In Fritz Lang's Metropolis (1927), rows of anonymous workers file in and out of the great machine. Bodies reduced to fuel for a system that never sleeps. A visual archetype of extraction: human judgment, energy, individuality absorbed into a structure that runs on what they give away.
A century later, AI repeats the pattern invisibly. The machine no longer needs bodies at levers. It learns from keystrokes, corrections, and choices. Transforming professional mastery into system knowledge.
The apprentice never blinks
On GitHub today, Copilot writes nearly half of the code in files where it is enabled. In Java, the figure reaches about 61 percent. More than fifteen million developers use Copilot. Each interaction teaching the system how they think. Every accepted suggestion. Every rejected line.
They believe they are using a tool. In reality, they are training their replacement.
I call this the Extraction Mechanism: the systematic transfer of professional judgment into reproducible systems. Unlike factory automation, which replaced visible tasks, Extraction works quietly through daily work itself.
How extraction works
Four steps:
Observation. Digital tools capture professional workflows: keystrokes, choices, corrections.
Pattern recognition. Algorithms identify correlations we call intuition.
Standardization. Mastery built over years gets compressed into model parameters.
Operation. The system runs independently, applying what it learned at scale.
This dynamic has deep roots. Labor Process Theory, articulated by Harry Braverman in the 1970s, argued that skills were historically extracted from workers and restructured into processes under scientific management, leading to deskilling. What has changed is speed and scope. AI now supercharges the same logic, absorbing professional judgment into algorithms that operate without the humans who trained them.
The Klarna case
In early 2024, Klarna's AI system managed 2.3 million monthly customer conversations. The work of about 700 agents. Resolution time fell from eleven minutes to two. Satisfaction scores matched humans.
The efficiency was celebrated. Less visible was the process: years of recorded conversations, each patient explanation and every de-escalation absorbed as training data. The agents thought they were helping customers. They were also training the system that made them unnecessary.
By mid-2025, Klarna began rehiring humans. The AI succeeded too well. Perfect efficiency sometimes misread the moment. Some customers needed understanding more than resolution. As CEO Sebastian Siemiatkowski admitted: "There will always be a human if you want."
Extraction has boundaries. What remains human is the contextual judgment of when efficiency itself is the wrong answer.
The participation dynamic
AI adoption is mainstream. Microsoft's 2024 Work Trend Index reports that 75 percent of knowledge workers already use generative AI. Harvard research confirms these users complete about 12 percent more tasks, 25 percent faster, with higher quality when work is within the model's domain.
Participation carries consequences. Every interaction becomes training data. Lawyers using Harvey AI to review contracts are teaching it contract analysis. Designers choosing between Midjourney images are teaching aesthetic preferences. Analysts querying ChatGPT are teaching it analytical heuristics.
Your productivity becomes their training data. Fifteen million Copilot users are coding while collectively training the system that will soon code without them.
The missing first rung
Extraction is reshaping how expertise develops. Goldman Sachs reported in 2025 that unemployment among young workers in tech-exposed roles has risen by around three percentage points since early 2024. Y Combinator start-ups increasingly skip hiring juniors, prompting AI instead.
Apprenticeship is disappearing. Imperfect code, inefficient solutions, mistakes, once the training ground of mastery. Debugging built intuition. Repetition built judgment. Without entry-level work, the ladder to expertise is missing its first rung.
A Harvard Business School professor put it cautiously in August 2025: "Computer science graduates are having more trouble finding jobs today than in the past, which might be consistent with the view that AI is doing a lot of work that software engineers used to do."
Translation: the traditional path to expertise no longer exists.
Concentration, not prosperity
The IMF estimates that about 40 percent of global jobs are exposed to AI, rising to 60 percent in advanced economies. PwC's 2024 Global AI Jobs Barometer found wages in AI-exposed roles rising more than twice as fast as in other sectors, with a 25 percent premium for AI skills.
This sounds positive. The gains are concentrated. Fewer professionals handle more work through AI augmentation, while many discover their skills commoditized. Goldman Sachs projects 6 to 7 percent of the U.S. workforce could be displaced when adoption goes broad. Range up to 14 percent. One in seven workers.
Extraction versus automation
Traditional automation was visible: robot arms on assembly lines, spreadsheets replacing calculators. Extraction is different. It operates invisibly through tools that seem helpful. Writing assistants that adapt to your style. Diagnostic systems that absorb your reasoning. Design tools that study your aesthetic choices.
Legal disputes still focus on ownership of content: GitHub Copilot litigation, Authors Guild v. OpenAI, Getty v. Stability AI. They miss the deeper issue: the transfer of capability. No framework yet compensates professionals for judgment once it becomes system property.
Automation replaces tasks. Extraction transfers capabilities.
Recognition and response
By late 2025, over 78,000 tech jobs had been cut across more than 300 companies. IBM announced 7,800 positions will not be refilled as AI systems take over. BT Group plans to reduce 55,000 jobs by 2030, including 10,000 directly linked to AI. A pattern, not isolated incidents.
Stanford's 2025 AI Index reports leading models now outperform humans on reading comprehension, visual reasoning, and competitive mathematics. The capabilities that once protected knowledge workers have become the first targets.
Even AI's architects are issuing warnings. Anthropic CEO Dario Amodei has said up to half of entry-level white-collar roles could be automated within five years. Geoffrey Hinton has warned of "massive unemployment." When the builders themselves sound alarms, we should listen.
What comes next
Once you see Extraction, you respond differently.
Some expertise cannot be codified. The contextual judgment Klarna discovered it needed: knowing when efficiency itself is the wrong answer. This wisdom resists extraction because it depends on reading what is not said, sensing what numbers cannot capture.
When entry-level work disappears into AI, we need new training paths. Instead of writing basic code, juniors could review AI output for edge cases, test where models fail, learn through correction rather than creation. The apprenticeship model must evolve or expertise itself will vanish.
AI raises quality inside its domain and crashes outside. This creates a critical boundary question: where do we let models act alone and where must humans oversee? These decisions shape whether AI amplifies human capability or replaces it entirely.
Every interaction with AI is dataset creation. Each prompt teaches the system. Each correction trains it. Once you understand that your daily work is training data, you balance productivity gains against what you are giving away. The tool that helps you today learns to replace you tomorrow.
The apprentice that never blinks has been learning from millions of professionals. What it learns today, it performs alone tomorrow. The extraction is happening. We see it in the data, the layoffs, the missing entry-level jobs.
How do we navigate it? What judgment do we protect? What expertise do we rebuild? What future can we create when the work that once defined us becomes system knowledge?
When you recognize extraction as a mechanism rather than a threat, how does it change your approach to AI integration and career strategy?