Google Talks $1.5 Billion Deal for AI Training Startup as Meta Tracks Employee Keystrokes

Google is in talks to invest more than $1.5 billion in Mechanize, a startup that builds virtual work environments designed to train AI agents, Business Insider reported. The deal would bring Mechanize's team into Google's orbit while the company licenses some of the startup's technology for model evaluation and development.

Meta, meanwhile, has been collecting employees' keystrokes, mouse movements, clicks, and other screen activity, according to Business Insider. The company says the goal is teaching AI how people actually use computers, from keyboard shortcuts to navigating internal business software.

These are two data points in the same story: the AI industry is moving past teaching chatbots to answer questions and toward teaching AI agents to do entire jobs.

From Predicting Words to Practicing Work

The first wave of large language models ran on two ingredients. Massive text scraped from books and the internet. Then human contractors rating whether chatbot answers were good or bad.

That recipe built impressive conversational AI. It has not been enough to build agents that can independently grind through complex, multi-step work over hours or days without falling apart.

The industry's answer is what's called reinforcement learning environments, or RL environments. These are simulated digital workplaces where an AI agent can attempt coding tasks, navigate business software, make decisions, and try to complete long-running jobs the way an employee would. Instead of grading a single answer, these environments let an AI try, fail, recover, and try again, then get judged on whether it finished the whole task.

Scale AI, a major supplier of AI training data, says nearly half of its new AI training projects now involve these RL environments, according to Chetan Rane, the company's head of product for agents and RL environments, in a company blog post cited by Business Insider. Rane wrote that frontier models increasingly need to learn through trial and error in realistic simulated environments rather than leaning solely on static datasets or human preference feedback.

What This Means for Workers

If the pitch works, AI agents will not just chat. They will operate spreadsheets, write and debug code across multi-step projects, and execute workflows currently handled by junior and mid-level employees in fields like software engineering, data analysis, and back-office administration.

That is the entire commercial bet behind Mechanize and companies like it: agents trained on simulated job tasks are worth more to enterprise customers than chatbots that just answer questions.

The Meta Keystroke Question

Meta's keystroke and mouse-movement collection raises a fair question that deserves a straight answer: whose data is this, and did employees know exactly what it would be used for?

Companies routinely monitor employee activity on corporate devices for security and productivity reasons, and that is legal and unremarkable on its own. Using that same data to train commercial AI products is a different purpose, and workers who signed up for a paycheck did not necessarily sign up to become the training set for the system that might automate their own job or someone else's. Business Insider's reporting does not indicate whether Meta obtained separate consent for AI-training use specifically, or whether this falls under existing employee monitoring policies. This gap matters, and it is worth watching whether Meta or any regulator addresses it directly.

At the same time, there is nothing inherently sinister about a company studying how its own employees use its own software to build better tools. Every company that makes workplace software has an interest in understanding how people actually click through it. The concern is not that Meta is doing research. It is transparency about scope, consent, and what happens to that behavioral data once it trains a model that outlives the employees who generated it.

The Bigger Shift

Google's talks with Mechanize and Meta's data collection are not isolated moves. They reflect a race among the largest AI companies to own the next training bottleneck. Text scraped from the internet is largely exhausted. Human-rated chatbot answers only go so far. The next edge comes from realistic simulations of actual work.

Business Insider's reporting frames this shift accurately as an industry-wide pivot, backed by Scale AI's own data showing roughly half its new projects now involve these environments. The price tag for workers whose daily habits, keystrokes, and workflows are becoming raw material for the machines meant to replace or augment them remains unresolved. Whether any company will offer clearer terms on that trade before the next round of layoffs gets blamed on "AI efficiency" is still an open question.

No terms of the Google-Mechanize deal have been finalized as of this reporting, and neither company has issued a public statement confirming the exact investment structure. Whether that deal closes, and at what valuation, is the next concrete milestone to watch.