OpenAI hires an army of contractors to teach AI how to code

Silicon Valley loves to talk about algorithms training themselves, but behind OpenAI’s slick tools is a massive army of real humans doing homework.

OpenAI has quietly hired around 1,000 remote contractors over the past six months across Latin America and Eastern Europe, according to a report from Semafor. About 60% are doing standard data labeling for computer vision and audio, but the other 40% are actual computer programmers tasked with teaching AI how to write software from scratch.

Explain your work, line by line

In the past, OpenAI trained its coding models like Codex by simply scraping billions of lines of public code from Microsoft’s GitHub. That taught the AI how code looks, but not how developers think.

This new hiring push is completely different. OpenAI is building a custom dataset that pairs working code with step-by-step human explanations written in plain English.

One South American developer who took a five-hour unpaid coding test said applicants had to tackle two-part problems. First, write down the entire thought process and approach in English. Second, provide the code.

Even bug fixing had strict rules. If applicants spotted an error, they couldn't just patch it. They had to document exactly what the bug was, why it happened, and how to correct it step by step. As the developer noted, OpenAI wants to capture the precise human reasoning behind every architectural decision.

The ghost fleet behind talent density

Just recently, OpenAI CEO Sam Altman bragged on Twitter that the company had only 375 full-time employees, praising its insane talent density compared to giants like Google or Meta.

That headcount conveniently excludes the sprawling workforce of outsourced labor across Latin America, Eastern Europe, and Kenya. Former Tesla AI chief Andrej Karpathy recently tweeted that English is the hottest new programming language, and these contractors are essentially drafting the dictionary.

Why it matters

This is a clever and necessary pivot for OpenAI. Current coding assistants are glorified autocompletes that often hallucinate because they mimic syntax without understanding problem-solving structure. By feeding models explicit human chains of thought, OpenAI is training Copilot to graduate from a code suggester to a full-blown autonomous builder that can diagnose its own mistakes.

For creators and founders with zero coding background, this brings the dream of building apps simply by describing them much closer to reality. But for junior developers, it should serve as a wake-up call that boilerplate coding is being commoditized fast.

Still, we think skepticism is warranted. We have seen this movie before with self-driving cars, which looked solved a decade ago before hitting a wall of corner cases. Teaching an AI to regurgitate basic Python is one thing, but whether underpaid contractors writing basic exercises can truly teach a model complex enterprise architecture remains an open question.