When we marvel at a chatbot writing working code, it’s easy to picture a machine that simply absorbed the internet. The reality is more human. Behind many leading models stands a quiet industry of specialists who write, grade, correct, and curate examples. Understanding that supply chain explains a lot about why AI gets better, and where it still stumbles.
Stage one: pretraining on huge piles of text
The first phase teaches a model language and general knowledge. It reads enormous collections of text and code, learning to predict what comes next. This stage is mostly automated, but even here people make choices: which sources to include, how to filter low-quality or harmful material, and how to remove duplicates.
Quality filtering matters more than raw volume. A smaller set of clean, well-structured material often beats a larger noisy one.
Stage two: teaching the model to be useful
A model fresh from pretraining can continue text, but it isn’t naturally a good assistant. The next phase uses curated examples of good behavior. People write ideal answers to prompts, or edit model outputs to improve them. This is called supervised fine-tuning.
For coding, this is where skilled engineers come in. They write solutions with clear explanations, fix buggy outputs, and craft tricky problems that test whether the model really understands. Their expertise determines how good the examples are.
Stage three: preference feedback
The model produces several answers to a question, and human reviewers rank them. Those rankings train the model to prefer responses people consider better: more accurate, clearer, safer, more helpful. This technique, often called reinforcement learning from human feedback, shapes the model’s tone and judgment as much as its knowledge.
For code, automated checks can supplement or replace human ranking. Tests either pass or fail, giving a clean signal. Increasingly, labs use these automatic verifiers to train models on thousands of programming challenges.
Why data vendors exist
Building a pipeline of qualified contributors is hard. You need to recruit people with the right skills, such as software engineers, mathematicians, doctors, or lawyers; train them on guidelines; check their work; pay them correctly; and manage quality across thousands of tasks. Specialist companies handle this for AI labs. They match experts with projects and run quality control.
The demand has shifted over time. Early projects relied on large numbers of general annotators labeling simple things. As models improved, the need moved to expert-level contributors who can judge complex reasoning, which is more expensive and harder to find.
What quality looks like
- Correctness. Code must actually run and handle edge cases.
- Diversity. Examples should cover many languages, frameworks, and problem types.
- Clarity. Explanations should be understandable, not just technically valid.
- Consistency. Reviewers must apply guidelines the same way, or the model receives mixed signals.
- Freshness. Libraries and best practices change; stale examples teach outdated habits.
Challenges and criticisms
Working conditions
Some data work is low paid and repetitive, and journalists and researchers have raised concerns about pay, job security, and exposure to disturbing content during safety review. Better industry standards are widely called for.
Bias
The people who label data bring their own perspectives, and guidelines embed value judgments. Models can reflect the demographics and assumptions of their trainers.
Synthetic data
Labs increasingly use model-generated data to supplement human examples. It’s cheap and scalable but can amplify mistakes if not carefully verified. Training on a model’s own output repeatedly risks degrading quality over time.
Copyright and consent
The sourcing of training material remains contested, with legal cases and policy debates around what content may be used and how creators should be compensated.
What this means for users
The next time an AI model solves a tricky coding problem, remember there is a long chain of human judgment behind it. It also explains uneven performance: strong in areas with plenty of verified examples, weaker where data is scarce or hard to check. If you work with AI tools, verifying output remains essential, especially in niche domains.
Takeaway
AI isn’t built from raw data alone. It depends on careful curation, expert feedback, and automated verification. As models grow more capable, the value of high-quality human input rises, not falls. The companies and workers who supply it are an essential, though often invisible, part of the technology we use every day.
Featured image: Programmer at work (Unsplash).jpg by Crew crew, CC0, via Wikimedia Commons.
