The AI industry's race to secure data is moving from training chatbot answers to training systems to learn how humans actually work, IT outlet Business Insider reported on Aug. 12 local time.

So far, large language models have developed based on data from internet text, books and human evaluators judging the quality of chatbot answers. The approach rapidly advanced conversational AI, but showed limits in building AI agents that independently carry out complex tasks over hours or days.

Many tech companies are focusing on digital environments similar to real workplaces, or reinforcement learning environments. There, AI agents try coding, using work software, making decisions and carrying out long-term tasks. They learn through failure and recovery. Instead of judging whether an agent gave a "good answer", the method gives reward signals based on whether it completed an entire workflow.

Recent industry moves show the shift clearly. Google is reported to be discussing a plan to invest more than $1.5 billion in Mechanize, a startup building virtual work environments for training AI agents. Mechanize said it aims for "full automation of the economy" and will create simulation and evaluation environments that capture what people do in real workplaces. Its first target is software engineering, with plans to expand to various white-collar tasks.

Elon Musk (일론 머스크), Tesla's CEO, also said he will collect work activity data from SpaceX employees and use it to improve the Grok AI model. Musk compared employees to AI's parents and said the model would inherit employees' thoughts and values. Meta is also seeking to collect employees' key presses, mouse movements, clicks and screen context so AI can learn how people actually use computers and work software.

Scale AI is also joining the shift. The company explained that "nearly half of the latest AI training projects are related to reinforcement learning environments that model coding, computer use and enterprise workflows." Uber is pushing an Agentic Pods project by putting AI engineers into its finance, legal, HR, marketing, procurement and customer support departments to observe how employees actually work and redesign that into AI-centered workflows.

The next stage of AI competition does not end with building smarter chatbots. Companies are digitally replicating real workplaces and having AI practise the many judgments and actions that make up modern work inside them. It is the start of a new data race to train AI agents to work like human employees.