A sustainability organization recently asked me to review a grant narrative an AI system had drafted for a major environmental foundation. The writing was clean. The structure was sound. The mission framing was compelling. It hit every element in the funder’s guidelines.
It was also going to get rejected before the program officer finished the first page.
The narrative was built around the organization’s carbon reduction commitments. Impressive numbers. Well presented. All completely undermined by one invisible error that a non-expert would never catch: The AI had conflated Scope Two and Scope Three emissions in the same impact argument.
For anyone outside sustainability reporting, that mistake sounds technical and minor. Inside the field. However, this is the equivalent of submitting a financial audit that confuses revenue with profit. Scope Two covers purchased energy emissions — electricity, heating, cooling. Scope Three covers the entire value chain — supply chain, business travel, product use and disposal. They’re measured differently, reported differently and matter differently to different funders.
Mixing them in a single impact claim doesn’t just weaken the narrative. It signals to any reviewer with real sustainability credentials that nobody who understood the field has touched the document.
The organization hadn’t noticed. Why would they? The writing looked authoritative. The numbers were plausible. The format was perfect.
That’s exactly the problem.
How Is AI Changing the Job Market for Knowledge Workers?
Generative AI is not displacing true expertise, but it is eliminating the market value of the appearance of expertise. While AI can rapidly produce authoritative formats and plausible data, it lacks genuine domain judgment. As a result, the labor market is experiencing a reinstatement effect, shifting demand away from superficial tasks toward high-depth roles like AI systems architects, LLM output editors and integration consultants that require deep human domain knowledge.
Everyone Is Having the Wrong Argument About AI
The public debate about AI and jobs is stuck in a simple frame. One machine replaces one person. Or one job category disappears. That framing isn’t exactly wrong, but it isn’t deep enough to be useful.
Economists Daron Acemoglu and Pascual Restrepo studied automation’s effect on labor markets carefully in their 2018 NBER working paper on robots and U.S. labor markets. Their findings are more nuanced than either the doom camp or the optimism camp wants to admit.
Automation has a displacement effect. That part is real. The Luddites weren’t wrong. Historically, mechanization displaced farm laborers. It eliminated switchboard operators. It hollowed out entire occupational categories that had employed millions.
But automation also has a reinstatement effect. Technology creates new tasks where human labor still has an advantage. When AT&T mechanized telephone switching between 1920 and 1940, operator jobs fell sharply. But the workers who came after this shift didn’t end up facing lower overall employment. Demand rose in clerical and service work. New tasks appeared that the old system had never required.
The serious question was never whether one machine replaced one person. It was always whether society was creating new tasks fast enough for people to move into them.
We’re in that gap right now with AI. The pattern is already visible.
What Has AI Actually Displaced?
AI didn’t displace expertise. It displaced the appearance of expertise.
In tech specifically, the gap between an output looking right and actually being right was genuinely hard to see from the outside. A product spec that hit the expected format. A technical brief that referenced the right frameworks. An architecture recommendation that sounded authoritative. A content strategy built around the right vocabulary.
AI produces all of those things now, and it does so faster, cheaper and at scale. But it also works without domain knowledge, judgment or any understanding of whether the output it’s generating is actually correct. What generative AI commoditized overnight was not the ability to do the work. It was the ability to look like you were doing it.
That’s what the meeting I described earlier revealed. It’s not that AI is dangerous. It’s that the performance of expertise — which had been a viable career strategy in knowledge work for a long time — just lost its market value entirely.
The people who built genuine depth are fine. They’re using AI to do work they already understand faster. The people who were coasting on looking like they had expertise are discovering that AI does that particular thing better and cheaper than they ever could.
What Is the Reinstatement Effect?
Here is what the job displacement narrative consistently misses. Before the internet economy, there were no search marketers. Before cloud computing, there were no DevOps engineers. Before mobile technology, there were no app developers. That work didn’t exist because the surrounding systems didn’t exist yet.
AI is building a surrounding system right now, and new tasks are already appearing. Most of them don’t have settled job titles yet. But the shapes are becoming clear.
AI Integration Consultant
This role isn’t about implementing AI tools. Instead, it means understanding how AI connects to existing systems, data structures, organizational processes and human decision points in a specific organization with specific constraints. This requires deep knowledge of both the technology itself and the business domain simultaneously. When this work fails, failure is immediately visible, so you can’t fake expertise here.
LLM Output Editor
Rather than proofreading, this means applying domain expertise to AI-generated output to identify where it’s wrong, where it’s hallucinating, where it sounds authoritative but isn’t and where the reasoning has gone subtly off track in ways that are invisible to a non-expert reader. You can’t do this work in a field you don’t understand deeply. The role exists precisely because AI produces confident output regardless of whether it is correct.
AI Ethics and Governance Practitioner
This is about constructing the accountability layer that organizations are forced to build as AI moves into consequential decisions. Who is responsible when an AI system produces a harmful output? How is bias identified, documented and addressed? What does the audit trail look like for a regulator or a board? This requires legal, organizational and technical literacy simultaneously.
AI Systems Architect
This role involves designing the full structure of how an organization uses AI across its operations. Not individual tools in isolation, but the entire system — data inputs, model selection, human review points, output validation, feedback loops, failure modes. This requires imagining the complete picture simultaneously and understanding both the technology and the business deeply enough to know where they interact badly.
None of these roles existed clearly three years ago. All of them require genuine expertise that cannot be pattern-matched or prompted into existence.
What This Means for Tech Professionals Right Now
The reinstatement effect is real. But it’s not universal, and it doesn’t reward everyone equally.
It rewards people who built something genuine underneath the surface. People who have gone deep enough in their disciplines to know where AI is right and where it is confidently wrong. People who can sit in a meeting, look at an AI-generated architecture recommendation and identify the three specific ways it is wrong before anyone else in the room has finished reading it.
That is not a skill you can acquire by learning to use AI better. Rather, it’s a skill you build by doing real work, badly, repeatedly, over time, until the judgment is yours and not borrowed from a tool.
Using AI as infrastructure for work you already understand deeply is a force multiplier. Using AI to produce work you do not understand is a debt that comes due the moment someone in the room actually knows the domain.
The tech industry has always rewarded people who build real things. The difference now is that the gap between building real things and looking like you’re building real things is no longer easy to hide.
AI didn’t create that gap. It just turned on the lights.
