AI policy

In short

1. In principle, we only use AI to interpret what you give it, not to invent things from scratch. In practice, that means using a model to generate flashcards from what you supply - everything else about Flora is simple, old-school technology.
2. Flora always leaves room for you to take control: you can use it without any AI features at all. Everything is yours to edit, overrule, or extend - what you add, and anything AI writes from it.
3. Everything you read on this site is written by a person, not generated.

What do we mean by “AI”?

Large language models (LLMs), and other models of the same scale that would be impossible to run on a laptop. Smaller machine learning models, or simpler algorithms, don’t count as AI here.

Flora uses AI to generate flashcards

When you give Flora a new source - a webpage, a PDF, a YouTube video - we send its text to a model, with precise instructions for pulling out claims that would make a good card.

Claims on generated cards don’t go beyond what the source says; each card is attached to a quote that backs up its question and answer. If a core claim isn’t in the document you supplied, Flora won’t fill the gap for you.

It doesn’t decide how and when you review cards

Today’s cards are ordered by a fixed scheduling algorithm, not a model - deterministic, classic software, built on Jarrett Ye’s research into spaced repetition.

Which model(s) does Flora use?

New AI models are released so often that marrying Flora to a single model doesn’t make much sense. Instead, Flora commits to a high level of performance against small set of tasks, like generating cards.

Behind the scenes, we’re continuously running experiments to determine which models most efficiently deliver the results that you expect. In general, Flora uses the smallest possible model which delivers the best possible performance against those internal benchmarking and evaluation systems, for the sake of cost, speed, and environmental impact.

None of the third-party models used by Flora are allowed to use your data for training.

What gets sent to an LLM

The text of the document you added, and a set of instructions for turning it into cards. Your personal data, your review history, and everything else about your account stay within Flora’s systems.

The LLM providers we work with are forbidden from training new models based on the prompts we send (ie your documents).

Where we use AI behind the scenes

Like most software developers these days, we use LLMs to help write the Flora codebase - this app is only made possible because of the enormous productivity boost which comes from collaborating with these systems. We’ve been developing software professionally with and about machine learning models for more than a decade, so we’re confident in our ability to avoid creating slop.

While we’re committed to writing everything on this site by hand, we also use LLMs as research assistants to search for, structure, and synthesise the academic literature and references which underpin Flora’s theory of change, to research around our writing on the app’s blog, and to proofread our writing before it’s published.