Adding Context for AI Forecasts

When Hinsley generates an AI forecast for a question, it draws on its own research into the question. Editors can add two kinds of context to that research:

  • Documents - Up to 10 web pages, such as articles or reports, that the AI forecasters always read.
  • Linked questions - Up to 5 other forecasting questions in your account whose current forecasts bear on this one.

Both kinds of context work for analysis-linked and account-level questions. Changes apply from the next AI forecast, whether it runs on the question’s schedule or is triggered manually. To set up AI forecasting on a question, see Creating Forecasting Questions.

Opening the Sources Tab

On the question page, open the gear menu and choose “Edit question”. This opens the “Manage question” drawer. Click its “Sources” tab.

The Sources tab is available to anyone who can edit the question: collaborators with edit access on an analysis-linked question, or account members who can edit an account-level question. It lists the question’s documents and linked questions, each with a count of how many of the available slots are used (for example, “3 of 10”).

Documents

Adding Documents

Click “Add URLs”, paste one or more URLs into the “Add URLs, one per line” box, and click “Add URLs” again. Hinsley fetches each new page in the background.

A URL is not added if it is not a valid web address, if it is already in the list, or if adding it would go past the limit of 10 documents. Hinsley lists each rejected URL with its reason and leaves it in the box, so you can correct it and try again.

If Hinsley’s research had already found a URL you paste, it is not added a second time. The existing source is set to High relevance, and the AI forecasters always read it from then on. Pasted URLs are also set to High relevance.

Document Status

Each document shows its title and a status badge. Below the title, a document you pasted shows its source domain and who added it and when. A URL that Hinsley’s research had already found shows “Found by our research” with the date it was provided. A document that is not readable shows the failure note in place of the domain and attribution.

The status badge is one of:

  • Fetching - Hinsley is still retrieving and reading the page.
  • Ready - The page has been read and is included in AI forecasts.
  • Not readable - Hinsley could not read the page, so AI forecasts skip it.

Documents that are still fetching or not readable are left out of AI forecasts. The same badge appears on the question page’s Sources tab until the document is ready.

Removing Documents

Click “Remove” next to a document and confirm. A URL you pasted is removed from the question. A URL that Hinsley’s research had found stays among the question’s research sources, and AI forecasts stop treating it as an editor-provided document.

Linked Questions

Linking a Question

Click “Link a question” and search by question name. With no search entered, the list shows recent questions. Each result shows the question’s current forecast. You can link any question in your account that you have access to, from any analysis or at the account level. The question itself and questions that are already linked are not offered.

Click “Select” on a result, optionally add a note describing how the two questions relate (up to 2,000 characters), and click “Link question”. The AI forecasters read the note beside the linked question.

Audience Warnings

Because an AI rationale can quote a linked question’s forecast, Hinsley shows a warning before you link a question that some readers of this question may not otherwise see:

  • Different analysis - The selected question belongs to a different analysis, which some people who can see this question may not be able to open.
  • Forecaster network - This question is posted to a forecaster network and the selected question is not. Network contributors can read this question’s AI rationale, including anything it quotes from the selected question.

The warnings do not prevent the link.

Managing Linked Questions

Each linked question in the list shows its name, its current AI and human forecasts (or its resolution, if it has been resolved or voided), and its note. Edit the note and click “Save” to update it. Click “Remove” and confirm to unlink the question.

The Linked Questions Card

On the question page, the Sources tab opens with a read-only “Linked questions” card. For each linked question it shows the name, its current forecasts or resolution, the date it closes or was resolved, the date it was linked, and its note.

Each viewer sees only the linked questions they have access to. If a viewer cannot open any of them, the card is hidden.

How AI Forecasts Use This Context

Each AI forecast reads every ready document, in addition to Hinsley’s own research. On later forecasts, the AI forecasters are told which documents were added since the previous forecast. Documents are treated as evidence to weigh and cite.

For each linked question, the AI forecasters read the question, its note, and either its current forecasts with their rationale summaries or its resolution. They treat these forecasts as other forecasters’ views, and reason about how each one bears on this question.

Citations to Linked Questions

When an AI rationale draws on a linked question, it cites it as [Q1], [Q2], and so on, numbered in the order they are first cited. Each citation links to the linked question. Hovering over it shows the question’s name, current forecasts, and dates, provided you have access to that question.

Question citations are removed from rationales shown on forecaster networks.