AI Companies Now Train Models on Your Workplace Keystrokes and Chat Logs, Not Just Your Prompts

The AI industry spent the last few years training chatbots to answer questions well. Now it's training AI agents to do entire jobs, and that requires a lot more data than a chat window ever collected.

According to Business Insider, Google is in talks to invest more than $1.5 billion in Mechanize, a startup building virtual work environments where AI agents practice coding, using business software, and completing multi-step tasks. The report says the deal would bring Mechanize's team to Google while the company licenses some of the startup's technology.

Elon Musk told SpaceX employees the company will collect and use their workplace activity data to improve Grok, the AI model built by his company xAI, according to Business Insider. Musk's own words, per the outlet: "It will be trained on you."

Meta has reportedly been trying to collect employees' keystrokes, mouse movements, clicks, and screen activity, according to Business Insider, with the stated goal of teaching AI how people actually navigate workplace software.

Scale AI, a major supplier of AI training data, says this is now industry-wide. Chetan Rane, the company's head of product for agents and reinforcement learning environments, wrote in a blog post cited by Business Insider that "frontier models increasingly need to learn through trial and error in realistic simulated environments rather than relying solely on static datasets or human preference feedback." Nearly half of Scale AI's new training projects now involve these simulated work environments, Rane said.

The Chatbot Data Problem Hasn't Gone Away

While companies build new ways to harvest workplace behavior, the older problem of employees pasting sensitive information into public chatbots is still very much alive.

Wald.ai reported that the Cybersecurity and Infrastructure Security Agency had to launch an internal review after acting director Madhu Gottumukkala uploaded sensitive "For Official Use Only" documents to the public version of ChatGPT. If accurate, that's a federal cybersecurity official making the exact mistake his agency exists to prevent others from making.

Wald.ai also cited industry figures claiming more than 77% of organizations are actively exploring or using AI tools, and that 58% of those companies have already dealt with AI-related security breaches. Those numbers come from industry surveys, not government audits, and should be read as such. The direction is consistent with other reporting: companies are adopting AI faster than they're building policies to control it.

Samsung banned employee use of ChatGPT after confidential source code reportedly leaked through the tool, according to Wald.ai. OpenAI has also faced a court order requiring it to preserve user data indefinitely as part of ongoing litigation, a development that cuts against OpenAI's own stated data-retention policies and gives users less control over what happens to old conversations.

What's Actually New Here

None of this means AI companies are doing something illegal. Employees generally sign agreements that let employers monitor company equipment and workplace software use. Musk's SpaceX and Meta are within their legal rights to collect workplace activity data from company-owned systems and accounts, and neither company has been accused of violating any law in these reports.

But there's a real difference between a company monitoring work output and a company building AI training pipelines out of every keystroke, mouse movement, and screen interaction an employee makes. Workers who signed on for a paycheck didn't necessarily sign on to become the raw material for a commercial AI product, and it's unclear from current reporting how much say, if any, employees at Meta or SpaceX have in opting out.

The strongest case for this approach is straightforward: AI agents that can actually complete real work, the kind companies are racing to build, need to learn from real work, not just internet text. Static datasets and contractor feedback got chatbots good at answering questions. They haven't produced agents that can reliably run a multi-hour task inside real business software. Scale AI's data suggests the industry believes simulated environments are the fix.

No regulator has announced an investigation into Meta's or SpaceX's employee-monitoring-for-AI-training practices as of now. Whether workers get any meaningful choice in the matter, and whether "trained on you" comes with any opt-out at all, remains something neither company has detailed publicly.