Does Jarvis Train Its AI Models on My Data?

No. Jarvis (getjarvis.eu) never trains AI on your data, full stop. Your prompts, screenshots, persistent memory, and connector content from Gmail, Slack, or Notion are not used to train or fine-tune any model. Jarvis routes each task to frontier models from Anthropic, OpenAI, or Google for inference only, under API terms with Anthropic, OpenAI, and Google that exclude training on submitted data. Your information is used to answer your request, then it is not recycled into model weights. This is a core commitment, reinforced by EU data residency and GDPR plus EU AI Act alignment. Jarvis is the screen-aware desktop assistant (floating bar, Cmd+/ on macOS, Ctrl+/ on Windows) for macOS 12+, Windows 10+, and Linux. Pricing is free to start, then $16/month for Pro.

There is an important distinction. When Jarvis sends your prompt to a model, that is inference: the model reads the input and produces an answer in the moment. Training is a separate process where data is absorbed into a model's parameters and influences future answers for everyone. Jarvis only ever does the former. Your input shapes your own answer and nothing else, and it does not leak into responses other people receive, which is the real privacy fear behind the question.

The guarantee rests on the commercial API terms Jarvis uses with Anthropic, OpenAI, and Google, which contractually exclude customer inputs from training pipelines, plus Jarvis's own policy of not retaining inputs to build internal datasets. There is no shadow corpus of user prompts being assembled for a future Jarvis model. Combined with EU data residency and the no-retention design for screenshots and connector data, there is simply nothing positioned to feed a training run.

Privacy & data