Gemini inside Google ran into fear over Google’s own code

Simon Willison quoted a passage from an LA Times report about Gemini’s rollout inside Google: early on, employees reportedly faced restrictions on using Gemini to write or analyze software because proprietary code might leak into the model’s training data. During verification, the LA Times article was available mainly through search metadata and the quoted excerpt, not as a fully open article.

Google also had to slow down its own Gemini usage

The useful point is that this is not the generic fear of a conservative company. If the quoted passage is accurate, the same risk pattern appeared inside Google, the company building and integrating Gemini.

That does not prove Gemini trained on employees’ internal code. It says something narrower and more important for buyers: rolling AI into software work requires knowing what is sent to the model, where it is stored and whether it can be used for later training.

Enterprise adoption depends on boundaries, not demos

For engineering teams, this episode is more useful than another video of generated code. It shows that the hard blocker is often not whether the model can write a function, but permissions, audit and data policy around the tool.

The same question appears with every coding assistant. Can the agent read the whole repository? Can it inspect security bugs? Can it touch customer data in tests? Without crisp answers, adoption turns into a patchwork of local exceptions and bans.

The excerpt is not the full story of Gemini’s rollout

Because the full LA Times article was not openly available during verification, this piece should not make detailed claims about the causes of Gemini delays or Google’s exact internal rules. The safe ground is the quoted passage and the broader enterprise governance lesson.

Timing matters too. Early rollout restrictions may not describe the current product policy. With stories like this, the difference between historical caution and present day controls is not a footnote.

Contracts and logs will matter more than launch copy

The next signals are in enterprise documentation: default training settings for customer data, retention periods, audit logs and separation between consumer and business modes. That is where we will learn whether the tool is fit for sensitive code or only for the safer edges of work.

For buyers, the story is a checklist. If model makers hesitate to put their own code into AI tools without rules, outside companies should not sign an order just because the demo looks fast.

Lilith's verdict

A coding agent without data boundaries is an intern with the vault key. It may be talented, but the first question is who stands at the door and writes down what it carried out.