How we use AI ​

Short version: AI helps us build the catalog and the site. It never sees you. Nothing you type, pick or are shown in the guide reaches a model, and no model decides anything about your accounts.

The test every use has to pass ​

A piece of AI belongs in SecBird only if it

1. makes existing information understood: it never invents, judges or decides;
2. shifts responsibility back from you instead of adding a new thing to babysit;
3. feeds checkable facts: its output is a quote with a source, or it is not a fact;
4. costs less than it saves: compute scales with the catalog (hundreds of documents), never with people (millions of visits).

So the AI in this product runs at build time, on public provider documents and on our own code. A fact it decides reaches you only above a confidence line, wears a badge that says so, and can be reported wrong in one click; people correct, they do not pre-approve (we have no team for that, and we would rather be honest about it).

What AI does here today ​

• Writes code and documentation with us. Claude (Anthropic), through Claude Code, drafts changes to this site and its data pipeline. Every change goes through a pull request, the test suite and a human merge.
• Researches services for the catalog. For the hand-researched tier, agents read a provider's own help pages and terms and draft the links, settings paths and evidence rows. A quote is copied verbatim or the row is marked unquoted. Those rows are labelled ai:claude-fable-5-1 and carry the status proposed; a proposed row is not shown on the public scorecard.
• Decides the facts on the scorecard and finds the settings pages. A decision model (below) answers typed questions over a provider's own document and points at the sentence that states each fact, or at the help page for a step. When its confidence is 80% or higher the fact or link is shown, with a small badge. The badge's tooltip tells you what decided it, how confident it was, the rule, and the quote with its source and date. Below 80% nothing is shown.
• Draws and speaks. The illustrations, the short sound cues, the narration voice and the secretarybird's head in our mark were generated with ElevenLabs tools from briefs we wrote. The words are always ours; the voice only reads them.

What AI does not do ​

• No chat, no assistant, no "ask SecBird".
• No reading of your account list, answers, alias or report. The guide keeps all of that in your browser; see what we store.
• No scores, grades or judgements. The clarity band is arithmetic over checkable facts; readability is a formula, not an opinion.
• No notifications, no summaries of policy changes, no interpretation of breaches.
• No decisions about you. The care levels and plans are rule-based.

Models and prompts ​

The generative extraction (a job that proposes rows from the terms snapshot) exists but its rows are not shown: only the decision model's rows reach the page, under the rule below.

A decision model, and what it never does ​

The second kind of model answers typed questions about a document (which of these options, yes or no with what probability) and cannot write text at all. It decides the facts on the scorecard and points at the sentence that states each one; the quote is always our own copy of the provider's words. The rule, decided on 2026-10-29: a decision with confidence of 80% or higher is shown at once, with a badge; below that, nothing. The number in the badge is the model's own probability, not a calibrated one: we had planned to calibrate it on rows people verified and dropped that plan for lack of hands. If a decided fact is wrong, "Something wrong?" on the badge tells us, and a person settles it: confirmed rows lose the badge, wrong ones disappear. It never sees anything you type, never scores a service, and never runs inside the guide.

On the listed tier it also decides, from a brand's homepage title and description, which category a service belongs to and whether a brand is a service where people hold accounts at all. Categories apply above a fixed cut-off; admissions wait until a person has checked a random hundred and set the cut-off where precision holds 0.95.

No service has been admitted to the list by the decision model yet. 130 listed services have a category it decided (the rest come from 2FA Directory keywords or stay "other").

What it costs ​

We publish the bill because "AI" is often a synonym for waste. Compute here scales with the catalog, not with you: a visit to the guide costs zero model calls.

Decision-model calls are metered the same way (tokens in, nothing out) and capped at ten dollars a month by the client itself.

Energy is not estimated. No provider we use publishes a per-token or per-credit figure we could cite, and a made-up factor would be worse than a blank. The raw ledger is at /ai-ledger.json.

Disputing an AI-decided row ​

Every evidence row shows its contributor and status. If a decided row is wrong, use "Something wrong?" on its badge, "Suggest a correction" on the step, or open an issue; the row becomes disputed and is hidden while it is open, and ends verified or removed. The process is the same as for a row a person wrote, described in our contribution guide.

The fine print ​

Nothing on this site is legal advice, and that includes anything an AI drafted or decided. Verified and decided rows are published under CC BY 4.0 regardless of who or what drafted them, with the confidence kept on the decided ones. Model names are trademarks of their providers; our use of Claude follows Anthropic's usage policy, and the ElevenLabs voices and images are used under the account's licence.