As artificial intelligence lowers the activation energy for starting projects, the bottleneck shifts from initiation to completion. The result isn't less work but dramatically more of it. This is the Jevons Paradox manifesting in knowledge work, flooding academic journals, overwhelming review systems, and paradoxically increasing workload despite efficiency gains.
The nineteenth-century economist William Stanley Jevons observed that more efficient steam engines didn't reduce coal consumption—they increased it. Cheaper steam power spurred new applications, driving overall coal use upward. This counterintuitive dynamic predicted that efficiency gains in resource use often trigger increased consumption rather than conservation. If AI makes knowledge workers five times more efficient, economic theory suggests we won't need fewer workers—we'll find five times more work to do.
Academic publishing provides the starkest evidence. An AI Task Force convened by the journal Organization Science documented a dramatic shift starting in 2023, after ChatGPT's release. Submission volume surged while the percentage of submissions using minimal AI plummeted from nearly 100% to closer to 30%. Submissions became far harder to read—longer words, more complex sentence structures, more jargon, more nominalizations—despite assumptions that AI produces cleaner text.
The quality decline carries consequences. Organization Science now desk-rejects nearly 70% of manuscripts showing heavy AI use, compared to 44% for papers written without AI assistance. Only 3.2% of high-AI papers ultimately gain acceptance, versus 12% of low-AI papers. Critically, editors making these decisions don't know which papers used AI—the patterns emerge in retrospective analysis. Generative AI makes individual researchers' lives easier in the moment while producing worse outcomes for the field, taxing the time and patience of the review community.
The mechanism becomes clear when examining workflow structure. In most knowledge work teams, only about 20% of effort involves individual production—writing code, drafting documents, creating designs, running analyses. The remaining 80% is collaboration: reviews and approvals, alignment meetings, decision-making, validation, waiting for partners to react. AI accelerates that 20%. A design spec that once required a day now takes an hour; a feature that demanded two days of coding comes together before lunch. But that extra output still flows through the same human gates.
Work in progress piles up at bottlenecks. Queues lengthen for reviewers, approvers, and decision-makers. Human throughput at those gates drops as cognitive overload, context switching, and fatigue slow reviews or push people to rubber-stamp—creating what insiders call slop and necessitating rework. Activity output metrics spike—messages, documents, tasks—while outcome metrics like shipped value or customer satisfaction barely move or even deteriorate. Rather than speeding up a slow system, AI overloads the parts already strained.
This coordination tax hits large enterprises hardest. Before AI, most big companies already operated with low execution costs but high coordination costs from complex stakeholder maps, compliance controls, cross-functional dependencies, and formal governance. In that environment, AI amplifies complexity and decision load. Teams spin up more artifacts with fewer shared standards; reviewers face longer queues and harder calls. Small misalignments multiply across functions, introducing outdated or conflicting information into institutional knowledge. Atlassian estimates this AI fragmentation tax costs the Fortune 500 roughly $161 billion annually.
When AI accelerates document drafting but human judgment gates remain unchanged, the result is backlogs, not breakthroughs.
The theory of constraints, developed by physicist-turned-management-guru Eliyahu Goldratt in the 1980s, explains why efficiency gains fail to translate into productivity improvements. Goldratt argued that every system has a limiting factor or bottleneck, and focusing improvement efforts anywhere except that constraint rarely increases overall throughput. In a chicken-coop assembly line, speeding up frame-building when roofing is the slowest step just creates piles of partially-constructed coops at the roofing station. When AI accelerates document drafting but human judgment gates remain unchanged, the result is backlogs, not breakthroughs.
This explains why predictions of AI-driven job automation have failed to materialize. Anthropic CEO Dario Amodei initially suggested AI would automate half of entry-level white-collar jobs; Microsoft AI CEO Mustafa Suleyman predicted human-level performance on most professional tasks within twelve to eighteen months. Yet Sam Altman recently expressed delight at being wrong about a jobs apocalypse, and Amodei now says AI won't replace jobs but will replace large parts of existing jobs. The entry-level job market that media outlets confidently declared wrecked by AI automation is now rebounding, with hiring projected to rise significantly.
What's actually happening differs markedly from industrial automation analogies. Small business owners interviewed about their AI deployment describe using it to hack together quick-and-dirty tools solving specific operational problems—managing timesheets, tracking inventory—rather than replacing entire job functions. This pattern resembles a high-school programmer creating custom web applications more than a power loom making textile workers superfluous. AI reduces friction for discrete tasks while leaving the broader work structure intact, a dynamic one researcher termed freestyle work. The technology is impacting the knowledge sector in ways both weirder and less dire than predicted.
The human cost of this proliferation manifests as intensified rather than reduced workload. When barriers to production fall while quality gates remain fixed, the systems governing sustained attention and decision-making face increased demands. Starting projects requires less activation energy with AI assistance, but completing them to standard requires the same or greater reserves of focused attention, judgment, and iterative refinement. The result: more projects in flight, more cognitive load in parallel, greater demands on workers managing multiple workstreams simultaneously.
This reframes what productivity improvement actually requires. Digital tools that speed communication or enable rapid content generation don't automatically improve performance if they don't address the true constraint. As Goldratt's theory predicts, optimizing non-bottleneck steps just creates pile-ups and distraction without boosting real productivity. The relevant question becomes not how to generate output faster, but how to identify and support the deep steps that actually move the needle—the human judgment, synthesis, and refinement that constitute genuine value creation rather than mere activity generation.
The path forward requires treating AI as a team coordination challenge rather than individual productivity enhancement. Organizations that have reduced their fragmentation tax by nearly half do so by embedding AI into shared workflows with explicit quality standards, rather than deploying it as isolated personal assistants. This means designing systems where AI accelerates bottleneck activities—review cycles, decision-making, validation—not just input creation. It means recognizing that efficiency gains triggering volume increases without throughput improvements impose real costs: reviewer burnout, quality degradation, coordination overhead, and the energy depletion of workers managing larger queues of incomplete work.
Brandon Sanderson, speaking about AI-generated art, offered a perspective worth extending to knowledge work generally: we have agency in defining what we value. Machines can generate manuscript after manuscript, but all we have to do is say no. The question isn't whether AI can produce more—it demonstrably can—but whether more production serves human ends. Sometimes there really is no shortcut to taking your time. The challenge ahead is building systems that recognize energy and attention as the true scarce resources, designing workflows where efficiency gains reduce demands on those resources rather than merely enabling their faster depletion through increased volume.