---
title: "The AI Layoff Trap"
aliases:
  - "AI Layoff Trap"
  - "Falk & Tsoukalas 2026"
tags:
  - theoretical
  - displacement
  - policy
  - arxiv
  - automation-tax
  - demand-externality
source: "/research/2026-03-arxiv-ai-layoff-trap.pdf"
extract: "/research/extracts/2026-03-arxiv-ai-layoff-trap.md"
authors: "Brett Hemenway Falk; Gerry Tsoukalas"
publisher: "arXiv (econ.TH)"
date: "2026-03-21"
---

# The AI Layoff Trap

> [!info] Source
> PDF (human): [2026-03-arxiv-ai-layoff-trap.pdf](/research/2026-03-arxiv-ai-layoff-trap.pdf) · Raw extract (machine): [2026-03-arxiv-ai-layoff-trap.md](/research/extracts/2026-03-arxiv-ai-layoff-trap.md)
> arXiv:2603.20617v1 — 21 March 2026 — Theoretical economic model (task-based automation), 53 pages
> **Keywords:** artificial intelligence, automation, labor displacement, Pigouvian tax

## TL;DR

A theoretical paper showing that even when firms can foresee that mass automation will erode the consumer demand they depend on, **competitive demand externalities trap rational firms in an automation arms race that exceeds the collectively optimal level**. Six policy instruments are evaluated: capital income taxes, worker equity participation, UBI, upskilling, and Coasian bargaining all fail to fix it. Only a Pigouvian automation tax restores the cooperative optimum.

## Key findings

- **The mechanism:** an automating firm captures the full cost saving but, under competitive pricing, bears only a fraction of the resulting aggregate demand destruction; the rest falls on rivals. Each firm's profit-maximising automation rate is a strictly dominant strategy that exceeds the cooperatively efficient level.
- **Competition deepens the distortion.** A monopolist fully internalises the externality; fragmented markets exhibit the widest gap. In the frictionless limit the game becomes a Prisoner's Dilemma — every firm displaces its entire human workforce, even though collective restraint would raise all profits.
- **The loss is deadweight, not transfer** — both workers and firm owners are worse off.
- **Robustness:** higher AI productivity *amplifies* (not resolves) the wedge — a "Red Queen effect". Endogenous wage adjustment, free entry, capital-income recycling, and richer market structures all fail to eliminate it.
- **Policy comparison:**
  - Capital income taxes: ineffective (operate on profit levels, not per-task margin).
  - UBI: ineffective (raises living-standard floor, leaves automation incentive unchanged).
  - Worker equity participation, upskilling, Coasian bargaining: narrow but cannot eliminate the wedge.
  - **Pigouvian automation tax**, set equal to the uninternalised demand loss per task, implements the cooperative optimum. Revenue can fund retraining, raising income replacement and shrinking the externality over time — potentially self-limiting.

## Notable real-world examples cited

- **Block (Feb 2026):** cut nearly half of its 10,000-person workforce; CEO Jack Dorsey said AI made many roles unnecessary and predicted "within the next year, the majority of companies will reach the same conclusion".
- **2025 tech layoffs:** >100,000 tech workers laid off; AI cited as primary driver in >half of cases. Concentrated in customer support, operations, middle management.
- **Salesforce:** replaced 4,000 customer-support agents with agentic AI.
- **Cognition's Devin** (Goldman Sachs, Infosys deployments): one senior engineer can do the work of a five-person team.
- **Eloundou et al. (2024):** ~80% of US workers hold jobs with tasks susceptible to LLM automation.

## Methodology in brief

Task-based automation model adapted from Acemoglu & Restrepo (2018), refocused from labor market to product market. N symmetric firms each choose automation rate α ∈ [0,1]; automated tasks cost c per task, human tasks cost wage w. Workers spend fraction λ of income on the sector; owners spend none. Fraction η of displaced wage income is replaced via reemployment/transfers. Quadratic integration friction k/2·α². Closed-form equilibrium with policy extensions.

## Implications for AdaptAI

### Calculator (`/calculator`)
- Reinforces the model: routine and high-`computerWork` jobs are at the front of the queue. No direct change to constants from this paper alone (it is theoretical), but supports keeping `repetitiveness` and `computerWork` weights as strong drivers.

### Big Picture (`/big-picture`)
- `bigPictureData.skillsDemand` and `industryExposure` — supports the framing that displacement is not self-correcting at high adoption rates.
- The narrative copy on `/big-picture` should acknowledge the **Red Queen effect**: better AI doesn't slow displacement, it accelerates it absent policy intervention.

### How to Adapt
- `/adapt/rights-protections` — most directly relevant. The paper provides theoretical backing for **policy intervention** (Pigouvian automation tax, retraining funded by it) over individual market remedies. Useful for the section advising workers on what to push for politically.
- `/adapt/career-transitions` — note the "demand cliff" risk: high-paid white-collar workers losing jobs reduces sector-wide demand. Worth mentioning when discussing realistic timelines.
- `/adapt/upskilling` — important caveat: the paper finds **upskilling alone cannot eliminate the externality**. Useful corrective to over-optimistic upskilling narratives.

## Related notes

- [[Anthropic Observed Exposure]] — the empirical measure of where the model's predicted displacement should appear first.
- [[Same Storm, Different Boats]] — early empirical evidence of the demand-side compositional shift the model warns about.
- [[IMF Skill Gaps SDN]] — policy contrast: this paper concludes upskilling alone cannot eliminate the externality; the IMF SDN argues training is the priority.

## Caveats & limitations

- Theoretical; no empirical estimates of α*, k, η for real economies.
- Assumes symmetric firms within a single sector — real economies have inter-sector spillovers that may diversify the demand base.
- Firm-side decision modelling does not include political/regulatory pushback as a self-correcting force.
- Pigouvian tax recommendation is conditional on accurate calibration of "uninternalised demand loss per task" — a hard empirical target.