Human Expertise, Training Data, and the Economics of AI Development

The popular picture of AI development is a story about computers: more chips, bigger models. Underneath it is a story about people. Every capable model has been shaped by the knowledge of people who wrote the text it learned from, demonstrated what good answers look like, ranked outputs, wrote the tests and graded the results. This article explains where that human contribution enters, what the evidence says about its value, and why consent and fair pay are not afterthoughts but part of the economics.

Five places where people enter

1. Pretraining text. Models learn from text that people wrote.
2. Demonstrations. People write examples of desired behavior. InstructGPT collected labeler demonstrations for supervised fine-tuning [1].
3. Preferences. People rank outputs, and the rankings train a reward model; this is the heart of reinforcement learning from human feedback.
4. Specifications. People write the tasks, rubrics and tests that define success, particularly in reinforcement learning with checkable answers [2].
5. Evaluation. People write and grade the exams. GPQA's 448 questions were written by domain experts; experts with PhDs reached 65% accuracy while skilled non-experts with web access reached 34% [3].

Quality can beat quantity

A striking result is how little curated data can do. The LIMA study fine-tuned a 65-billion-parameter model on just 1,000 carefully curated prompts and responses, without reinforcement learning, and reported remarkably strong performance. The authors took this as evidence that most knowledge is learned in pretraining and that limited, high-quality tuning data can teach the model how to respond [4]. And in the InstructGPT study, a 1.3-billion-parameter model fine-tuned with human feedback was preferred by human evaluators to a model with 175 billion parameters [1].

These results are specific to their setups and should not be stretched, but the direction is clear: who produced the data and how well matter at least as much as how much.

Will human data run out, or be replaced?

Two findings bear on this.

Supply. Villalobos and colleagues forecast that if current trends continue, models will be trained on datasets roughly equal to the available stock of public human-generated text sometime between 2026 and 2032, and discussed synthetic data, transfer learning and efficiency as routes beyond it [5].

Synthetic data's limit. Shumailov and colleagues found that indiscriminate use of model-generated content in training causes irreversible defects, which they call model collapse, in which the tails of the original data distribution disappear [6].

Our interpretation: together, these suggest that fresh, high-quality human contributions, particularly about rare situations, expert judgement, local knowledge and languages with little web presence, remain valuable. They do not prove that any particular dataset will sell or how large the market is. We have not found independent, verified figures for the market and do not offer any. The honest test of value is measurement: did training on this data improve performance on held-out tasks that matter?

The part that is easy to ignore: the people

Data work is work. Hara and colleagues analysed earnings on Amazon Mechanical Turk and reported a median hourly wage of about two US dollars, with only a small share of workers earning the US federal minimum wage [7]. Irani and Silberman described Turkopticon, a tool through which workers on the platform reviewed employers [8]. If expertise is the scarce input, treating experts that way is poor economics as well as poor ethics: skilled people do not spend scarce time on opaque, low-paid tasks, and quality falls.

Fair arrangements include clear pay, explained acceptance and rejection, a way to dispute decisions, and the freedom to choose what to take part in.

Rights are part of the cost

Data are only usable if the rights are real. A Data Provenance Initiative audit of more than 1,800 text datasets found inconsistent documentation of licences and attribution [9]. A study of 14,000 web domains found that restrictions on AI use were spreading and often inconsistent between a site's terms of service and its robots.txt [10]. Consent has a legal shape in many places; under the EU's data protection regulation, for example, it must be freely given, specific, informed and unambiguous [11].

The practical consequence is that a data business is also a rights business: for each source, who owns it, what was agreed, which purposes are allowed, and what must be deleted on request.

Where social intelligence fits

The hardest and most valuable data to produce may be the kind that teaches models about people: how a mediator defuses a conflict, how a teacher explains something gently, how a community in one region says no politely. That knowledge is tacit, culturally situated and often subjective. It needs expert judgement, local contributors and rubrics that accept disagreement. It is also where private information is most at stake, which is why HUTE keeps private HUMAN conversations out of its data business entirely. See Intelligence Infrastructure and Responsible AI.

For the full review, with sources, read Research Report 003. To see how real work becomes training material, read what are AI production traces.