Governance
A short card for each public build: what data it uses, what could go wrong, where a person checks the work, and what it shouldn't be used for.
AI governance frameworks ask organisations to write these things down before something goes wrong. These tools are small, but several of them write or check text automatically and some hold people's email addresses, so the same questions apply. Writing the answers down is also how I practise what I'm studying for the AIGP.
Statute Atlas
- Data
- New Zealand legislation from the Parliamentary Counsel Office. Only derived statistics are stored; no legislation text is republished. No personal data, no accounts.
- Main risks
- A count from pattern matching read as an exact tally; a chart's headline overstating what thin or incomplete data shows.
- Safeguards
- Each pattern-based measure is spot-checked on a random sample by two independent AI reviewers, and the result is shown under the chart. Headlines are worded from the data's direction, and superlatives say "so far" until the full statute book is read.
- Human review
- I settle any case the two reviewers disagree on, and approve each change to how a measure works.
- Not for
- Legal advice, or the legal status of any particular Act. Check legislation.govt.nz for that.
Evidence Reviews
- Data
- Public documents: council papers, consultations, official statistics and published research. Reader accounts hold an email address and saved preferences.
- Main risks
- A wrong or unsourced fact presented as checked; a review read as taking a political side.
- Safeguards
- Every fact is checked against its source before release, and reviews that fail the check don't publish. Reviews never name a political party, and central-government topics are held during a general election period.
- Human review
- Nothing goes on the site or out by email until I release it.
- Not for
- Legal advice, or the final word on a council decision.
Second Look
- Data
- What charities file on the public Charities Register, and funders' own published rules and dates. Accounts, if used, hold an email address and saved funds.
- Main risks
- A charity judged unfairly from its numbers; a group missing a funding round because a date was wrong.
- Safeguards
- Reports explain what the figures can and can't show. Funding details are read from each funder's own page and fact checked separately, and each fund links to its source.
- Human review
- New and changed funders are reviewed before they go live.
- Not for
- A verdict on whether a charity is well run, or a substitute for reading the funder's own guidelines.
Verso
- Data
- Auction house catalogues and published references on artists and provenance. Newsletter readers give an email address.
- Main risks
- A wrong attribution, date or provenance detail that affects what someone bids.
- Safeguards
- Catalogue facts are re-checked against the auction house's listing, and anything that can't be confirmed is marked as unconfirmed rather than stated.
- Human review
- I approve each Sunday List before it sends.
- Not for
- A valuation, or a replacement for the auction house's condition report.
More cards will be added as builds go public, and Lab prototypes get one if they become builds.