Meta will record staff keystrokes, mouse movements and clicks on some internal apps to train AI agents, the company told TechCrunch. It said the data will help models learn how people actually complete tasks inside software, rather than relying on synthetic examples.
What Meta says it will collect
Meta is rolling out an internal tool that will capture staff interactions with certain company applications, the company confirmed. The system will record mouse movements, clicks and keystrokes so engineers can use real examples of how people navigate software when training AI agents.
The company told TechCrunch that the aim is to teach models how people actually complete tasks on computers — selecting menu items, filling forms, clicking buttons — rather than relying only on synthetic or labelled examples. Meta said safeguards are in place to protect sensitive material and that the captured inputs are intended solely for model training.
Reuters first flagged the story earlier this week, and TechCrunch published the company’s response after seeking comment.
Why internal logs matter to AI builders
Training modern AI models takes vast amounts of data. Raw text and images have long been the main fuel, but companies building agents that act inside apps say they need different kinds of examples — sequences of human actions tied to on‑screen outcomes.
- Mouse paths and typed input reveal human workflows: where people hesitate, which dropdowns they use, and how they correct mistakes.
- Those signals help an AI learn not just what to say, but how to act inside software interfaces and perform multi-step tasks.
- Meta says internal telemetry is a pragmatic, lower-cost way to simulate realistic interactions compared with hiring testers or crafting synthetic behaviour.
Privacy and legal concerns
The plan drew quick attention because it raises hard questions about employee privacy and corporate data governance. Recording keystrokes has a fraught history in workplaces when used for monitoring rather than legitimate product research.
Privacy advocates warn that even if a company intends narrow use, keystroke capture can expose highly sensitive content — passwords, private messages, and personal notes often flow through the apps people use at work. Even with filters, the risk of accidental capture remains.
Meta said there are safeguards to protect sensitive content and that the data won't be used for other purposes beyond training its models. The public disclosure prompted fresh scrutiny of how broadly that promise can be enforced and audited.
Industry context: feeding the AI pipeline
Meta's move comes as tech firms search for new sources of training material. The industry has already turned to a wide variety of datasets — public text, licensed content and, controversially, private corporate archives in some cases.
Recent reporting has documented buyers of older startups' internal records, such as archived chats and ticketing systems, being repurposed as training data. Collecting internal usage traces is a logical next step for companies that want to teach agents to act inside software, but it shifts the debate from public‑facing data to employees’ day‑to‑day work, where legal protections and expectations differ.
Why this matters: the move brings the privacy debate squarely into the workplace rather than the public internet, and it raises a concrete question the company must answer: how will Meta demonstrate that filters and audits reliably prevent sensitive material being captured or used for other purposes?
Related Articles
- BMJ Open audit finds nearly half of AI health answers problematic
- Amazon to invest up to $25bn in Anthropic
- Anthropic secures $5bn from Amazon and vows over $100bn AWS spend
Meta said the tool will capture mouse movements and keystrokes on certain apps to train agents and that safeguards will prevent the data's reuse for other purposes. The immediate question is how Meta will demonstrate those safeguards can be independently audited and enforced.
This article was created with AI assistance.
