Human Labor as Software Infrastructure
Scale: the $290M/year Mechanical Turk of machine learning
The important shift was turning one off human judgment into software infrastructure that could be called on demand. Mechanical Turk let a developer send a task to a distributed workforce through an API, while reCAPTCHA embedded tiny labeling jobs into normal web traffic, using millions of users to help read scanned text that OCR systems could not reliably decode. That pattern matters because it showed human labor could be routed, measured, and productized like compute.
-
Mechanical Turk launched in November 2005 as a web services API for what Amazon called artificial artificial intelligence. In plain terms, software could post a job like identify the object in this photo, wait for a person to answer, then pull that answer back into the application.
-
reCAPTCHA took the same basic idea but changed the interface and the economics. Instead of paying workers directly, websites collected human effort from login and signup flows. Carnegie Mellon materials describe reCAPTCHA as helping digitize old books, with each solved prompt contributing human corrections where OCR struggled.
-
This created the template that Scale later specialized for machine learning. General crowdwork platforms had labor supply, but Scale paired labor with task routing, quality checks, and labeling software built for image, video, and 3D sensor data. That is why it could outperform broader vendors like Appen and sell into autonomous driving workflows.
Going forward, the winning versions of this model keep moving up the stack. Raw crowd labor becomes less valuable on its own, while systems that combine humans, software, and model assisted review become more valuable. The market keeps rewarding companies that turn messy human judgment into a reliable production workflow for high stakes AI use cases.