Language model
reading, finding claims, explaining.
A language model on your own account reads messages, e-mails, linked pages and screenshots, finds checkable claims, writes the short summary of a finished check, and powers the assistant. You choose the provider: Anthropic's Messages API or any Chat Completions compatible API, including a model you host yourself.
01What it does
The console's Assistant section configures one model, used by all of these:
| Use | What the model gets | What it returns |
|---|---|---|
| Message and e-mail read | The redacted text, inside a marked data block it cannot leave | Strict facts through a tool: what it asks for (money, sign-in details, a code, an installation, personal data, moving to another app or number), pressure, which brand it claims to be, the scam type |
| Page read in the link check | The page's text, redacted | What the page asks the visitor to do |
| Screenshot read | The image (PNG, JPEG or WebP up to 7 MB) | The text in it and what kind of screen it is (chat, e-mail, post, payment confirmation, other) |
| Claim finding | The redacted text, cut at 12,000 characters | At most three checkable claims |
| Summary | The findings only, no links | Two sentences; replaced by a fixed template when it reassures, names another level, is too long or holds a link |
| Assistant chat | The person's questions and the verdict, redacted | Explanations and proposals the person must tap |
The model's answers can only add findings; they never lower a level set by rules or providers, and the verdict words always come from code, not from the model. Without a model, messages, e-mails and screenshots get only the phrase rules, which never give Safety OK on their own ("AI read: not configured"), images skip the screenshot read ("Screenshot read: not configured"), and no claims are found.
The screenshot read sends the image to the same model. A model without image input fails that read, and the image then goes to the image checks only.
02The assistant
The assistant answers questions about a verdict, a message or what to do next. It can start checks and suggest a playbook or an official contact, but it only proposes: nothing happens until the person taps. Suggested links come only from curated places (your playbooks' official contacts, the built-in brand list and reviews on known fact-checker sites), never from the checked content or pages found on the web. A fixed level sentence starts every answer at Caution or Danger, and answers that sound too sure are replaced. People can report an answer ("Report this answer"); reports appear in the console under Reported answers. The app shows "Answers come from an automated assistant and can be wrong."
03Settings
Console, Assistant:
| Field | What to enter |
|---|---|
| Assistant | The switch. It is on by default, but nothing runs until a model and key are set. Turning it off also turns off the reads, claim finding and summaries above. |
| API type | "Chat Completions (compatible providers and local models)" (default) or "Messages API" (Anthropic). Owner only; changing it clears the saved key. |
| API base URL | Empty: the vendor's own address (https://api.openai.com/v1 or https://api.anthropic.com). For another provider or a local model, its address. Owner only; changing it clears the saved key. |
| Model | The model id exactly as your provider names it. |
| Model for Pro installs | Optional. Empty: Pro installs use the model above. The app then shows "Assistant Pro". Claim finding and summaries of Pro installs use it too; the message, e-mail, page and screenshot reads always use the main model. |
| API key | Your provider's key. Owner only. |
| Provider name shown to users | Named in the app's first-use notice that says where assistant messages are sent. |
| Extra instructions | Added to the built-in instructions. They cannot remove the built-in rules. |
| Assistant messages per day, all installs | A cap on your bill (default 5,000). 0 = no cap. |
Unlike the research and speech-to-text addresses, the assistant's address may point at your own network, so a model on your own hardware works. Its host then appears on the consent screen of every user (see Consent).
04Get a key
- Anthropic: Claude Developer Platform, API keys. Choose "Messages API" as the API type.
- OpenAI or another Chat Completions provider: the provider's own dashboard. Keep "Chat Completions" and enter its base URL if it is not OpenAI.
- A local model: any server that offers the Chat Completions API, for example on your own machine; enter its address as the base URL.
05Cost
Billed per token to your provider account, at its price. Most checks call the model at least once (the read and the summary), so it is usually the largest share of a light check's cost. A model you run on your own hardware has no per-token bill.