Skip to main content
Ledger
Ledger is nonpartisan. Ledger does not endorse candidates, campaigns, or parties.
LegalAI Policyv2.1.0
Effective: October 5, 2026View current version

The Ledger iOS app is covered by its own policies: iOS app Terms of Service · iOS app Privacy Policy · iOS app AI Policy

AI Policy

Quick summary

  • Ledger has no AI assistant. There is no chat, no explain feature, and no AI-written answer anywhere on the board.
  • Ledger uses AI in four places, all inside its own review and publishing process: helping founders draft proposed polls, checking proposed policy polls against Ledger's editorial rules, an advisory screen of stock photographs before a person vets them, and choosing which already-vetted library photograph accompanies a poll.
  • AI never publishes. A Ledger founder reviews and approves every poll and every social post. A model's output is a recommendation to that founder, nothing more.
  • No personal data goes to a model. The models see poll text, library photographs, and the photographs' descriptions, and a drafting assistant sees public news pages and a founder's instructions, never anything a user entered.
  • The editorial rules are the same for people and models: neutral questions, literal story text, sourced topics, no persuasion.

1. What this policy covers

This policy describes how Ledger uses AI. Ledger is a nonpartisan public board of public-policy polls: each policy poll is a current policy question tied to a sourced news story, answered with a two-way choice, and each race poll asks who the person expects to win a 2026 federal race. The polls, their text, their images, and their results are what visitors see. None of it is generated for a visitor on request.

Ledger's AI provider for checking polls and for photographs is Anthropic (Claude models). Those model versions are pinned and changed only through a reviewed code change. A founder drafting a poll may use Anthropic's Claude or OpenAI's models.

2. Checking proposed polls

Every proposed policy poll passes two checks before a founder sees it:

  1. Fixed checks, no AI. The draft must be complete, cite only sources from Ledger's vetted list of outlets, and concern a current story. A draft citing any other source is rejected.
  2. An AI review. A model reads the draft's own text against five of Ledger's editorial rules — the question is a neutral support/oppose with no loaded premise, it is tied to a concrete policy action rather than a headline or a personality, the summary states that action's effect neutrally, no candidate or official is characterized, and nothing nudges the reader — and returns pass, borderline, or fail with a short explanation. If the model cannot complete its review, the result is recorded as an error, never as a pass or a fail.

A poll about a race, whose options are the candidates on the state's certified list, is instead checked against that list, and no model is involved.

The model does not judge whether the headline and summary are faithful to the source, or whether a story came from a campaign or advocacy group; the founder checks those. The founder sees the model's verdict and explanation beside the draft, then decides. Only a founder's approval publishes a poll. A poll never goes live because a model passed it.

Automated news drafting is switched off. Ledger's code contains a feature that drafts candidate poll questions from news feeds with an AI model. It is switched off in production. If Ledger turns it on, a new version of this AI Policy will say so first, and every drafted poll will still pass the checks above and a founder's publish decision.

Founders may draft polls with an AI assistant. A Ledger founder may use a general-purpose AI assistant (Anthropic's Claude or OpenAI's models) to find a current story at a vetted source and draft the neutral question and the statement of what the policy would do. The assistant works from the founder's instructions and public news pages, and receives no user data. A founder edits every draft, every draft is checked (by the model check, or against the certified list for race polls), and only a founder's decision publishes it. This is separate from the news-drafting feature, which stays switched off.

3. Stock photographs

A poll may carry one photograph from Ledger's stock library. When a photograph is added to the library, a model checks it against Ledger's rule for poll imagery — symbolic or topical only, never a real person or a specific real event as the subject — and writes one plain sentence describing what it shows. The check is advisory: a person vets every photograph before it can be attached to a poll. If the model is unavailable, the photograph waits for that person's review like any other. Ledger does not use AI to generate images.

When a proposed poll reaches founder review, a model chooses which photograph from the library accompanies it. The model sees the poll's text and the one-line descriptions of the photographs it can choose from, and a person has already vetted every one of those photographs. The model never adds a photograph to the library.

4. What AI does not do on Ledger

  • Answer questions from visitors. There is no assistant, chat, or explain feature.
  • Publish polls or social posts. Founders do.
  • Decide what counts as neutral on its own. It may flag; a founder decides.
  • See personal data. No email, name, answer, reaction, or IP address is ever sent to a model.
  • Train on your data. Ledger does not use your answers or reactions to train any model, and Anthropic's API terms apply to the drafts and images Ledger sends.
  • Choose which results you see or in what order. Result ordering on the board is mechanical — by response count and by 24-hour change — and is never set by a model.

5. Versioning and review

This document is versioned. Ledger publishes only the current version; the version history is summarized at /legal, and the full text of any earlier version is available on request to the support address on the contact page. Revisions happen when the rules for Ledger's AI use change. Routine wording cleanups do not trigger a version change.

6. Effective date and version

  • Version: v2.1.0
  • Effective date: October 5, 2026
  • Prior versions: v2.0.0, v1.2.0, v1.1.0, v1.0.0.

What changed in v2.1.0. A Ledger founder may now draft proposed polls with a general-purpose AI assistant, Anthropic's Claude or OpenAI's models. This version adds that use, says what the assistant sees, and confirms that a founder edits every draft and that only a founder publishes. It also describes race polls, which ask who a person expects to win a 2026 federal race, and notes that a race poll is checked against the state's certified candidate list, with no model involved. It also corrects §2: only sources on the vetted list are accepted, and the model judges five rules from the draft's own text while the founder checks source fidelity and campaign origin.