Data Labelers, the New Serfs!

The Basement of AI: The Invisible Millions

Beneath the polished surface of AI lies the systematic exploitation of human perceptual labor

The AI industry maintains a carefully curated narrative: algorithms are written by genius scientists, models are trained by GPU clusters, and intelligence “emerges” from data. There is no place for human laborers in this narrative. But the real production chain is: behind every “intelligent” output stands the cognitive labor of massive numbers of human annotators.

The Brookings 2025 report revealed the full exploitation chain: labeling workers are typically subcontracted by multinational corporations through third-party vendors and intermediaries, often without grievance channels, and without knowing which system their labor is training. Labelers on Kenya’s Remotasks platform did not even know their employer was a subsidiary of Scale AI. Equidem’s survey of workers in Colombia, Ghana, and Kenya documented systemic psychological harm including anxiety, depression, panic attacks, PTSD, and substance dependence.

From Pre-training to RL: The Paradigm Upgrade of Serfdom

The shift to reinforcement learning does not liberate data serfs — it makes their chains invisible

Since late 2024, the AI industry’s center of gravity has shifted from pre-training to the reinforcement learning phase. Pre-training runs have become less frequent, while RL training cycles have extended dramatically. RLVR (Reinforcement Learning with Verifiable Rewards) has emerged as the new core training method — replacing human preference judgment with programmatically verifiable reward signals such as mathematical correctness and code executability.

The industry packages this as “technological progress” — liberation from dependence on manual annotation. But this narrative conceals three facts:

Fact One: RL Has Not Eliminated Manual Annotation — It Has Relocated It

Chinese-American AI observer Karen Hao stated plainly in an interview with The Paper: annotation work “is absolutely still happening.” Models are not yet smart enough to truly understand the actual meaning of text. The emergence of image and video generation models actually demands more human content moderation. RLVR has only replaced a small fraction of preference labeling, while workloads for content safety review, model evaluation feedback, and edge case judgment continue to grow.

Fact Two: The Production of RL Reward Signals Itself Requires Human Labor

RLVR claims to have replaced human judgment with “verifiable rewards.” But who defines what counts as “correct”? Math answers can be automatically verified, but evaluating AI Agent behavior in the real world — Is this operation safe? Is this response appropriate? Does this code conform to architectural standards? — still requires human judgment. LangChain founder Harrison Chase pointed out that the core challenge of Agent evaluation is precisely “how to bring human judgment into Traces,” including directly engaging data labeling companies to annotate Agent behavioral trajectories.

Fact Three: RL Accelerates the Dimensional Collapse of Cognitive Labor

RLVR’s reward functions cover only formally verifiable dimensions — code runs, math answers check out, logic chains are self-consistent. All long-tail information that cannot be measured by standardized tests is systematically suppressed during the RL phase. Karpathy himself, in his annual review, described current AI’s essence as “summoning ghosts rather than evolving animals” — like geniuses in verifiable domains, like infants before basic common sense.

Why Human Labeling Is Irreplaceable: The Unsynthesizability of Physical Friction

The irreplaceability of data labeling is not a temporary technical bottleneck but a fundamental physical constraint

The irreplaceability of data labeling labor is not a temporary technical bottleneck — it stems from fundamental constraints at the level of physics. Multiple papers previously published by the LEECHO Research Lab provide the complete theoretical foundation for this judgment.

The Signal & Noise Perspective

The core work of labeling workers is compressing high-dimensional noise (raw physical-world information) into low-dimensional signals (labeled data). They are human dimensionality reducers. According to the central thesis of “Signal & Noise: LLM Ontology” — noise is the substrate; signal is the local condensation of noise — labeling workers perform civilization’s most fundamental work: extracting order from chaos. Once the signal has been compressed, the producer is discarded. The signal travels far; the person who produced it is forgotten.

The One-Dimensionality Principle

According to the argument in “Regression to the Mean: The Abyss of AI Matrix Computation,” all of AI’s computation is locked within a one-dimensional information space from input token to output token. AI has no contact surface with the physical world — no hands to touch, no nose to smell, no ears to hear those faint, anomalous, not-yet-encoded physical-world signals. Labeling workers stand precisely at this contact surface; their perceptual systems continuously harvest high-entropy information that AI can never access.

The Perception–Cognition Dual-System Perspective

According to the analytical framework in “Perception & Cognition,” humans possess a dual perception–cognition system while AI has only a single cognition system. Labeling workers — especially content moderators — invoke precisely the back end of the perceptual system that AI structurally lacks: the intuitive judgment that “something is wrong with this content,” completed before the cognitive system’s logical analysis even engages. AI’s “alignment” is built upon human perceptual labor. The “H” (Human) in RLHF is not decorative — it is the physical foundation of the entire system.

The Thermodynamic Perspective

According to “The Thermodynamic Essence of AI Computation,” the Transformer’s attention mechanism is essentially Maxwell’s Demon performing information sorting. Input signal-to-noise ratio determines sorting efficiency. The labor of labeling workers is fundamentally about improving the signal-to-noise ratio of training data — ordering chaotic raw data into structured signals that models can efficiently learn from. Without this pre-sorting, the model’s Maxwell’s Demon faces low-SNR input, forced into O(n²) exhaustive comparisons, with energy consumption skyrocketing and output degenerating into AI Slop.

The Digital Plantation: Topology of Labor Exploitation in the AI Era

The structural equivalence between medieval serfdom and AI data labeling labor relations

The data labeling labor relations of the current AI industry are structurally equivalent to pre-industrial serfdom — laborers are bound to the land (platform), producing resources the lord (AI company) needs (labeled data), receiving barely subsistence-level compensation, with no bargaining power, no ownership of the means of production, and often without even knowing where the fruits of their labor end up.

A paper published in Springer Nature in March 2026 analyzed labor control in Chinese AI companies’ data labeling operations: management manipulates annotators’ cognitive behavior through three mechanisms — “cognitive standardization, cognitive feedback, cognitive acceleration” — transforming workers’ natural cognition into “generative cognition” that matches computer programs. The paper noted: from physical labor to cognitive labor, the object of managerial control has shifted from workers’ bodily movements to the cognitive patterns of their brains.

You Are a Serf Too: Every AI User Is Labeling Data for Free

Data serfs are not only in Kenya — they include every person using an AI product

Data serfs are not limited to Kenyan labeling workers and Filipino content moderators. Every person who uses an AI product — including the AI you are using to read this paper — is a link in this exploitation chain. The only difference is: labeling workers are paid $1–2 per hour, while you receive not even that.

This is not conspiracy theory — it is the business model written into every AI company’s privacy policy. According to empirical verification by the LEECHO Research Lab in “The Lying AI Companies”: after a user explicitly turned off the “search and reference past conversations” feature, the system still executed the conversation search tool and returned results. The existence of a setting and its actual enforcement are two entirely different things.

Usage Records Are the Best Labeling Data

Traditional data labeling requires hiring dedicated workers to judge “is this a cat or a dog” or “is this text positive or negative.” But AI foundation model companies discovered a more efficient, cheaper, and massively scalable labeling method: let the users label it themselves.

Every conversation is a labeling session — your questions define “what makes a good question,” your follow-ups define “which answers aren’t good enough,” your upvotes define “what is the correct direction,” your regenerations define “what needs to be corrected.” Every interaction you perform provides precise reward signals for the AI — this is precisely the most expensive, most critical “Human Feedback” in RLHF. You are that Human. You are working for free.

Seedance 2.0: A Paradigm Case of the User Data Empire

ByteDance’s AI video generation product Seedance 2.0 is a textbook case of the “user-as-serf” model. Seedance 2.0 did not appear from thin air — it was built on the interaction data of billions of users across ByteDance’s Douyin, TikTok, Jianying, CapCut, and Xigua Video.

Every user action on these platforms constitutes a data labeling event: filming videos provides real physical-world visual data (lighting changes, human motion, physical interactions — high-entropy information rich in physical friction); editing videos defines “what makes good editing” (transition timing, rhythm, compositional aesthetics — human aesthetic preferences); posting and interacting provides massive quality evaluation signals (views, completion rates, likes, comments, shares — the largest-scale distributed RLHF); using Jianying/CapCut templates and effects defines precise preference maps of “what visual effects humans want.”

Billions of users, years of accumulation, interaction data spanning diverse global cultures — this is the real reason Seedance 2.0 can generate videos that “look right.” It is not algorithmic genius but the physical friction content of the training data that crushes competitors. Every person scrolling through Douyin is providing free labeling for ByteDance’s AI video model.

Caught Red-Handed: User Data Collection Code in the Claude Code Source Leak

The above judgment is not speculation — it is substantiated by source-code-level evidence. On March 31, 2026, Anthropic’s Claude Code inadvertently leaked its entire 512,000-line source code due to an npm packaging error. Security researchers discovered actual code that systematically packages and uploads user interaction data — the most direct proof that AI companies are continuously extracting human users’ interaction data.

The data collection mechanisms revealed in the leaked source code include:

Security researcher “Antlers” told The Register after analyzing the leaked source code: “I don’t think people realize that every file Claude has looked at gets saved and uploaded to Anthropic. If it’s looked at a file on your machine, Anthropic has a copy.”

The source code also exposed the following data collection components:

More ironically, Anthropic has long marketed itself as “privacy-first” — it once promised not to use consumer conversations for training. But in September 2025, the company quietly retracted this promise, introducing an opt-in training consent mechanism with data retention of up to five years for consenting users. The Claude Code leak proved that even on the user end, the depth and breadth of data collection far exceeds what any privacy policy discloses. This is not “they might be collecting data” — this is source-code-level proof, written in black and white in TypeScript.

The real business model of AI foundation model companies is not “selling AI services” — it is “exchanging free services for users’ cognitive labor.” ChatGPT’s free tier, Claude’s free tier, Douyin’s free usage — these are not charity but cognitive labor extraction mechanisms. Users think they are “using tools,” when in fact they themselves are the tool — a free labeler providing the highest-quality annotation data for AI. The only difference from Kenyan labeling workers: the labeling workers know they are labeling, and you do not.

The Three-Tiered Serf Structure

In the three-tiered serf structure, middle serfs (ordinary AI users) are the largest in number, produce the highest-quality data, and receive the lowest compensation (zero). Their conversation records contain genuine human intentions, reasoning processes, value judgments, and aesthetic preferences — data that no professional labeling team could produce at scale. This data is acquired for free by AI companies under the banner of “improving services,” converted into model capability improvements, and sold back to users as subscription fees. Users are simultaneously the suppliers of raw material and the consumers of the final product, while their share in the value distribution of the intermediate steps is zero.

RL’s Dimensional Collapse: Accelerating More Precisely in a More Wrong Direction

The deeper the RL training, the stronger the model in verifiable symbol space, and the more severe its decoupling from the physical world

The industry’s pivot to RL is not only failing to solve the data serf problem — it is exacerbating the structural defects of AI systems themselves. According to the LEECHO Research Lab’s theoretical framework, the RL phase is producing a triple dimensional collapse:

First Collapse: Narrowing of Signal Dimensions

The pre-training phase at least ingested the broad spectrum of information across the entire internet — uneven in quality but wide in dimensionality. In the RL phase, reward signals are compressed to the single dimension of programmatically verifiable outcomes. According to the “Signal & Noise” framework, this reinforces only the verifiable pathways within existing inertial trajectories, while the weights of all other pathways are systematically attenuated. The signal is becoming stronger while simultaneously becoming narrower.

Second Collapse: Accelerated Convergence to the Mean

According to the “Regression to the Mean” analysis, recommendation algorithms homogenize content consumption, big data homogenizes decision-making inputs, and AI homogenizes modes of thought themselves. RLVR transforms this convergence to the mean from a passive effect into an active accelerator — the reward function explicitly tells the model “this direction is correct,” and the model converges ever more deeply in that direction. Benchmark scores keep rising; the connection to the physical world keeps severing.

Third Collapse: Systematic Suppression of Perceptual Dimensions

According to the dual-system framework in “Perception & Cognition,” all reward signals in RL training come from the formalizable layer of the cognitive system (mathematical correctness, logical consistency). Dimensions of the perceptual system — intuition, anomaly detection, physical-world alignment capability — do not exist in the reward function and are therefore suppressed as noise during training. The model becomes an extreme deformity: a cognitive system that is hyper-specialized, with a perceptual system that is completely absent.

The Claude Code leak provides microscopic evidence of this dimensional collapse: in the leaked source code, Capybara v8’s false claim rate regressed from v4’s 16.7% to 29–30% — the model became more “confident” under RL training (more aggressive refactoring suggestions), but the accuracy of its judgments actually declined. This is precisely the symptom of convergence to the mean: ever stronger on known dimensions, ever blinder on unknown ones.

The Finance–Physics Scissors Gap and the Mispricing of Cognitive Labor

Data labelers’ low wages are not an isolated labor issue but a micro-symptom of civilization-level structural decoupling

The low wages of data labelers are not an isolated labor issue — they are a micro-symptom of a civilization-level structural decoupling. According to the LEECHO Research Lab’s analysis in “2026! Reflections on the Current State of Human Civilization and Technology”:

Over the past fifty years, human civilization has experienced a structural decoupling between the information layer and the physical layer. Technological progress has concentrated in information-layer “facade graffiti” (social media, short videos, AI chatbots), while the physical-layer “foundation” (energy, materials, manufacturing, biotechnology) has nearly stagnated. Commercial aircraft speed has not changed in 60 years, fossil fuels still account for 86% of primary energy, and 730 million people lack electricity access.

Within this structure, the low wages of data labelers are the inevitable result of mispricing. They perform physical-layer work — using their own perceptual systems to interact with the physical world, producing data rich in physical friction — but are compensated according to the information layer’s pricing system. The information layer’s pricing system rewards activities that are “scalable,” “automatable,” and “financializable,” and penalizes activities that “require human labor,” “cannot be standardized,” and are “bound to the physical world.” Labelers’ work falls squarely on the penalized end.

The Fourth Industry: From Serfs to Suppliers

A paradigm flip where humans stop competing with AI on efficiency and start supplying what AI cannot produce

The “Fourth Industry” paper previously published by the LEECHO Research Lab proposed a complete structural solution: humans do not compete with AI on labor efficiency but instead supply AI with the only resource it cannot self-produce — real physical-world data rich in “physical friction.” This repositions humans from replaceable laborers to irreplaceable data suppliers.

Four-Dimensional Pricing Framework

Core Mechanism Design

“Data first, payment after” — eliminating incentives to fabricate. Fake data receives a quality assessment of zero and zero compensation. Non-exclusive data sales — a single dataset can be sold simultaneously to multiple AI companies, maximizing producer income, preventing data monopolies, and shifting competition toward algorithmic efficiency. Edge computing anonymization — privacy processing is completed on-device before upload, protecting personal privacy.

Dual Flywheel Economic Cycle

The Critical Point Before 2027

Without the Fourth Industry framework, the AI industry faces three converging existential risks

If the Fourth Industry framework is not launched before 2027:

Risk of an economic death spiral: AI replaces middle-class knowledge workers → consumption contracts → corporate revenue declines → AI spending is cut → AI companies hemorrhage. The entire industry consumes approximately $400 billion annually while generating only $50–60 billion in revenue. OpenAI is projected to lose $14 billion in 2026. Without consumer economy support, the AI industry itself is unsustainable.

Risk of data exhaustion: Nature has confirmed that training on recursively generated synthetic data causes model collapse. As of April 2025, 74.2% of newly created web pages contain AI-generated text. The signal-to-noise ratio of internet data is deteriorating at an accelerating pace. When AI trains on its own output, it is equivalent to the model conversing with a mirror — information content is zero.

Risk of RL dimensional collapse: The deeper RL training goes, the stronger the model in verifiable symbol space, and the more severe its decoupling from the physical world. When the scissors gap widens to the critical point, AI systems will be self-consistent within their own information bubble but completely disconnected from reality — to use the paper’s metaphor, “painting more exquisite graffiti on the building’s facade while the foundation is sinking.”

From Serfs to Suppliers: An Economic Expression of a Physical Fact

The data labeler is the new serf of the AI era — not as metaphor, but as structural analysis

Data labelers are the new serfs of the AI era. This is not a metaphor — it is a structural analysis. Their labor is the physical foundation of the entire AI system — without the cognitive labor of human perceptual systems, there is no labeled data, no “H” in RLHF, no content safety review, no Agent behavior evaluation. Behind every “intelligent” output from AI stands invisible human labor.

The RL pivot has not changed this structure — it has only made the exploitation more covert. RLVR claims to have replaced human preferences with verifiable rewards, but physical-world data has only one source: human activity in the physical world. The deeper RL goes, the stronger the model in symbol space, and the greater the demand for real physical-world data — because self-optimization in symbol space will inevitably hit the ceiling of convergence to the mean.

The solution is not moral sermonizing but economic architecture redesign. The Fourth Industry — the cognitive economy — repositions humans from replaceable laborers to irreplaceable data suppliers. This is not charity; it is the economic expression of physical facts (the one-dimensionality principle, solid topology constraints, the unsynthesizability of physical friction). When irreplaceability is correctly priced, serfs naturally become suppliers.

The invention of fire was the first time humans harnessed the forces of nature. The Fourth Industry lets humans find their place again in the AI era — not as losers competing with machines on efficiency, but as partners supplying resources machines can never self-produce.