AI depends on things people have written, noticed, and judged. As demand for that material grows, the question reaches beyond model training: who produces it, and what do they receive for their contribution?
We Have But One Internet
“Data is the fossil fuel of AI,” Ilya Sutskever told the audience at NeurIPS 2024. “We have achieved peak data.”
Epoch AI has quantified the problem: the stock of high-quality, human-generated public text sits at roughly 300 trillion tokens. At current consumption rates — datasets growing 2.5x annually, compute scaling 4x — that stock will be exhausted between 2026 and 2032. Possibly sooner.
Three hundred trillion sounds like a lot until you realize it is everything humans ever typed into the internet. Every blog post, every research paper, every forum argument. One finite deposit.
And it is shrinking from both sides. Between 2023 and 2024, owners of 25% of the web’s highest-quality data sources blocked AI scrapers entirely. Harvard legal scholars now compare pre-2022 uncontaminated datasets to “low-background steel” — metal forged before nuclear testing irradiated the atmosphere, irreplaceable because you cannot un-contaminate what has already been mixed with synthetic material.
The internet is not a renewable resource. We just assumed it was.
The Synthetic Mirage
If human data runs out, why not generate synthetic data? Train models on their own output. Let AI feed itself.
Nature published the results of that experiment. When AI models train on recursively generated synthetic data, they suffer what researchers call “model collapse.” The rare patterns, minority perspectives, and unusual observations that researchers call the tails of the distribution vanish. Outputs converge to a bland average, then to nonsense. The degradation is irreversible. Training on AI-generated data is like photocopying a photocopy. You lose information with every pass.
This is not a universal failure. Synthetic data works brilliantly where verification oracles exist. A mathematical proof is right or wrong. Code compiles or it doesn’t. AlphaProof solved International Math Olympiad problems through pure self-play. AI systems design novel crystal structures by searching combinations of elements against known physical laws. But these are searches within defined parameter spaces, verified against established rules — fundamentally different from navigating cultural context, social reality, or situations where no oracle can confirm the answer.
The world is mostly not a math proof.
Market dynamics, cultural shifts, how customers behave in a store in Nairobi, what it feels like to use a product that technically works but subtly fails — this is knowledge AI cannot verify against a formal oracle and therefore cannot generate for itself. No synthetic dataset captures why a pricing model works in Stockholm but not in Lagos.
The domains where AI is genuinely superhuman are the ones with scorecards. The domains where humans remain irreplaceable are the ones where nobody agrees on how to keep score — or where the score takes years to arrive.
Content Floods, Attention Doesn’t
The content side of this equation is already visible.
Instagram post reach fell 31% from 2024 to 2025. Reels reach dropped 35%. Not because audiences left, but because AI-augmented production tools flooded the supply side while human attention stayed exactly where it has always been: finite.
YouTube responded in July 2025 with a policy overhaul that would have been unthinkable two years earlier. Fully automated AI videos no longer qualify for monetization. The platform now requires “genuine creative effort” and “clear human involvement.” Text-to-speech over stock footage, formulaic reaction content generated at industrial scale — demonetized.
But platforms are thriving, a skeptic might note. YouTube posted $36.1 billion in ad revenue in 2024, up 13%. If content homogeneity were a crisis, wouldn’t people leave?
They would not, because platforms monetize attention, not content quality. A platform can post record revenue while its creators starve. The relevant signal is not platform health — it is what happens to the value of each piece of content when anyone can produce unlimited material at near-zero cost.
Production capacity stops being the scarce resource. Novelty becomes the scarce resource. And novelty requires a human who has been somewhere, observed something, thought something that was not already in the training data.
That scarcity is not new. It has been embedded in AI systems from the very beginning — we just never called it what it was.
Artificial Artificial Intelligence
In 2005, Jeff Bezos launched Amazon Mechanical Turk and described it as a marketplace for “artificial artificial intelligence.” The name was borrowed from an 18th-century hoax: a chess-playing automaton that toured Europe defeating human opponents, concealing a human chess master crouched inside the cabinet. A machine that was not a machine.
Modern data-labeling services still depend on human work behind the software.
reCAPTCHA asked users to prove they were human by deciphering distorted text. What they were actually doing was digitizing books for Google — proving their humanity by performing labor for a machine.
Today, Scale AI operates a workforce of over 240,000 human data labelers spread across Kenya, the Philippines, and Venezuela. They draw bounding boxes, classify images, evaluate AI outputs, and provide the reinforcement signals that keep large language models from drifting into hallucination. The pay ranges from $0.90 to $2.00 per hour in the Global South, up to $25 in the United States — for identical work.
Reddit, meanwhile, made a different calculation. Twenty years of messy, opinionated, irreplaceably authentic human conversation turned out to be worth $203 million in data licensing deals with Google and OpenAI. By late 2025, Reddit was pushing for dynamic pricing — compensation that increases as its content proves more essential to AI performance. The forum that everyone mocked as the internet’s id became the internet’s most valuable mine.
The pattern runs in one direction. In the old model, humans needed machines to be productive. In the emerging model, machines need humans to remain capable.
Two Futures for the Human Sensor
Greater demand for human data does not necessarily mean better conditions for the people producing it.
Calling people “sensors” and “APIs” carries a real risk: it can make exploitation sound like an ordinary design choice. The Human API argument has to account for the people doing this work, rather than romanticize their role in the technology. The current evidence supports that objection: Kenyan data laborers describe their working conditions as “modern day slavery.” Scale AI’s Remotasks shut down its Kenya operations in March 2024, stranding thousands of workers.
We named this industry after a con trick — a machine that hid a human inside — and then built the con trick at planetary scale.
But the market is also bifurcating. While Remotasks collapsed its low-wage operations, Surge AI built a curated platform of 50,000 expert annotators — linguists, lawyers, medical doctors, writers — earning premium rates and commanding a valuation trajectory toward $25 billion. According to Amplify Partners’ analysis, data labeling costs surged 88x from 2023 to 2024, while compute costs rose 1.3x. That ratio reflects a shift from commodity labeling to expert annotation, not a uniform cost increase across all data work.
The binding constraint on AI progress is shifting from silicon to people. As models grow more capable, evaluating their outputs requires increasingly specialized humans — senior practitioners who take weeks, not days, to assess whether a legal analysis is sound or a medical recommendation is safe. This is the same expertise cliff that is breaking apprenticeship pipelines across industries, now showing up as an economic signal in AI’s own supply chain. You cannot gig-work your way to that judgment.
Two futures, then. In one, the Human API is Mechanical Turk 2.0: billions of people reduced to callable data endpoints, performing micro-tasks for micro-pay with no protections and no leverage. In the other, the scarcity of genuinely irreplaceable human-generated data gives workers something they have not had in the AI era — bargaining power.
Which future materializes is not a prediction to watch unfold. It is a choice organizations are making right now, whether they realize it or not.
Recognizing the people behind the data
If your organization uses AI, it depends on human-generated data that AI cannot produce for itself. The question is whether you know where that dependency lives.
Start by tracing that dependence through the work your people already do.
Map where humans in your organization generate irreplaceable observations — field reports, expert judgments, customer interactions, the operational knowledge that lives in Slack threads and hallway conversations rather than databases. This is what teachers discovered first: the value is not in transmitting knowledge but in generating the novel observations that no training set contains. Most organizations produce far more valuable human data than they realize, and treat it as exhaust rather than asset.
Score how substitutable each data stream is. If AI can generate a 70% adequate alternative, that stream is a commodity. If it requires embodied experience, cultural fluency, or situated judgment that no synthetic process can replicate, it is an appreciating asset. Reassess this regularly: last year’s expert judgment may be this year’s commodity labeling, because the bar for what counts as “expert” keeps rising.
Choose your play. Structure irreplaceable data for licensing — though recognize that Reddit had two decades of authentic public discourse at massive scale; your version may be smaller but no less valuable if the data is genuinely unique. Deploy your people as deliberate data collectors in domains where AI cannot self-verify. Build infrastructure that converts human experience into structured data products on an ongoing basis.
But before you build any of it, run the exploitation check. Are the humans generating your data compensated at a rate that reflects the value they create? Do they know their output feeds AI systems? Would you be comfortable if a journalist described your human data supply chain in detail?
If you fail two of those tests, you are building Mechanical Turk 2.0 regardless of your intentions.
A field report or an expert review comes from someone spending time and exercising judgment. These systems depend on people continuing to contribute that work. Their pay and working conditions belong in the design of the system, not outside it.
References
-
Villalobos, P., Ho, A., Sevilla, J., Besiroglu, T., Heim, L., & Hobbhahn, M. (2024). “Will we run out of data? Limits of LLM scaling based on human-generated data.” Epoch AI. https://epoch.ai/blog/will-we-run-out-of-data-limits-of-llm-scaling-based-on-human-generated-data
-
Shumailov, I., Shumaylov, Z., Zhao, Y., Papernot, N., Anderson, R., & Gal, Y. (2024). “AI models collapse when trained on recursively generated data.” Nature, 631. https://www.nature.com/articles/s41586-024-07566-y
-
Catanzaro, S. (2025). “Annotation for AI Doesn’t Scale.” Amplify Partners. https://www.amplifypartners.com/blog-posts/annotation-for-ai-doesnt-scale
-
Pogrebna, G. (2024). “AI is a multi-billion dollar industry. It’s underpinned by an invisible and exploited workforce.” The Conversation. https://theconversation.com/ai-is-a-multi-billion-dollar-industry-its-underpinned-by-an-invisible-and-exploited-workforce-240568
-
Burden, J., Chiodo, M., Grosse Ruse-Khan, H., et al. (2025). “Model Collapse and the Right to Uncontaminated Human-Generated Data.” Harvard Journal of Law & Technology. https://jolt.law.harvard.edu/digest/model-collapse-and-the-right-to-uncontaminated-human-generated-data
-
Schwartz, O. (2019, updated 2021). “The Untold History of Amazon’s Mechanical Turk.” IEEE Spectrum. https://spectrum.ieee.org/untold-history-of-ai-mechanical-turk-revisited-tktkt
-
TechCrunch (2025). “YouTube prepares crackdown on mass-produced and repetitive videos as concern over AI slop grows.”; Search Engine Journal (2025). “YouTube’s AI Slop Problem and How Marketers Can Compete.”
-
Reddit data licensing revenue and dynamic pricing: TechCrunch (2024), WinBuzzer (2025), Columbia Journalism Review (2025).
-
Surge AI valuation and expert annotation model: Sacra (2025), Amplify Partners (2025).
-
Instagram reach decline (2024-2025): SocialInsider (2025).