In March 2026, Goldman Sachs updated its labor displacement model and walked back some of the more alarming projections from its 2023 report. The revision got much less press coverage than the original. That pattern keeps repeating: dramatic AI job displacement headlines, quiet corrections, rinse, repeat.
This doesn’t mean the concern is wrong. It means the picture is more uneven and slower-moving than a lot of the coverage suggests, and that the unevenness matters a lot for what you actually do about it.
What the employment data shows so far
The Bureau of Labor Statistics Occupational Outlook data through 2025 shows net job growth in most technology categories. Software developer roles are projected to grow 25% through 2032. Data entry clerk and some paralegal roles are declining. The displacement is real, but it’s concentrated in specific categories, not spread uniformly across knowledge work.
The LinkedIn Economic Graph research team published analysis in 2024 showing that roles requiring AI skills grew 2.5x faster than roles that didn’t, and that upskilling into AI tooling was the most common path out of declining role categories. Whether that rate holds into 2026 is uncertain, the model is still evolving.
Many people in roles adjacent to code generation (junior copywriters, certain data analyst functions, content operations) are finding real compression. Many are not. This is not universal. It is a general experience that a specific subset of knowledge workers are having, and it’s worth being honest about which subset that is.
Where displacement is actually happening
The clearest examples are in high-volume, rule-based cognitive work. Document summarization that used to require a junior analyst. First-draft legal research that used to go to a paralegal. Tier-1 customer support where scripts and FAQs cover 80% of cases. These functions have seen real headcount compression, and the companies doing the compressing are generally not hiding it.
Klarna’s CEO Sebastian Siemiatkowski said publicly in 2024 that their AI assistant was doing the work of 700 customer support agents. Duolingo cut a significant portion of its contractor translation workforce in late 2023. These aren’t rumors; they’re documented in earnings calls and trade press.
The pattern in both cases is the same: high-volume work with predictable inputs and measurable outputs. That’s where AI tools are genuinely faster and cheaper than human labor right now.
Where displacement is slower than the headlines suggest
Roles that involve building trust with specific people, sales relationships, clinical care, legal counsel, engineering management, are moving much more slowly. This isn’t because AI can’t do the task at an abstract level; it’s because clients, patients, and colleagues still largely want a human they can hold accountable.
There’s also the coordination overhead problem. Integrating AI tools into existing workflows requires people who understand both the business process and the AI capability. That’s not a small thing. A lot of companies have discovered that deploying a capable AI tool doesn’t automatically translate into productivity gains without humans managing the transitions. That work creates its own demand.
If you’d asked me in 2023 whether software engineers would be largely displaced by 2026, I’d have said that seemed very aggressive. I still think that was right, though the pace of code generation improvement in the last 18 months has made me less confident than I was.
The skills that are actually holding value
System-level thinking. Knowing that a function works is different from knowing where it sits in a distributed system, what it costs at scale, and what breaks if it’s slow. AI tools are not good at this yet.
Judgment under uncertainty. Deciding which direction to go when you have incomplete information and real stakes. This is the work that senior engineers, product managers, and executives do most of the time, and it doesn’t compress easily.
Explaining technical decisions to non-technical people. This sounds mundane. It’s one of the most durable skills in the current market, in part because AI-generated explanations tend to be generic and because trust in explanations is relational, not just informational.
What this means practically for job seekers
The interview market in 2026 is harder than 2021 and easier than 2023. Hiring is happening, but it’s concentrated in teams that are actively using AI tools, not hiring people to do work that AI tools now handle. The question “can you use AI tools effectively” has moved from a nice-to-have to a basic expectation in many technical roles.
For people actively interviewing, this shows up in how technical interviews are structured. More companies are asking how you’d approach a problem with AI assistance, not as a way to evaluate the AI, but as a way to evaluate your judgment about when to use it and when not to.
Craqly is built on the premise that the interview itself is a place where AI assistance, used well, changes outcomes. That’s a real bet on a specific part of the picture. Whether it reflects the broader trend in your field depends on what you do and who you’re interviewing with.
The uncomfortable part of this conversation
The people most likely to be reading articles like this one are probably not in the most affected categories. The roles with the highest displacement risk, data entry, certain content operations, basic document processing, are not roles where people spend Friday afternoons reading tech blogs about AI trends.
That gap between who reads the displacement coverage and who experiences the displacement is worth holding onto. The anxiety in tech circles about AI replacing knowledge workers is often more diffuse and anticipatory than it is immediate. That’s not a reason to ignore it. It’s just a reason to be specific about what’s actually happening, to whom, and on what timeline.