Breathe Better. Live Smarter.®

Archived version · pinned v2026-10-15

Language Model Disclosures

Effective October 15, 2026 · Version 2.1 · Reference page. The Terms of Service and Privacy Policy control.

This page explains where Terrestream uses language models (software that writes text), what we send to the companies that run them, and which features are written from fixed templates instead. The Privacy Policy refers to this page.

1. Summary

  • Support assistant. Our help search at terrestream.com/support and the Ask Terrestream conversation on public website pages can ask a language model from Anthropic for short answers grounded in our help content.
  • No account or sensor access. We do not attach account details, sensor readings or home location to model requests. The question, recent conversation context and public page category are sent as described in section 2. Suggestions, notifications, Discovery cards and Year in Air use our own rules and fixed templates (section 4).
  • We don't use your data to train language models.

2. Support assistant

2.1 What it does

The help search at terrestream.com/support finds help articles that match what you type. If you type a question of at least eight characters and pause, our server may also ask a language model to write a short answer using only the matching Terrestream help content, such as setup guides, articles and sections of our legal documents. The answer comes with links to the help sources it was given.

The Ask Terrestream conversation sends a request only when you press Send. Opening the panel, typing or choosing a suggested question does not ask the model. Each reply uses fresh help content; previous messages explain your issue but are not a source of product facts. Both interfaces show a language-model notice and ask you not to include personal information.

2.2 Provider

Anthropic (United States), using its Claude models. Our server sends the request; your browser doesn't contact Anthropic directly.

2.3 What is sent

  • Your help-search question (up to 300 characters), or your submitted conversation question (up to 1,000 characters).
  • For conversation replies, up to six recent user and assistant messages, with up to 1,000 characters per message and 6,000 total, plus a public page category such as setup or product. We do not attach the full page URL, query string or form contents.
  • Excerpts from the matching Terrestream help content.
  • Instructions that tell the model to use only that content, not to invent anything, and not to give legal advice or medical diagnosis or treatment advice.

We do not attach your account details, sensor readings or home location to the request. Your own messages may contain information you enter. We screen for some sensitive patterns, but this filter cannot detect every kind of personal information. Do not enter personal information or credentials in either support interface.

2.4 What we keep

For help-search questions, our server caches generated answers for up to 24 hours and fallback answers for up to five minutes. The copy contains the answer and source links under a one-way code derived from the question; the question itself is not stored.

Conversation replies do not use this answer cache, and we do not keep a server transcript. Up to six recent messages stay in the current browser tab. Start over clears them; after 30 minutes without interaction they expire and are removed when the support code next runs, including when a suspended tab resumes. The Privacy Policy describes provider processing; these application retention limits are not a promise of zero retention by the provider.

If you choose human help, you can review and edit a short summary of your own messages before opening a WhatsApp draft. Our team receives it only when you send it in WhatsApp. WhatsApp applies its own terms and privacy policy. With analytics consent, we count fixed support actions and placements, without questions, transcripts, summaries or draft URLs.

2.5 Limits

Answers can be wrong or incomplete. They aren't legal, medical or safety advice. The linked help articles and the Terms control. If the answer doesn't help, contact support.

3. Climate-news summaries

The current apps and web dashboard do not offer language-model climate-news summaries.

3.1 What it does

The news panel shows publisher text, as described in section 4.

3.2 Providers

No language-model provider is used to write the news panel shown in the current apps or web dashboard.

3.3 What is sent

That panel does not send your information to a language model.

3.4 Checks

If a future release adds model-written news, we will update this disclosure before offering it.

4. Where we don't use language models

These features don't use a language model. They come from our own rules, calculations and fixed templates, applied to your home's data:

  • readings, bands and the overall air rating for your home;
  • suggestions, alerts and notifications;
  • weekly Discovery cards;
  • Year in Air, including its letter;
  • weekly insights, room patterns and printable reports;
  • the local-conditions card on terrestream.com.

The news panel in the web dashboard shows each publisher's own headline and, where the publisher allows it, a short excerpt. No language model writes or summarizes it.

5. Your choices

The help search may request an answer when you pause while typing a question at terrestream.com/support. The conversation panel requests an answer only after you press Send. To avoid sending text to a language model, browse help articles or choose WhatsApp or the contact form. Start over clears the conversation in the current tab.

6. Limitations

  • Text written by a language model can be wrong, out of date or phrased more confidently than the facts support.
  • It isn't medical advice and doesn't diagnose anything.
  • It can't tell you whether smoke, carbon monoxide (CO), gas or another hazard is present. The sensor doesn't detect carbon monoxide — keep your CO alarms.
  • Don't use it as the only basis for a decision about health, safety, housing, insurance or emergencies.

Terms §5 covers outputs and suggestions in full.

7. Changes to this page

We update this page when we add, remove or change a use of a language model, or change a provider. Providers are also listed at /legal/sub-processors, and changes to that list follow the notice rules there.

8. Version history

  • October 15, 2026 · 2.1. Added the public website conversation, explicit Send behavior, bounded tab history, uncached conversation replies and reviewed WhatsApp handoff.
  • September 27, 2026 · 2.0. Retitled "Language Model Disclosures". Describes the uses that exist: the support assistant; climate-news summaries are not offered in the current clients. Discovery cards and Year in Air are written from templates. Removed the descriptions of trained models, model-written letters and a settings opt-out, none of which exist. Moved from /legal/ai-models.
  • June 1, 2026 · 1.0. Initial publication at /legal/ai-models.

Earlier versions are kept at /legal/archive.

9. Related documents

WhatsApp support