We’ve all heard the reassurances over the last year. First, that AI won’t take your job; someone using AI will. More recently, that AI will handle task execution while humans supply the taste and judgment to direct it. Many of us have repeated these ideas to audiences of anxious employees, or to leadership teams weighing AI investments. It felt safe to say because the story felt familiar, like history repeating. But what I’ve started to notice is that the origins of that story may be leading us astray.
It came from software. The comfort we offer about AI is really a memory of what tools like spreadsheets, desktop publishing, and computer-aided design did over the last few decades. Software made execution faster, cheaper, and more accurate. A spreadsheet recalculates a thousand cells without a slip. A design tool renders in seconds a change that used to take hours or more. And in making execution cheap, the tools raised the value of human taste, judgment, and review. When the mechanical part gets easy, the discerning part gets valuable. That’s the premise behind “you will be freed to do the higher-value work.” It’s a software story.
The other history is not that story
There’s a second automation history that we tend to blur into the first. Physical and industrial automation, the machines and robotics that transformed manufacturing, did something different. It didn’t mostly elevate the workers whose tasks it absorbed. It displaced them. A machine that welds, sorts, or assembles runs without breaks, without fatigue, without a shift change, and once it’s good enough the economic logic isn’t partnership but replacement. The human roles that survived moved largely into maintenance and repair, keeping the machines running, rather than into higher judgment about the work the machines were doing.
So we’ve got two automation lineages. One made human judgment more valuable. One made human labor redundant and kept people around to service the equipment. We’ve been applying the first story to AI, confidently, because AI is software, and software is the lineage that elevated us. The comparison is understandable. It may also be a category error, because AI doesn’t sit cleanly in either history, and increasingly the two histories are converging as software merges with robotics.
The software wave wasn’t painless either
I want to acknowledge that the software story wasn’t all positive either, because that matters for where we are today. Software elevated judgment, but it also displaced a great many people along the way. Desktop publishing gutted typesetting, the skilled trade that set and arranged every printed page before software could do it. Spreadsheet software eliminated hundreds of thousands of traditional accounting clerk and bookkeeper jobs. Numerous other jobs have disappeared as the technology improved. The elevation of human judgment was real, and so was the disruption underneath it.
That’s the part we overlook when we make the software comparison. The promise that people move up is a promise about the category of work, not a guarantee about the individual worker. The software wave genuinely created more valuable human work and genuinely closed off career paths. We need both perspectives in mind as we consider what’s happening now.
The line that’s moving
So where are we now? Until very recently, the working division of labor with AI was familiar and reassuring, and it fit the software story. The AI executed. The human judged. The model drafted the content, generated the deck, produced the code, and a person supplied the taste that decided whether it was any good. Execution became faster, judgment stayed human, and the old promise seemed intact.
In the last several months that line’s been moving. AI systems are no longer just executing a task someone gives them. They’re increasingly doing the orchestration, deciding which steps to take in what order, delegating work. They’re generating options and then recommending among them. A word processor may have flagged your misspellings or poor sentence structure, but it never critiqued your argument or suggested an alternate framing. What we’re seeing now is a degree of autonomy that looks less like a tool waiting to be used and more like a junior colleague who proposes the plan, not just executes it. The boundary the software story depended on, that machines handle the doing and humans do the deciding, is now shifting.
Of course, software has been making consequential decisions for a long time. Credit scoring, loan approvals, and insurance risk assessment were all automated decades ago. But those systems needed a human to specify the decision procedure in advance: every variable, every weight, every threshold. The system could only consider what someone had thought to include, and it couldn’t handle anything outside those strict bounds. That constrained automation to narrow, predictable domains where the decision could be fully specified ahead of time. What’s changed is that while firms still train AI systems on their own methods, policies, and business context, the specification no longer has to be exhaustive. The system can extend past its training data to handle cases nobody documented and domains the older systems could never reach. It’s reaching into the judgment layer the software history promised would remain ours.
This time feels different
I don’t want to overstate this, because “this time is different” doesn’t have the greatest track record, and skepticism has often been merited. AI isn’t better than human judgment at everything. In many domains it isn’t even close, and despite rapid advances it may not be for some time. Anyone claiming certainty about the timeline, in either direction, is likely selling something.
But the operative word in “not as good yet” is yet, and the rapid pace of AI advancement makes me think it’d be foolish to wave the “yet” away. The progress isn’t arriving on one front. It’s arriving on several at once: not only larger language models but new computing architectures and new chip technologies, and the convergence of AI with the physical machines of the other automation history. When advances compound across many directions simultaneously, our previous experience with any single technology becomes a much less reliable roadmap for where we’re headed. It’s hard not to feel uncertain about how long the reassuring division of labor holds.
Missing and mixed signals
Another common story we’ve heard about new technologies is that they may eliminate a lot of jobs, but they create new ones for people to move into. Many people today make a living doing work previous generations wouldn’t have recognized. New career creation is documented history.
That doesn’t mean new types of work arrive on the same timeline as the displacement, or that they affect the same people. And with AI specifically, it’s early. We don’t yet see the mass displacement some have predicted, nor have we seen a wave of genuinely new AI careers. What we do see is rising demand for some existing roles, for example major technology companies funding training programs for the skilled trades needed to build data centers, along with AI specializations layered onto existing roles, like data science and engineering.
There’s one early signal getting a lot of attention. A Stanford Digital Economy Lab study of ADP payroll data found no widespread displacement, but did find reduced hiring among 22-25 year-olds in the most AI-exposed occupations. The authors don’t claim causation, and among its caveats the study is based on one payroll provider’s records rather than the whole labor market. Still, it points to a different outcome than the one we’ve been bracing for. Instead of people starting at the bottom of the career ladder and climbing, the bottom rung itself may be getting pulled up out of reach, or at least far enough that it’s a lot harder to get started.
What we climb into
The promise of the software wave was simple and specific. By making execution cheap, it made human judgment more valuable, and judgment is the rung people climbed onto when the rung below them was automated away. That’s the move the comfortable story describes. You lose the task, you keep the discernment, you step up.
The long view I’ve been circling is what that move looks like when the rung being automated is judgment itself. Every prior wave left a higher place for people to stand. If AI is the software that reaches into the deciding and not only the doing, then the reassurance we keep reciting reflects a history that may no longer apply. And if the bottom rung is moving out of reach at the same time, the pressure is squeezing both ends of the ladder at once. So I don’t see a tidy answer for what people move up into when the role we’re supposedly climbing toward is one the technology is also learning to fill.
Sources
- Erik Brynjolfsson, Bharat Chandar and Ruyu Chen, “Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence,” Stanford Digital Economy Lab, August 2026 revision.