Last updated: October 20, 2026
Northern Inference Inc. ("we", "us", "our") is a Canadian company committed to data privacy and transparency. This policy covers both our website (northerninference.ca) and our API platform. We explain what data we collect, where it goes, and what your rights are.
Account and identity data:
API usage data:
Scheduled monitors. If you create a scheduled monitor, we store the prompt you provide, encrypted at rest in a form our servers can read, so it can run while you are away. This permission applies only to the scheduled prompt you save for that monitor. It does not extend to storing your other chat messages or answers in server-readable form.
Chat agent in the portal:
If you turn on location in the iPhone app, each message you send includes your approximate city, your precise location or a place you typed, as you choose, so the model can answer questions about places nearby. Northern Inference does not store it, apart from the full prompts that Content Debug Mode logs when it is on for your account. It goes to the model provider with your message, and on a route without zero data retention that you allowed, the provider may keep it under its own policy. If you use web search, the model can include the place in a search query, which is sent to Brave in the United States and kept by Brave for up to 90 days, and which is saved with the chat if you save it, as described under Web search and web pages in chat. To name the place from your device's location, the app uses Apple's geocoding service. A place you type is kept on your device until you clear it.
Website and portal analytics:
Support data:
Voice input and sharing a chat:
Web search and web pages in chat. If you turn on web search and page reading, or allow it for one answer, the search queries the model writes are sent to Brave Software, Inc. in the United States, and may be processed in other countries where Brave's service providers operate. A query can include words from your conversation, including names or details you wrote. You can see every query in the chat. Brave keeps search query logs for up to 90 days for billing, troubleshooting and abuse prevention. We don't send Brave your account details or your network address; Brave receives the network address of our search server. Information sent outside Canada may be accessible to courts, law enforcement and national security authorities of those countries under their laws. When you turn on or allow web search, your search queries are sent to the United States, including in chats with Tier 3 models. When the model reads a web page, our servers in Canada request the page from the website that hosts it. That website receives the page address and our server's network address, not yours, and can see that the request came from Northern Inference.
We don't keep the search results or the text of the pages read; they exist in memory while your answer is written. If you save the chat, your queries, the addresses of the pages read, and the address and title of each search result the answer used are saved with it, encrypted so that we can't read them, so you can open an answer's sources when you come back to it. The text of pages is not saved. We keep a record that each search or page fetch took place, when, and what it cost (a page fetch is free), without the query, the results or the page address, like the record of any other request. We also keep a record of each choice you make about web search (turning it on or off, allowing it once or declining), with the time and the version of the wording you saw. Our search and page-reading servers keep technical logs (timing, sizes and status, and for page reads a pseudonym for you that changes every day) for 14 days. To answer abuse complaints, we keep a coded fingerprint of the address of each page read, with that day's pseudonym, for 30 days; we can match it only against an address someone reports to us. While Content Debug Mode or the Audit and Compliance tier is on for your account, what it keeps of a chat message includes the search queries and the addresses of the pages read, but not the search results or the page text.
Chat agent: documents, ratings, export and search:
Chat and Tier 4 models. The chat agent offers Tier 3 models, processed in Canada, and Tier 4 models, processed outside Canada. The model list groups them under those headings, and each answer shows its tier and the jurisdictions recorded for it. When you choose a Tier 4 model, your message, the earlier messages of the chat, your attachments, and any instructions, project files and memories the chat adds are sent to a provider that may process them outside Canada. Courts, law enforcement and national security authorities in those places may be able to obtain access to information held there under their laws. The same can apply to a provider that operates in Canada but is headquartered abroad, as the Tier 3 row below notes.
Image generation in chat (where available). Where the chat agent offers image generation, it uses a Tier 4 model provided by Google through Vertex AI. We send Google your prompt and any images you attach, so they may be processed outside Canada, and Google's own terms govern what Google does with them (see "What providers see" below). The generated image comes back through our service to your browser. We don't keep it. It is stored only if you save the chat, sealed like any other part of the chat, in the place you chose, except in the cases this policy describes for request content, such as Content Debug Mode. Generating an image is an API request, so we record the model, cost, route tier and request ID as for every request, and not the prompt or the image.
Our core product is routing inference requests to LLM providers. Where your data goes depends on the model route you select per request:
| Tier | Provider | Data Location | Jurisdiction |
|---|---|---|---|
| Tier 1: Self-Hosted Hardware (coming soon) | Hardware you own, or rent or lease from NI, routed by NI | Your premises; models run only on hardware you own, or rent or lease from NI (for example a data centre in a shipping container), and NI's router handles each request in memory to forward it | Yours |
| Tier 2: Sovereign Hardware (coming soon) | NI-owned, single-tenant bare metal | Canadian-owned data centre; inference on NI-owned hardware, not a US-owned cloud | Canada |
| Tier 3: Managed Cloud (live) | AWS Bedrock / Azure OpenAI in Canadian regions | Canada (ca-central-1 / Canada East); enforced at the routing layer and fails closed | Canada (US CLOUD Act applies to providers) |
| Tier 4: Global Access (live) | Provider-default and non-Canadian routes, including non-Canadian AWS Bedrock routes and Google Vertex AI | US, EU, APAC, or multi-region; the exact route is disclosed on every response | Disclosed per request |
What Northern Inference sees: Your prompts and completions pass through our routing infrastructure in Canada (ca-central-1) in memory only, for the duration of routing. We do not persist request content by default. Chain of custody records (entity, jurisdiction, timestamp per hop, but not content) are stored for 90 days. The only ways content is stored are the four opt-in exceptions described in section 1: PII Substitution mappings (transient), the Audit and Compliance tier (redacted exchanges), Content Debug Mode (full content, opt-in, described below), and the prompts you save for scheduled monitors (encrypted at rest in server-readable form).
Northern Inference does not use your content to train or fine-tune any model. That covers request content, chat messages, and everything in your chat documents.
Messages you send to the chat agent follow the same path as other API requests, including the exceptions in section 1. Under each answer the chat agent shows its privacy tier, the jurisdictions recorded for it in our chain of custody, and what it cost, with the full chain of custody one click away. These come from our routing records. They are not a trace of the physical network path or of how a provider processes a request inside its own systems.
What providers see: Each provider you route to receives your prompt and returns a completion, subject to their own privacy policies. Each provider has published its own position on training on customer content. AWS states in the Amazon Bedrock FAQ that neither AWS nor third-party model providers use inputs to or outputs from Amazon Bedrock to train models. Microsoft states in its data, privacy and security documentation for Foundry Models sold by Azure that prompts and completions are not used to train foundation models without your permission or instruction, and that the Microsoft Products and Services Data Protection Addendum governs the processing. Google states in the Training Restriction in the Service Specific Terms for its cloud services that it will not use customer data to train or fine-tune models without prior permission or instruction. These are the providers' own statements and terms, not promises by Northern Inference, and we cannot verify how a provider operates inside its systems. We record the document each statement rests on in each route's data-use posture, and the chat agent offers a model only if the record for its route shows a no-training commitment.
What providers keep. Retention differs by provider and model, and each route's position is on our data-use pages. Microsoft Azure may store prompts and completions that are flagged for human abuse review, and Microsoft's primary documentation does not publish a maximum retention period for that storage; this does not apply where your deployment is approved by Microsoft for modified or zero abuse monitoring. For some models on AWS Bedrock, AWS may keep inputs and outputs for abuse detection. For Anthropic Claude Fable 5 and Claude Fable 5.1, AWS states that all traffic is retained for up to 30 days for automated offline abuse detection and that flagged traffic may be reviewed by people at AWS; for some other models AWS states that it retains flagged traffic for up to 30 days. Google states that its Gemini models cache inputs and outputs in memory for up to 24 hours by default, and that it may log prompts to detect abuse.
Tier 4 routes use provider-default or non-Canadian residency; custody headers show the exact upstream provider, region, and jurisdiction for each request. Northern Inference's custody record reflects where we routed each request; once a request enters a provider, where inference runs and how the provider handles your data rests on that provider's own published documentation and commitments, which we cite per route in each model's data-use posture. We cannot independently execute inside a provider's infrastructure.
Prompt caching. To speed up repeated or long-context requests, Northern Inference and some providers use prompt caching, which temporarily stores a copy of the recurring portion of your prompt (the cached prefix) so it does not have to be re-processed on the next matching request. Any such cache lives in the same jurisdiction as the inference (a Canadian-resident route caches in Canada), is short-lived (typically a few minutes, set by the provider's cache time-to-live; Google states that its Gemini models keep cached data in memory for up to 24 hours by default), and is automatically purged when it expires. Prompt caching does not change where your data is processed and does not add long-term retention; it is a transient, in-jurisdiction performance optimization. You can see whether a request used the cache in its usage record (cache-read and cache-write token counts).
The following third parties process data on our behalf:
| Sub-processor | Purpose | Data Location |
|---|---|---|
| Amazon Web Services | Infrastructure, database, storage, email (SES), LLM inference (Bedrock). Tier 3 (Canada) and Tier 4 (non-Canadian routes). | Canada (ca-central-1) for our infrastructure, storage and Tier 3 inference; Tier 4 Bedrock routes may process data outside Canada. |
| Microsoft Azure | LLM inference (Azure). Tier 3 (Canada East) and Tier 4 (global). | Canada (Canada East) for Tier 3; data may be processed outside Canada for Tier 4. |
| Google Cloud | LLM inference (Vertex AI) and, where available, image generation in chat. Tier 4 only. | Not attested as Canada-only. Some routes use a Montreal regional endpoint, but Google has not committed that processing stays in Canada, and some use global endpoints. Treat as processed outside Canada. |
| Stripe | Payment processing | United States |
| Brave Software, Inc. | Web search in the chat agent, when you turn it on or allow it: receives the search queries the model writes, which can include words from your conversation | United States, and other countries where its service providers operate |
| Google Fonts | Font delivery (website only) | Global CDN |
Some routes are served through more than one organization: the provider we contract with, and in some cases a partner that hosts the model on that provider's platform. For each such route we publish the data-handling position of each organization in the chain as far as it is visible to us, cite the document each statement rests on, and retain a dated copy of the version we relied on. Where the chain extends beyond our visibility we say so and make no representation about that segment. These statements describe our practice and the documents we relied on at the stated date; they do not warrant the conduct of any third party, and each third party's own terms govern its processing. The route-by-route record is at northerninference.ca/data-use.
When enabled on your API key or per-request, our PII Substitution feature automatically detects personal information (names, emails, phone numbers, credit card numbers, IP addresses, locations) in your prompts and replaces them with realistic fictional substitutes before the request leaves our infrastructure. The originals are restored in the response you receive.
Content Debug Mode is an opt-in tool that lets Northern Inference support log the full content of your API prompts and responses to diagnose an issue (for example, a client configuration that makes a model misidentify itself). It is off by default and is never enabled silently.
While Content Debug Mode is on for your account, it also captures messages sent to the chat agent, because they are API requests made through your account. The chat agent shows a notice while it is on: it checks before each message, and again when each answer starts.
We do not use third-party tracking cookies. Authentication uses Secure, HttpOnly session cookies backed by server-side session records.
We use browser localStorage for:
Browsers can clear this storage on their own, for example after a week without a visit to the site.
Under the Personal Information Protection and Electronic Documents Act (PIPEDA), you have the right to:
Sealed chats and chat documents: we can't open them, so we can't include their content in a copy of your personal information. You can read your instructions, projects and memories in the chat, and you can export your saved chats to a file yourself. Ratings and the other records this policy describes are included in what we can give you.
To exercise any of these rights, email privacy at northerninference dot ca. We will respond within 30 days. For deletion requests, we will confirm deletion within 30 days of verifying your identity.
The Service is for adults. We do not knowingly collect personal information from anyone under 18, or under the age of majority where they live if that is higher. If we learn that an account belongs to someone under that age, we close it and delete its personal information, except what the law requires us to keep. To report such an account, write to us at the address in the Contact section.
For questions about this privacy policy or our data practices:
Email: privacy at northerninference dot ca
Northern Inference Inc., Ontario, Canada
We will post material changes to this policy on this page with an updated date and notify registered users by email at least 14 days before changes take effect.