Pensio makes one promise about the AI that reads your journal: what it sends is not stored, and it is not used for training. Every request asks for both, and if no model can promise both, the request is refused rather than sent somewhere that could keep it.

This month, the service Pensio reaches AI models through told me that Google's models would stop qualifying on 18 October, because Google may keep requests that its safety checks flag. That is a fair rule for them. It is not the promise I made you.

So before that date, everything that ran on Google's models moved to models that still promise both. In Explore, the suggested follow-up questions and the conversation titles now come from Mistral AI, a company in France. I tested each replacement first on Explore conversations in English, Portuguese and Spanish, and only switched when it did the job as well.

One more change this month: the model that reads your entries is newer. In my tests it added far fewer things you never wrote, and it reads feelings a little more gently, with intensity about 13% lower on average. Entries it already read keep their reading, so your chart may dip slightly from this week.

The privacy page now names every AI company that handles your words, and what each one does with them.

Open the Privacy Policy to see who handles what.

— Gabriel