TL;DR — Simon Willison breaks down Arvind Narayanan & Sayash Kapoor’s Normal Tech essay: New York’s WARN Act data shows zero AI-related layoffs among 160+ companies in the first year of mandatory disclosure. The real bottlenecks in software engineering aren’t typing code — they’re deciding what to build, verifying it’s correct, and the deep human understanding that binds it all. Published June 14, 2026.
The evidence that should kill the replacement narrative
New York became the first U.S. state to require an AI-disclosure checkbox on WARN Act layoff filings. March 2025 – March 2026: 160+ companies filed WARN notices. Zero checked the AI box. Per Hunton Employment Labor Perspectives.
This isn’t a lagging indicator — it’s the hardest data we have. If AI were replacing engineers at scale, someone would have checked that box. Nobody did.
Supporting signal: The U.S. Bureau of Labor Statistics’ JOLTS data for the same period shows software developer job openings declining from a 2022 peak, but the quit rate (workers voluntarily leaving) also fell — suggesting stability, not displacement BLS JOLTS.
The three bottlenecks AI doesn’t touch
Narayanan & Kapoor identify three irreducible constraints that make mass replacement structurally unlikely:
1. Deciding & specifying what to build
Requirements gathering. Stakeholder alignment. Translating “we need something like Uber but for dog walkers” into actionable specifications. This is product strategy, not code generation. PMs, architects, and senior engineers own this — and AI assists but doesn’t replace the judgment calls.
2. Verifying & being accountable for what’s delivered
Code review. Security auditing. Testing strategy. Legal and professional accountability — engineers sign off on outcomes. When a medical device fails or a payment system leaks data, a human engineer’s license and reputation are on the line. An LLM has no skin in the game.
Real-world anchor: The 2024 CrowdStrike outage that crashed 8.5M Windows machines stemmed from a kernel driver update that skipped staged rollout testing — a process failure, not a code-generation failure. No AI agent would have caught the missing canary deploy. Microsoft Post-Incident Review.
3. Deep human understanding (the meta-bottleneck)
The deep human understanding — of the codebase, the business, and the environment — required to carry out both of these.
This encompasses:
– Codebase context: Historical decisions, technical debt, architectural nuances
– Business context: Domain knowledge, user needs, organizational constraints
– Environmental context: Infrastructure, compliance, operational realities
AI Insider take: I’ve watched this play out in my own work. AI writes the boilerplate beautifully. But when the production incident hits at 2 AM and the stack trace points to a decision I made six months ago, the fix comes from context the model never saw. This is why running local LLM coding agents still requires human oversight — the tool executes, you direct.
Why “coding” was never the bottleneck
| Activity | % of engineering time | AI acceleration |
|---|---|---|
| Writing implementation code | ~20-30% | High (2-5x) |
| Deciding what to build | ~20-30% | Low (assistive only) |
| Verifying / testing / review | ~25-35% | Medium (tooling helps) |
| Deep context / tacit knowledge | ~15-25% | Near zero |
The essay’s logic chain:
AI replaces coding → Coding isn't the bottleneck →
Real bottlenecks: (1) Specification, (2) Verification/Accountability, (3) Deep Contextual Understanding →
These require human judgment, liability, and tacit knowledge →
Mass replacement narrative unsupported by evidence
Simon Willison’s commentary: AI helps, but context is king
I’m finding AI assistance also helps me with the deciding and verifying steps, but it’s the “deep human understanding” that remains key to the value I provide. Give me all of the AI assistance in the world and the value I produce will still be reliant on how deeply I understand both the problems and the solutions that the agents are building for them.
This matches what I see in the field. The engineers who get 10x value from AI aren’t the ones asking “write me a React component” — they’re the ones saying “here’s the architectural constraint, here’s the business requirement, here’s the legacy weirdness, now help me reason through the tradeoffs.”
Beyond software: the regulatory cushion
| Profession | Regulatory Barriers | Automation Cushioning |
|---|---|---|
| Software Engineering | Very few | Baseline (still resilient) |
| Healthcare | Extreme | Even more cushioned |
| Law | Extreme | Even more cushioned |
| Finance | High | Even more cushioned |
| Education | High | Even more cushioned |
Given that this is true even in a sector with very few regulatory barriers, most other professions are likely to be even more cushioned.
If software — the most automatable, least regulated knowledge work — shows zero AI-driven displacement, the “AI will replace everyone in 3 years” crowd has a data problem.
What this means for your career
If you’re a junior engineer: The “write code fast” path is compressing. Invest in the judgment layer — system design, tradeoff analysis, asking “why this feature, why now, what breaks?”
If you’re a senior engineer: Your value increases. AI handles more implementation; your contextual judgment becomes scarcer and more valuable.
If you’re a manager/hiring: Stop optimizing for “coding speed.” Optimize for “specification clarity” and “verification rigor.” The bottleneck moved.
The counterargument (and why it’s thin)
“But what about Devin / Cursor / Claude Code agents that build entire features?”
They’re impressive demos on greenfield, well-specified tasks. Drop them into a 5-year-old codebase with half-documented APIs, regulatory constraints, and a product manager who changes requirements twice a week — they stall at the same bottlenecks: specification ambiguity and contextual understanding. For a reality check on what developers actually ship with these tools, see what developers build with AI coding agents.
“But capabilities are improving exponentially!”
They are. But the bottlenecks aren’t capability problems — they’re accountability and context problems. No amount of intelligence gives an AI legal liability or access to the unwritten tribal knowledge in your Slack history.
Signals to track
- NY WARN Act Year 2 data (March 2026 – March 2027) — will anyone check the box?
- State-level AI disclosure laws spreading — more data points coming
- Agentic AI benchmarks (e.g., NVIDIA AgentPerf) — measuring end-to-end task completion, not coding speed
Bottom Line
Verdict: The “AI replaces software engineers” narrative fails the evidence test. Zero WARN Act AI layoffs. Three structural bottlenecks (specification, verification/accountability, deep context) that require human judgment and liability. AI accelerates implementation — the easy part — while the valuable part remains stubbornly human. Don’t fear replacement; fear irrelevance if you ignore the judgment layer.
FAQ
Q: Did New York actually find zero AI-related layoffs?
A: Yes. From March 2025 to March 2026, 160+ companies filed WARN notices in New York. Not a single one checked the AI-disclosure box. Source: Hunton Employment Labor Perspectives.
Q: If AI codes 5x faster, why aren’t engineers being laid off?
A: Coding is only ~20-30% of engineering time. The bottlenecks — deciding what to build, verifying it works, and deep contextual understanding — aren’t accelerated by code generation.
Q: What about AI agents like Devin that build entire features?
A: They work on greenfield, well-specified tasks. In real codebases with legacy constraints and shifting requirements, they hit the same specification and context bottlenecks.
Q: Should junior engineers worry about their careers?
A: The “write code fast” path is compressing. Juniors should invest in system design, tradeoff analysis, and asking “why this feature?” — the judgment layer that AI can’t replace.
Internal links for deeper context
- What developers actually build with AI coding agents in 2026 — real use-case survey
- How to run local LLM coding agents with Ollama — self-hosted setup guide
- Open source AI models worth running in 2026 — model comparison
Source: Simon Willison, “Why AI Hasn’t Replaced Software Engineers, and Won’t,” published June 14, 2026. Primary analysis: Arvind Narayanan & Sayash Kapoor, “Why AI Hasn’t Replaced Software Engineers, and Won’t,” Normal Tech, published June 2026. NY WARN Act data: Hunton Employment Labor Perspectives, March 2026. BLS JOLTS data, June 2026. Microsoft CrowdStrike post-incident review, July 2024.
