The story that AI will trigger mass layoffs in software engineering keeps getting louder. The data tells a different one. According to Simon Willison, who highlights a new essay from researchers Arvind Narayanan and Sayash Kapoor, there’s now enough evidence to reject the idea that once AI hits a certain capability threshold, the layoffs follow. What stands out here is the logic of their test case: they picked software engineering on purpose, because it’s the profession most exposed to AI disruption and has almost no regulatory walls protecting it. If the narrative breaks down even here, it breaks down nearly everywhere.
Let me sort the myths from what’s actually happening.
Myth 1: AI is already causing mass tech unemployment
The paper points to a clean piece of evidence. In March 2025, New York became the first U.S. state to add an AI disclosure checkbox to its WARN Act filings, the notices companies must file before large layoffs. In the full first year, more than 160 companies filed WARN notices. Not a single one checked the AI box.
That’s not proof AI has zero effect on hiring. But it’s a direct shot at the claim that AI is the engine behind tech job cuts right now. Companies cutting staff are naming other reasons.
Myth 2: Writing code is the job
Here’s the core misread. AI is genuinely fast at the typing-code-into-a-computer part. The mistake is assuming that part is the job. Narayanan and Kapoor argue software engineering is about a whole lot more than producing lines of code.
When they dug into what engineers actually spend time on, the surveys pointed to meetings and debugging, not raw coding. So the real question becomes: what’s happening in those meetings, and why can’t AI do it?
Myth 3: As models improve, the bottlenecks disappear
This is where the analysis gets useful. The authors identify three bottlenecks that resist automation:
- Deciding and specifying what to build. Turning vague business needs into a precise spec is judgment work, not typing.
- Verifying and being accountable for what’s delivered. Someone has to own whether the thing actually works and is safe to ship.
- Deep human understanding. Knowing the codebase, the business, and the environment well enough to do the first two.
More capable models speed up the typing. They don’t hand you accountability or context. Willison, who writes from years of hands-on AI-assisted development, adds a sharp personal note: AI now helps him with the deciding and verifying steps too. But the deep understanding is what still drives his value. As he puts it, give him all the AI assistance in the world and his output still depends on how deeply he understands both the problems and the solutions the agents are building.
Why this matters now
The timing isn’t academic. Layoff announcements get framed as AI efficiency, hiring freezes get blamed on automation, and the narrative shapes real budget and career decisions. This essay gives practitioners and executives a grounded counterweight built on filings and task-level analysis rather than vibes.
There’s also a broader read. The authors note that if software engineering, a low-regulation, high-exposure field, is this cushioned, most other professions are likely to be even more protected. Doctors, lawyers, and accountants carry licensing, liability, and accountability structures that slow displacement further.
What to do with this
- For engineers: Lean into the parts AI can’t own. Specification, verification, and context are where your value concentrates. Treat agents as fast typists you direct, not replacements you fear.
- For managers: Stop budgeting for headcount cuts on the assumption that AI replaces engineers wholesale. The bottlenecks it doesn’t solve still need humans who understand your systems.
- For everyone selling the layoff story: Check the filings. So far they don’t back it up.
The honest version of the future is less dramatic than the headlines. AI changes how engineers work without removing the need for them. You can read the full essay from Narayanan and Kapoor, with Willison’s commentary, at the original source.