services those salaries support.
Robustness and sensitivity
The ordering of outcomes survives every variation tested: India gains less than the US, cognitive workers lose, other workers gain, and the labour share falls. The compensation ratio is the most fragile result. In the extreme scenario the transfer needed ranges from 0.6 to 2.1 times residents' gain, driven mainly by how easily India adds capital and how much of its cognitive labour income there is.
| Variation | Subst. GDP % | Subst. transfer ÷ resident gain | Extreme GDP % | Extreme cognitive wage % | Extreme cognitive unemployment % | Extreme transfer ÷ resident gain |
|---|---|---|---|---|---|---|
| Baseline | 2.7 | 0.62 | 9.2 | −14.8 | 20.1 | 1.01 |
| Capital supply elasticity ε = 1 | 2.0 | 1.21 | 6.1 | −20.8 | 19.6 | 2.14 |
| ε = 6 | 3.0 | 0.48 | 10.6 | −12.2 | 20.3 | 0.79 |
| ε = ∞ (capital at a fixed world return) | 3.4 | 0.35 | 12.4 | −8.5 | 20.6 | 0.59 |
| Flexible cognitive wage (ξ = 0) | — | — | 11.7 | −36.4 | 9.0 | 0.95 |
| Rigid cognitive wage (ξ = 0.9) | 2.4 | 0.64 | 7.7 | −1.8 | 25.3 | 1.10 |
| Offshored AI penetration ζ = 1.0 | 2.8 | 0.50 | 10.0 | −13.5 | 18.7 | 0.79 |
| ζ = 2.0 | 2.6 | 0.75 | 8.5 | −16.1 | 21.6 | 1.27 |
| Cognitive share of labour income 0.30 | 1.9 | 1.15 | 6.3 | −19.7 | 21.2 | 1.73 |
| Cognitive share of labour income 0.50 | 3.4 | 0.37 | 12.6 | −10.3 | 18.9 | 0.62 |
| Foreign share of AI capital income 30% | 2.7 | 0.52 | 9.2 | −14.8 | 20.1 | 0.85 |
| Foreign share of AI capital income 90% | 2.7 | 0.76 | 9.2 | −14.8 | 20.1 | 1.24 |
| Export workers keep full earning power | 2.9 | 0.55 | 9.9 | −17.1 | 21.1 | 0.94 |
| Indian firms adopt at US speed | 4.2 | 0.60 | 12.3 | −22.5 | 24.1 | 1.02 |
| India captures new AI-services work (export ρ = 0.5) | 2.8 | 0.51 | 10.3 | −12.9 | 18.1 | 0.71 |
| Lower automation share (ψ − 0.15) | 2.4 | 0.55 | 8.8 | −12.0 | 18.7 | 0.89 |
| Larger export island (6% of GDP) | 2.5 | 0.74 | 8.1 | −16.1 | 20.4 | 1.25 |
| US-like job matching (μ̄ = 0.17, finding rate 0.20) | 2.6 | 0.63 | 9.8 | −12.9 | 18.9 | 0.90 |
| Labour share 0.45 | 2.5 | 0.68 | 8.6 | −16.7 | 20.1 | 1.09 |
Four patterns stand out.
- Capital supply matters most, as in the original. At ε = 1 the rental rate absorbs more of the gain and the extreme-scenario cognitive wage falls 20.8%. At ε = ∞, which describes compute imported at a fixed world price, India keeps more of the gain as wages.
- Wage rigidity decides whether the cost is lower pay or unemployment. Fully flexible wages hold cognitive unemployment at 9.0% but cut the cognitive wage by 36%; very rigid wages hold the wage near its no-AI path but push unemployment to 25%.
- Faster domestic adoption is a trade-off, not a free gain. Adopting at US speed raises GDP by a third in the extreme scenario but deepens the cognitive wage loss to 22.5% and unemployment to 24.1%.
- Capturing new AI-services work is the single most protective lever for exports. With half a new task per automated task, export employment falls 9.5% instead of 55% in the extreme scenario.
The export channel's size depends almost entirely on ζ. Export employment by 2030 falls 23%, 55% and 86% below its no-AI path at ζ = 1.0, 1.5 and 2.0 in the extreme scenario, and 5%, 11% and 17% in the substantial one. ζ is also the parameter with the least direct evidence behind it, which makes it the first thing worth measuring.
What the model leaves out
On balance the omissions understate the costs to India's cognitive workers; their net effect on aggregate GDP could go either way. The original's own caveats still apply: no policy responses, business cycles, financial disruption, robotics or catastrophic risks, and nothing beyond 2030.
| Omission | Likely direction of bias | Why |
|---|---|---|
| No earnings loss when a domestic cognitive worker takes non-cognitive work | Costs understated | Cognitive pay is about 1.9x other pay in this calibration; only export workers carry an explicit loss (50%) |
| No underemployment margin | Gains to non-cognitive workers overstated | Extra demand for physical work could be met by underemployed farm and casual workers working more days, which damps the wage rise. The small fall in overall unemployment in the substantial scenario likely reflects this |
| No aggregate-demand feedback | Downside understated | Tech salaries support urban consumption, rents and local services in a few cities; their loss would spread beyond cognitive workers |
| No exchange rate | Export loss somewhat overstated | Lower services exports would weaken the rupee, cushioning remaining exports; it would also raise import costs |
| Data-centre investment demand not modelled | Near-term GDP understated | The buildout is investment spending in India even where the capital is foreign-owned |
| Export demand tied to the US scenario | Ambiguous | Europe, the Middle East and Asia-Pacific clients may adopt at different speeds |
| AI savings on export work assumed to pass to clients | Ambiguous | Indian providers could keep more margin (upside) or face price cuts beyond headcount cuts (downside) |
| No age, gender, caste or regional detail | Distribution understated | Young urban women already face 26.6% unemployment; the hiring margin falls on the young |
| Indian labour-market flows assumed, not measured | Unknown | No Indian equivalent of the CPS matched flows used in the original |
| No survey of Indian expectations | Not applicable | The original's survey of US adults has no Indian counterpart here; the explorer lets users enter their own |
| Capability discount of 0.9 for Indian languages and context | Ambiguous | A judgement call; the gap may close quickly |
Two omissions deserve emphasis. The first is the absence of an earnings penalty for domestic workers who switch groups, which the original paper's reviewers already flagged for the US. In India the pay gap that switching crosses is larger, so this is the omission most likely to make the scenarios look kinder than they would be. The second is the lack of an underemployment margin: India's non-cognitive labour market adjusts more through hours and days worked than through unemployment, which the two-group search model cannot represent.
Indicators to watch
By early 2028 the three scenarios separate enough to tell apart in Indian data, mainly through export headcount and educated unemployment. As in the original, almost all divergence comes after 2027, so the next 18 months of data are the cheapest test of which world India is in.
What each scenario predicts by early 2028
| Marker | Modest | Substantial | Extreme |
|---|---|---|---|
| Export headcount vs its pre-AI trend | about −1% | about −3% | about −10% |
| Cognitive unemployment rate (normal 8.5%) | 8.6% | 9.0% | 10.5% |
| GDP vs no-AI path | +0.3% | +0.8% | +2.4% |
Where to look
| Indicator | Latest reading | Model object it measures | Source |
|---|---|---|---|
| Tech revenue growth minus headcount growth | 6.1% vs 2.3% in FY26 | export demand path and ζ | Nasscom Annual Strategic Review, each February |
| Fresher intake at large IT firms | quarterly company results | hiring margin of cognitive adjustment | company filings |
| Urban youth and graduate unemployment | urban youth 19.0%, September 2026 | cognitive unemployment, search frictions μ | PLFS monthly bulletin |
| Share of AI use outside software, and per-capita AI use rank | 45.2% of tasks software-related; 101st of 116 per capita | domestic diffusion d | Anthropic Economic Index India brief |
| AI use by Indian firms outside IT | no official series | domestic diffusion d | gap: no Indian equivalent of the US Census AI supplement |
| Net factor income paid abroad (GDP minus GNI) | 1.2% of GDP in FY26 (₹346.36 vs ₹342.05 lakh crore) | foreign ownership wedge | MoSPI |
| Imports of computer and AI services | RBI services trade data | foreign ownership wedge (AI bought as an import) | RBI |
| Labour share in the new national accounts | compensation of employees near one third of GVA | labour share path | MoSPI |
| Data-centre capacity and its ownership | about 1% of global capacity | ε and foreign ownership | Goldman Sachs via Business Standard |
The largest measurement gap is domestic firm adoption. The original calibrates diffusion from the US Census Bureau's business survey; India has no comparable series, so adding AI-use questions to an existing enterprise survey would close the weakest link in this calibration.
Implications for policy
The model does not choose policies, but it changes the question India faces. The original concludes that if extreme change arrives, the central problem is sharing the gains of a much richer economy. For India the gains available to share are thinner, so policies that enlarge or retain the gain matter as much as policies that redistribute it. Each option below has a real case against it.
| Option | What the model says it moves | Case for | Case against |
|---|---|---|---|
| Capture new AI-services work (AI integration, evaluation, agent operations, Indian-language data) | export reinstatement ρ: export losses fall from 55% to 9.5% in the extreme scenario | the most protective lever for the export sector; builds on existing client relationships; NITI Aayog sees up to 4 million new roles | AI-services revenue was about USD 10–12bn in FY26, a small base; clients may bring this work in-house or use AI-native firms |
| Large-scale AI skilling (NITI's proposed AI Talent Mission) | search discount μ and the hiring margin | the adjustment falls on first-job seekers, whom training can reach before they enter the queue | evidence on large retraining programmes is mixed; in the extreme scenario the binding constraint is demand for cognitive work, not skills |
| Faster AI diffusion to domestic firms, public services and Indian languages | domestic adoption d | raises GDP and resident income most directly (+33% in the extreme scenario at US adoption speed) | deepens cognitive displacement: the cognitive wage falls 22.5% instead of 14.8% |
| Domestic ownership of compute and models | foreign share of AI capital income | shrinks a wedge of up to 2.5% of GDP; resident gain rises from 5.5% to 8.0% of GDP as the foreign share falls from 90% to 30% | does not raise GDP; risks subsidising capacity foreign providers supply more cheaply, and the model has no return-on-public-capital penalty |
| Income support for displaced cognitive workers or first-job seekers | redistributes the cognitive loss | the loss is concentrated and visible, and India has little income insurance for formal workers | full compensation would cost up to 6.2% of GDP; cognitive workers are among India's better-paid, so transfers toward them compete with poorer claimants; non-cognitive workers gain without support |
| Services-trade diplomacy (data flows, recognition, mobility) | export demand | protects demand that AI does not destroy; Goldman Sachs flags protectionism as a risk to services exports | trade rules cannot offset clients' own automation decisions |
| Wait for evidence | none | the data so far fit the modest or early substantial path; the World Bank judges fears of widespread job losses in developing countries overstated | divergence comes fast after 2027, and the hiring margin hits cohorts that cannot wait |
Two results cut across these options. Raising the gain and protecting cognitive workers pull in opposite directions on the domestic side: faster adoption makes the economy richer and the cognitive labour market worse. And the export channel is the one place where a single lever, capturing the new work AI creates, improves both.
Appendix
A. Replication of the original US results
The model was re-implemented from the paper's equations (Proposition 1, the system (39), the posting, matching and wage-rigidity rules, and the steady state (38)) on a monthly grid from January 2024 to January 2030. With the paper's Table 1 parameters it reproduces the published Table 3 to within rounding.
| US outcome in 2030 (M / S / E) | Published | This implementation |
|---|---|---|
| GDP, % above no-AI path | 1.6 / 8.3 / 32.4 | 1.61 / 8.29 / 32.42 |
| GDP growth, % a year | 2.4 / 5.4 / 15.4 | 2.40 / 5.37 / 15.46 |
| Cognitive wage, % | 0.4 / −0.3 / −11.5 | 0.41 / −0.33 / −11.49 |
| Other wage, % | 1.1 / 5.9 / 33.6 | 1.14 / 5.86 / 33.61 |
| Capital stock, % | 2.3 / 13.8 / 56.3 | 2.31 / 13.83 / 56.29 |
| Labour share, % | 59.4 / 56.1 / 45.2 | 59.42 / 56.16 / 45.52 |
| Cognitive employment since mid-2026, % | −0.5 / −3.9 / −21.5 | −0.46 / −3.87 / −21.45 |
| Cognitive unemployment, % | 2.9 / 4.5 / 17.9 | 2.92 / 4.50 / 17.90 |
| Overall unemployment, % | 3.9 / 4.6 / 11.9 | 3.86 / 4.53 / 11.86 |
| Measured TFP, % | 0.7 / 3.1 / 13.4 | 0.71 / 3.07 / 13.41 |
The one visible gap, the extreme-scenario labour share, is internal to the paper: its own labour income (+0.5%) and GDP (+32.4%) imply 0.60 × 1.005 / 1.324 = 45.5%, which is what this implementation reports.
B. Changes to the model for India
- The cognitive group contains a domestic part and an export island. The island's employment target follows the export equation in the extensions section, built from the matching US scenario's m, d, a, ψ, ρ and its non-cognitive demand shift.
- Export firms shed workers first by not replacing departures, then by layoffs. Displaced export workers enter the cognitive unemployment pool, but half of their earning capacity is lost on net shrinkage of the island.
- Proposition 1 gains a domestic labour-supply shift ΔlnL equal to the portable share of displaced export labour. Both domestic targets scale by e^ΔlnL, and capital demand rises with ΔlnL.
- Domestic cognitive wages are set by the original sticky-wage rule, with export-displaced unemployed counted in the attached cognitive force.
- GDP adds export value added to domestic output. Resident income subtracts 60% of the income earned on domestically automated instances.
- The ideas stock follows the US scenario's path rather than India's own research block.
- The labour block's inflow into cognitive search includes graduates searching for a first job; 35% of the inflow responds to job prospects (55% in the original).
C. A survey of Indian expectations
The original surveyed 10,980 US adults. An Indian counterpart would need mixed modes (phone plus face-to-face) to reach adults without smartphones, at least ten languages, and weights from PLFS. The five questions carry over, but the eight benchmark tasks should include India-relevant ones, such as replying to a GST notice, handling a customer call in a regional language, and maintaining a client's software system. Three questions should be added: whether the respondent's employer serves foreign clients, how long a displaced IT or BPO worker would take to find work, and whether AI services the respondent uses are Indian or foreign.
D. Code
The companion explorer runs the full model in the browser; its page source contains the complete implementation, including the US model used to drive the export channel. The browser version matches the Python implementation used for this report to two decimal places on every reported outcome.
Sources
- Korinek, Jones, Sacher, Cotter and McCrory (2026), Economic Scenarios for Transformative AI, The Anthropic Institute Working Paper 2026-02
- Anthropic, Scenarios for our Economic Future, scenario explorer v1.0, September 2026
- Anthropic, India Country Brief: The Anthropic Economic Index, February 2026
- Nasscom, Annual Strategic Review 2026 press release, 24 February 2026
- NITI Aayog, Roadmap for Job Creation in the AI Economy, October 2025
- MoSPI, Provisional Estimates of Annual GDP for 2025-26, 5 June 2026
- Penn World Table 11.0 via FRED, Share of Labour Compensation in GDP, India
- The Tribune (ANI), PLFS monthly bulletin for September 2026, 15 October 2026
- Business Standard, World Development Report 2026 on AI and outsourcing, 4 August 2026
- Business Standard, Goldman Sachs on Gen-AI and India's non-farm workforce, 29 July 2026
- OfficeChai, State-wise software exports 2024–25 (MeitY STPI and SEZ data), 5 April 2026