AI Use Policy

1. What We Use AI For

Chatterbox uses AI to accelerate the processing of public information — not to generate or invent it. Specific uses:

• Source-grounded summarization (Claude Sonnet 4.6). When an indexed source document is too long for a briefing card, we use Claude Sonnet 4.6 to produce a summary that is constrained to facts present in the source text. Every summary must cite a passage from the underlying source. If the model cannot produce a grounded summary, it refuses rather than inventing content.
• Sensitivity and topic classification (Claude Haiku 4.5). We use Claude Haiku 4.5 to classify whether an item falls into a sensitive category (crime, arrests, deaths, minors, schools, elections, lawsuits, public health) and to assign a topic label. Classification is used to route items to the review queue; it is never used to suppress items without human review.
• Embedding for deduplication and search. We generate vector embeddings to identify duplicate items across sources and to power semantic search on location pages. Embeddings operate on source text only and do not generate new content.
• Geocoding assist. We use AI to extract and normalize geographic references from source text (street addresses, ZIP codes, county names) to attach items to the correct location nodes in our geography graph. Geocoding results are validated against our canonical geography database.

2. What We Never Use AI For

• Inventing facts. AI models are never prompted to generate information that is not present in the indexed source text. Prompts include explicit instructions to refuse if the source does not contain the requested information.
• Rewriting partner content beyond fair-use summarization. We do not use AI to paraphrase or rewrite articles from partner publishers in ways that would substitute for reading the original.
• Treating social-media claims as facts. AI classifiers and summarizers are not fed social-media content as a factual input. Social-media posts are treated as tips requiring independent verification, not as sources to summarize.
• Making decisions on sensitive content without human review. AI classification routes sensitive items to a human review queue. The AI does not approve, reject, or publish sensitive items. Human review is mandatory.

3. Source-Grounding Requirement

Every AI summary must cite a specific passage from the underlying source document. The summarization prompt explicitly instructs the model to: (a) base the summary only on the provided text; (b) cite the passage that supports each claim in the summary; and (c) refuse — returning a structured refusal response — if the source text does not contain enough information to answer the request. Refusals are logged and the item is held unpublished until a human resolves the gap.

4. Human Review

Human review is mandatory for any item that: (a) falls into a sensitive category (crime, arrests, deaths, minors, schools, elections, lawsuits, public health); (b) has an AI confidence score below the configured threshold; or (c) is a community submission. Human editors see the full classification output, the confidence score, the underlying source, and the AI-generated summary before making a publish decision. Editors may edit, reject, or approve; they may not bypass the queue to publish directly.

5. Disclosure

Every item that involved AI assistance in its summarization carries the AI-Assisted Summary provenance label, visible on the item card and in briefing emails. The label links to this policy. You have the right to know when AI was used to process the information you are reading, and we treat that disclosure as non-negotiable — not as fine print.

6. Backup Providers

Our primary AI provider is Anthropic (direct API). In the event of a service outage or degraded performance, we may route inference through OpenRouter (which proxies Anthropic models) or AWS Bedrock (which hosts Anthropic models under AWS's infrastructure). In all cases, the same model family (Claude Sonnet 4.6 for summarization, Claude Haiku 4.5 for classification) and the same source-grounding prompts are used. We list these vendors here so you understand who may process source text on our behalf.

7. Data Sent to Providers

We send AI providers only the specific source text needed to complete the requested task (summarization, classification, geocoding, or embedding). We never send user personally identifiable information (PII) — including email addresses, location preferences, or subscription data — to any AI provider. Prompt payloads are constructed from indexed source documents and our structured prompts only.

8. Auditability

For every published item that involved AI, we record in our audit log: the model used (including version), the provider route (direct, OpenRouter, or Bedrock), the prompt template version, the confidence score returned, and the reviewing editor (if applicable). This record is retained for the life of the item and is available to editors for review.

9. Updates to This Policy

We will publish material changes to this policy — including adding new AI use cases, changing providers, or changing the scope of what AI can decide — before they take effect. The “Last updated” date above reflects the most recent revision. We will not silently expand AI use to new categories without updating this document.