AI Transparency

1. How the DFO AI works

The Development Finance Observatory features an AI-assisted search workflow that helps you navigate, filter, and synthesize development finance data. The AI agent is designed to answer only from a focused set of pre-vetted, official data sources — including OECD DAC, World Bank IDS, and IMF databases — and does not draw on the open web or general knowledge to generate figures.

If a question falls outside this whitelisted set of sources, the assistant will tell you so rather than provide an answer. It is not able to guarantee the complete elimination of errors, and we encourage users to verify any figure used for high-stakes decisions against the cited primary source.

The same data is available to other AI assistants, such as Claude and ChatGPT, through the DFO’s connector. There the DFO supplies the figures and their sources, and the assistant you chose writes the answer under its provider’s own terms.

2. Data provenance and source grounding

We prioritize precision over automation. Every figure the assistant returns is paired with a citation back to the primary source institution (e.g., OECD, World Bank, IMF) and, where applicable, the specific dataset and reporting year. The assistant is restricted from generating a standalone answer when no matching source can be retrieved — in that case, it will say the data isn’t available rather than estimate or infer a figure.

If you believe the DFO produced an inaccurate or unsourced answer, report it to [email protected].

3. Data privacy commitments

Your privacy is a core part of how this platform is built:

  • No file collection. The DFO does not accept or process external file uploads from users, so no external proprietary data or files enter the system pipeline.
  • How questions are processed. The text of your question is sent to Google’s Gemini models, which read it to find the data you asked for and, in the Studio, to write the answer.
  • What a search records. When you search on this site, we record the first 500 characters of your question together with the data it matched, to learn what people look for and where the data falls short. The record is tied to your account if you are signed in, and to a random identifier in your browser if you accepted usage statistics in the cookie banner. These records are held in the DFO’s own database and in PostHog, an analytics service hosted in the EU.
  • Studio conversations. The Studio stores each conversation so that it can continue from one message to the next. If you are signed in, the conversation is also saved to your account until you delete it or delete your account. A technical record of each conversation, including its text, is kept in Langfuse, a monitoring service hosted in the EU, so that we can find and fix faults. Usage statistics record that a prompt was sent and how long it was, and leave out its text.
  • Connected assistants. When you use the DFO from another assistant, the DFO receives the question and the data requests that assistant sends. It needs no account and is not told who you are. Charts and tables the assistant creates through the DFO are stored so that they can be shown, and anyone who has a chart’s link can open it.
  • Internet addresses. Our servers’ request logs record the internet address of each request, as on any website. Usage statistics do not include it.

For a question about your data, or for help with the DFO, write to [email protected]. ONE’s privacy policy covers the rest of this website.