Why we make local the default for private AI

We make local the default because most people will never find the right privacy toggle. The third of the five beliefs on our About page (opens in a new tab) says it with a joke: "Teaching everyone data privacy is a lost cause (sort of)."

The "sort of" matters, because people deserve to know where their words go. Still, a privacy plan that depends on everyone reading the fine print and remembering a setting every time will fail most of them. So the belief proposes a better default: "local first, with heavier work offloaded to more capable models only when the owner approves it."

People know the risk and paste anyway

That belief cites Cisco's 2024 Consumer Privacy Survey, released October 30, 2024, which asked more than 2,600 consumers in 12 countries about privacy and AI. Among generative-AI users, 84% said they were concerned that what they enter could go public. Even so, 30% of all the generative-AI users surveyed said they enter personal or confidential information.

Our About page lists what that looks like in daily life: "Teachers throw student data into personal chatbot accounts. Doctors type in patient symptoms. Tired people upload their tax documents."

Four cases from 2025 and 2026 where users' settings had little say

Our privacy page (opens in a new tab) keeps four receipts, each with a named source.

Deleted chats, kept (May 2025)

In May 2025, in the New York Times copyright case, a federal judge ordered OpenAI to preserve ChatGPT output logs it would otherwise delete, including chats users had deleted. The blanket hold ended that September, and what was captured, except logs from the EEA, Switzerland and the UK, stays held for the litigation.

20 million conversations as evidence (January 2026)

In January 2026, in the same case, a federal judge affirmed an order for OpenAI to produce 20 million de-identified consumer ChatGPT conversations, full prompts and full answers, to the publishers suing it. The people who wrote those conversations are not parties to the lawsuit.

Training, on by default (as of October 2026)

On consumer ChatGPT accounts, the "Improve the model for everyone" setting starts switched on, so conversations may be used to train OpenAI's models. OpenAI's Data Controls FAQ (opens in a new tab) explains how to turn it off under Settings, then Data controls. Turning it off applies only to conversations from that point on.

A million log lines in the open (January 2025)

In January 2025, Wiz Research found a DeepSeek database reachable from the internet with no authentication. It held over a million log entries, including chat history and API secrets. No user setting reaches a misconfigured server.

In all four cases, once a conversation sat on someone else's server, a court order, a default setting or a mistake decided what happened to it next.

What a local default changes

Our privacy page puts the alternative in five words: "Nothing is logged by us." We log nothing because there is no account, no telemetry and no cloud to log to.

Base Drive, our first product, works that way. Its chat app talks to the model only on localhost, on your own laptop, and keeps working after you turn Wi-Fi off. So the court orders, the training default and the open database above have nothing of ours to act on. As of October 2026, Base Drive is coming soon at a planned $19.99 and cannot be ordered yet. How Base Drive earns the trust a USB drive needs covers how you can check those claims yourself.

A local default also moves responsibility to you. Our terms (opens in a new tab) say we "cannot protect data on a machine that is already compromised, shared, unencrypted, or physically accessible to someone else." Your laptop's password and disk encryption become part of your AI privacy.

How outside help would work

A small local model will sometimes fall short on the hardest questions. Our fifth belief calls bigger models "tools to call upon," and our privacy and how-it-works pages describe a three-part plan for those moments:

• A consent prompt before any outside help, naming exactly what would be sent
• A written record of every approval afterward
• One switch that turns outside help off for good

As of October 2026, the consent prompt, the approval record and the off switch are plans. In our site's words: "Outside help is not in the shipping build." As it stands, Base Drive answers entirely on your laptop, with no network involved at all, and a local-only mode is planned to keep it that way for anyone who wants it.

Outside help would start switched off and stay off until you turn it on. It would go only to United States-hosted providers under zero-data-retention terms, and if no such provider is available, the request would stop. What zero data retention means, and where it stops covers those provider rules and their limits.

We still think a consent prompt beats a settings toggle. A toggle asks you to protect yourself once, in advance, for every future chat. A consent prompt asks about one request, when it happens, and shows you what would be sent before you say yes.

Where our promise gets conditional

The same privacy page has a section called "The honest part." It says: "The moment you allow outside help or remote access, 'data never leaves your machine' becomes conditional."

That condition applies to everything above. Once outside help exists and you approve a request, that request would leave your machine under the provider's terms. Remote access is a planned, optional Base Router feature. It would be encrypted end to end, but setting up the link may need some outside coordination. So it would start switched off, and you would have to turn it on. Base Router itself is in development at a planned $400 and is not available to order. Base Router, the always-on AI box we are designing has the details.

Our own demos sit outside the promise too. Our terms say anything you paste into the demos on our site leaves your machine by definition. /try (opens in a new tab) says "These chats aren't stored," and /open (opens in a new tab) sends model calls only to zero-data-retention providers, though its web search and PDF parsing fall outside that guarantee. Both still run on computers that we or our providers operate.

As designed, data leaves your own hardware only when you choose it: a request you approve, remote access you switch on, or a demo you type into on our site.

Our fourth belief is about friends of ours who want nothing to do with AI. Building AI for people who are wary of AI covers how we build for them.