"Why AI hasn't replaced software engineers, and won't" is an essay published June 10, 2026 by Arvind Narayanan and Sayash Kapoor (Princeton) on AI as Normal Technology (normaltech.ai), the successor venue to their AI Snake Oil writing. Subtitled "Coding agents as normal technology," it is the first in a planned series and extends the authors' normal-technology framing to the labor question, using software engineering as the test case because AI capability and adoption are furthest along there. The essay argues that the evidence is already sufficient to reject the thesis that AI capability, once past a threshold, causes mass layoffs, and that this holds even in a sector with few regulatory barriers (Source: normaltech.ai).
Summary of argument
The essay makes two linked claims: empirically, that reported AI-driven software layoffs are largely mislabeled, and mechanistically, that the structure of software work resists automation of the kind that would displace engineers. It organizes knowledge work as a "decide-execute-deliver sandwich" (with deep human understanding of the codebase, business, and environment as a prerequisite for all three layers). AI compresses the middle "execute" layer but leaves the two outer layers — deciding and specifying what to build, and verifying and being accountable for what is delivered — largely intact, a limit the authors argue capability improvements alone will not overcome.
Key claims
- "AI washing" of layoffs. The authors characterize prominent AI-attributed software layoffs as mislabeled. Block's February 2026 cut of 4,000 (founder Jack Dorsey citing AI "enabling a new way of working") followed pandemic-era tripling of headcount and financial pressure; a Cash App data scientist described "very limited gains in productivity." Snap's roughly 1,000 April layoffs (CEO Evan Spiegel citing AI and "65% of new code") followed an activist-investor (Irenic) cost-cut campaign, and the cuts did not concentrate in AI-exposed roles. Intuit's May cut of 3,000 was described by its CEO as having "none of it had to do with AI," targeting "coordination-heavy roles."
- Survey evidence of mislabeling. A resumetemplates.com survey found 59% of U.S. hiring managers emphasize AI when explaining layoffs because it "plays better with stakeholders." A Harvard Business Review survey of over 1,000 executives found 21% made large headcount cuts "in anticipation of" AI versus only 2% citing actual AI implementation — a roughly tenfold gap. Forrester's J. P. Gownder reported that nine in ten companies preparing supposedly AI-driven layoffs had no mature AI application ready to fill the roles.
- WARN-act disclosure data. In the first year of New York's WARN-act AI-disclosure checkbox (added March 2025; the first U.S. state to do so), only one company — Nespresso — checked it, accounting for 46 of roughly 25,000 laid-off workers, about two-tenths of one percent.
- Layoffs are the wrong signal. Citing research that AI's effect operates "through slower hiring rather than increased separations" (SSRN 5425555), the authors argue firing destroys tacit knowledge and is costly relative to natural turnover.
- Employment still growing. A Federal Reserve paper ("AI and coder employment") finds U.S. software employment still growing post-ChatGPT, about 3 percentage points per year slower than a no-AI counterfactual; the authors note the methodology cannot capture self-employment and argue entrepreneurship may absorb part of the slowdown.
- The sandwich model. The "AI-written-code" percentage is described as disconnected from labor displacement because writing code was never the bottleneck — a 2019 Microsoft Research study found developers spend 9–61% of time coding. The real bottlenecks are deciding/specifying, verifying/accountability, and the deep understanding both require. As AI absorbs lower-level decisions, the authors argue the value of human decision-making migrates upward rather than the "decide" layer thinning.
- Writing code vs. shipping code. An NBER working paper (w35275) across 100,000 GitHub developers found AI agents produced an eight-fold increase in lines of code but only 30% more releases, which the authors read as evidence that the Decide and Deliver bottlenecks remain.
- Vibe coding is not agentic engineering. The essay distinguishes "vibe coding" (the user does not supervise, review, or evaluate the agent's output) from "agentic engineering" (a term attributed to Simon Willison; the human stays in control and accountable). It cites SWE-chat (arXiv:2604.20779), a logged dataset finding only 44% of agent-produced code survives into user commits, that vibe-coded commits introduce vulnerabilities at nine times the human-only rate, and that the most common user intent is understanding existing code (19%) rather than generating new code (13%).
- Demand may rise. Invoking price elasticity (and, loosely, Jevons' paradox), the authors argue cheaper software increases demand for software and therefore derived demand for engineers, noting programmer employment grew from near zero around 1950 to millions. They distinguish this from past democratization claims for FORTRAN, COBOL, and SQL that did not eliminate the profession, arguing the barrier is skilled judgment and accountability rather than syntax.
- Real but different job losses. The authors acknowledge AI can reduce demand for a product (Chegg, Stack Overflow) or prompt restructuring at firms that sell AI (IBM, SAP), framing both as distinct from AI replacing engineers.
The essay closes by invoking Fred Brooks's "No Silver Bullet" (essential versus accidental complexity) as a 40-year-old articulation of why the "decide" layer resists automation, and previews a follow-on essay arguing that aggregate demand can stay healthy even as individual engineers' careers are disrupted by firm type, geography, seniority, and pace of adaptation.
Reception
The essay's empirical findings on the New York WARN checkbox, Nespresso, and the Federal Reserve growth figures were reported contemporaneously and corroborate items in the developments record. As an op-ed advancing the authors' normal-technology position, its forecasts (that mass layoffs "won't" come and that demand may rise) are arguments rather than established findings; the authors themselves note data gaps, including the difficulty of detecting upticks in poorly supervised high-stakes AI use.
Provenance
Authored by Arvind Narayanan and Sayash Kapoor; published June 10, 2026 on normaltech.ai. Foundational essay advancing an original framework (the decide-execute-deliver sandwich) and named distinction (vibe coding versus agentic engineering, the latter attributed to Simon Willison). Source authenticity verified at gap-scan (2026-06-12) against the fetched text and the authors' bylines.
Relationships
- depends-on: AI as Normal Technology — extends the authors' normal-technology framing to labor.
- contradicts: The Adolescence of Technology — rejects the AI-driven-mass-layoffs trajectory advanced by Dario Amodei.
- related: AI Labor Disruption — primary concept the essay bears on.
- related: AI Coding Agents — the sandwich model and vibe-coding/agentic distinction.
- related: AI Displacement vs. Augmentation — augmentation-over-displacement reading.