Economic scenarios — methodology

A what-if engine for AI's labor & growth impact. The invariants say what must hold; this page says why.

0. What this is

A what-if engine, not a forecast. The economic scenarios module lets you move five drivers of AI's labor-market impact and watch GDP, jobs, wages and the labor–capital split respond, beside live US baseline data. It is an independent, open implementation inspired by and calibrated against the published reference scenarios in Anthropic's *Economic Scenarios for Transformative AI* (Korinek, Jones, Sacher, Cotter, McCrory, 2026). It is not endorsed by Anthropic and reproduces none of their text, charts or assets.

Every number on screen carries a provenance flag — LIVE, FILED, GUIDANCE, ESTIMATE, STALE or USER — and live values set the sliders' defaults; they are never locks.

1. The chain, stage by stage

Jobs are bundles of tasks (the O*NET taxonomy). For each knowledge-work task, AI can leave it unchanged, augment it (assist a human), automate it (do it without a human), or a new task is created. The five drivers:

Driver	Meaning	Range
Capabilities	share of knowledge tasks AI can do	0–100%
Adoption	of those, the share where AI is actually used	0–100%
Autonomy	of AI-used tasks, the share done without a human	0–100%
Productivity	output multiplier on AI-affected tasks	1×–10×+
Adjustment speed	months for a displaced worker to find new work	1–36+

From these the model derives task penetration, a concave GDP response, a labor→capital share shift, wage changes by worker group, and worker-flow and unemployment outcomes. The full equations and calibrated coefficients are documented in the repository's docs/ECON_MODEL_SPEC.md.

Worked scenarios

Modest — AI resembles the internet: gradual gains within historical norms. ~+1.6% GDP by 2030, labor share barely moves.
Substantial — AI can do half of knowledge work by 2030, mostly autonomously, but adoption lags; the economy grows at about twice its normal rate. ~+8% GDP, labor share toward 56%, knowledge wages roughly flat.
Extreme — AI outperforms on nearly all knowledge work, autonomously, creating no new human tasks. ~+32% GDP, labor share toward 45%, knowledge wages fall by more than 10%.
From my capex winter — arrive from the capex model after cutting capex conviction, and the bridge lowers the capability and adoption defaults: less compute built, a smaller labor-market shock.

2. Why the invariants are economics

Each invariant encodes one claim that must always hold:

INV-E1 — zero-AI reconciles to baseline. With capabilities or adoption at zero, every delta is zero: no uplift, no share shift, no displacement. The base case is honest.
INV-E2 — factor shares sum to one. Labor share + capital share = 1, always.
INV-E3 — workers are conserved. Stayed, crossed over, still-to-move and displaced sum to 100% of workers.
INV-E4 — friction is monotone. A longer adjustment time never *lowers* unemployment, all else equal.
INV-E5 — provenance is never silent. Every displayed value carries a flag; an ESTIMATE is never shown as LIVE; a failed fetch flips to STALE and the model keeps solving.
INV-E6 — clamps hold. At every slider extreme, shares stay in [0,1], unemployment ≤ 100%, wages above −100%.
INV-E7 — more capability or adoption never reduces GDP uplift.
CAL-1 — calibration. With inputs set to each published reference scenario, GDP uplift lands within ±1 point and labor share within ±1.5 points of the report's headline numbers, with wage direction matching.

3. How the two models connect

The capex model answers *who gets paid to build AI*; this one answers *what the build does to jobs, wages and growth.* They are linked by one explicit bridge: the capex model's solved system-capex ratio feeds a transparent, monotone elasticity that sets this module's capability, adoption, autonomy and productivity defaults. The bridge moves only defaults — you can always override them — and it is flagged ESTIMATE: *this link is a judgement call, not an empirical estimate.* There is no feedback the other way; the labor model never changes the capex solve.

Read distributionally, the labor-share loss and the fall in knowledge-worker wages are the household-side reading of the same "mortgage on the future" that the capex model's depreciation wall measures on the balance sheet. See the capex methodology §6.

4. Sources and known limitations

Live and derived data

FRED (same path as the capex model): real GDP, unemployment, labor share, employment, and JOLTS job openings/hires/separations. Each carries an as_of date and refreshes on cadence (monthly/quarterly), never real-time.
BLS OEWS (processed): occupation employment, used to derive the knowledge-worker share of employment via a documented, editable mapping of SOC groups — a judgement call, flagged ESTIMATE.
**O*NET** (sample): task statements, to make concrete what a "task" is.
Anthropic Economic Index (open dataset): AI-usage shares, which inform the adoption default; any exposure we derive is ESTIMATE.

What the model does not do

Following the report's stated limitations and adding our own: no policy responses, no business cycles, no aggregate-demand effects, no data-centre build-out demand, no robotics or physical automation, no individual-worker trajectories, no catastrophic risks, and no feedback from the labor side back into the capex model. The factor-share and wage coefficients are calibrated to three published points, not estimated from micro-data — they are ESTIMATE and editable in the assumptions panel, where an override flips the field to USER.

Required disclosures

Inspired by and calibrated against the published reference scenarios in Anthropic's report; krzy.ai's implementation is independent and not endorsed by Anthropic.
Every parameter that is an estimate is labelled as one, and can be overridden.
This is a what-if engine. It explores consequences of assumptions; it does not predict the future.