- The real protection for many knowledge jobs has been organizational inertia, not a lack of AI capability, and that inertia is now weakening.
- Snap’s layoffs were notable because leadership explicitly credited AI (over 65% of new code generated by AI), breaking the usual “restructuring” euphemisms and potentially giving other CEOs cover to do the same.
- Multiple datasets suggest entry-level tech is collapsing faster than Anthropic’s initial 14% estimate, with steep drops in junior employment and postings alongside large layoff waves increasingly attributed to AI.
- Anthropic’s follow-up survey shows a U-shaped anxiety curve: people who gain the most productivity from AI are also highly worried about job loss because they can see how quickly their work can be compressed.
- “Scope expansion” and higher output demands mean fewer specialists and fewer new hires: existing workers absorb more tasks, headcount may not fall immediately, but entry-level opportunities shrink.
In March, Anthropic published data showing a 14% decline in entry-level hiring for AI-exposed occupations. I wrote about it at the time. The core argument: a 61-percentage-point gap existed between what AI could theoretically automate in computer and math occupations (94%) and what people were actually using it for (33%). That gap was the moat protecting most knowledge workers. Not their skills, not their irreplaceability, but organizational friction.
I wrote: "The moat protecting your job isn't your skill. It's organizational inertia. And inertia, by definition, is temporary."
Seven weeks later, the inertia is breaking.
Snap said the quiet part
On April 15, Snap laid off 1,000 people, 16% of its workforce. The CEO explicitly said AI now generates over 65% of the company's new code. This was the first time a major tech company connected layoffs directly to AI capability rather than hiding behind "restructuring" or "refocusing."
Snap's stock went up.
The significance isn't the layoff itself. Tech layoffs happen constantly. The significance is the stated reason. Every previous AI-related layoff used ambiguous framing designed to maintain plausible deniability. We're "refocusing on core priorities." We're "streamlining operations." Everyone understood the subtext, but nobody said it out loud because saying it out loud changes things.
Snap said: AI does the work these people did. We don't need them anymore. That's a fundamentally different kind of announcement. It breaks a taboo every other CEO was carefully maintaining, and it gives permission to every other CEO who was thinking the same thing but didn't want to be the first to say it publicly.
The entry-level collapse
Stanford's AI Index data shows employment for software developers aged 22 to 25 has fallen nearly 20% since 2024. Entry-level tech job postings in the US have dropped 67%. In the UK, tech graduate roles fell 46% in 2024, with projections for a further 53% drop by 2026. Recent graduate unemployment hit 5.7% in Q4 2025, worse than at any point during the 2008 financial crisis.
Anthropic's paper measured a 14% decline using Claude data alone. The broader market, reflecting all AI tools combined, shows the decline is steeper and accelerating.
Q1 2026: roughly 95,000 tech layoffs, nearly half attributed to AI. Oracle cut 10,000+. Amazon cut 16,000. Meta cut 8,000. In a 17-day window in April alone, 19,000 confirmed layoffs cited AI as a factor. The trend line stopped being gradual.
The paradox in Anthropic's follow-up
Anthropic surveyed 81,000 Claude users in April. The finding that should concern everyone: workers reporting the largest AI productivity gains are also the ones most worried about losing their jobs. A U-shaped curve. Workers slowed down by AI are anxious. Workers dramatically sped up by AI are also anxious, because they can see the implication: if the tool does my work this fast, how long until someone decides the tool doesn't need me?
Only 60% of early-career workers felt they personally benefited from AI, compared to 80% of senior workers. The people with the least job security are the least convinced the technology helps them.
48% of respondents reported "scope expansion," doing tasks outside their previous capability. A product manager building dashboards. A marketer writing SQL. Sounds positive for the individual. At the organizational level, scope expansion means fewer specialists needed. The jobs that would have existed for new entrants are being absorbed by people who already have jobs.
10% of respondents said their employers were using productivity gains to demand more output rather than reducing headcount. AI doesn't reduce workloads, it expands them. People aren't laid off. They're stretched. The headcount stays the same. The workload increases. The hiring of additional staff stops.
The moat is breaking
In the original paper, I described the 33% observed-vs-theoretical utilization as the key number. The gap between capability and adoption. That was the moat.
Two data points suggest the friction is decreasing faster than expected. Snap's 65% figure means one company has already crossed the midpoint where AI does more of the coding work than humans do. The 33% average across the economy obscures individual organizations that are much further along.
Entry-level posting declines of 67% show where friction disappears first. Not laying off existing employees (high friction: severance, morale, knowledge loss). Not restructuring teams (medium friction). Just not posting the next junior role. Nobody notices except the person who would have gotten the job.
The paper's scenario of a "Great Recession for white-collar workers" was framed as something "absolutely possible" if utilization moves from 33% to 66%. Some sectors are already there. All three indicators from Anthropic's early warning system, entry-level funnel, utilization rate, companies making the implicit explicit, moved in the same direction over seven weeks.
In March, the responsible takeaway was: start building immunity while you still have time. The thermometer showed a rising fever but the number wasn't scary yet.
The April update: the number is getting scary for specific populations. If you're a junior developer, a recent graduate in a knowledge work field, or in a role where your primary output is text or code that AI handles well, the timeline has shortened.
The gap between "could handle" and "does handle" isn't just shrinking. In some corners of the economy, it's already closed.
Assume the entry-level bottleneck is real and plan around it: build proof-of-work (a portfolio, shipped projects, measurable outcomes) that shows you can use AI to deliver end-to-end results, not just write code or complete isolated tasks. Watch for signals inside your company—hiring freezes, rising output expectations, and “AI-first” mandates—because they often precede role consolidation even without layoffs. If you’re early-career, prioritize roles and teams where you own a business metric and can expand scope safely, while continuously upgrading your AI workflow so you’re the person who directs the tools rather than the work the tools replace.