Why GoSosh is private

Every app says it cares about privacy. We want to be specific about what that means here, because the differences between privacy stories matter — and they’re not visible from outside.

The short version

GoSosh is the only AI-powered conversation-tracking app we know of where everything the AI does happens on your iPhone. Not “we delete the data after,” not “we anonymize it,” not “we don’t sell it to third parties.” Your tracking preferences and reading patterns do not leave the device — with one specific exception that we describe in detail below.

We could not see what topics you watch, which stocks you follow, or which conversations you’re tuned into even if we wanted to. There is no server holding your reading history. There is no account. There is no email address we could match you to.

That’s a different category of privacy than what most apps claim, and it’s worth understanding why.

How most “AI-powered” tracking apps work

When you use Perplexity, ChatGPT, Google Discover, Apple News+ AI summaries, or any of the dozens of “AI” apps that have launched in 2024–2026, here is what’s happening behind the screen:

1. You tap an item or query.
2. The app sends what you tapped — or your question, or your context — to a server.
3. That server forwards it to OpenAI / Anthropic / Google / a similar cloud model provider.
4. The model generates a response. The provider logs the input for safety, debugging, and (in some terms of service) training future models.
5. The response comes back to your phone.

Steps 2 through 4 are where what you watch leaves your device. Most apps will tell you they don’t sell your data. Fewer will tell you that the cloud LLM provider sees every request. Almost none will tell you that, in many cases, “we delete it after 30 days” actually means the provider keeps it longer than the app does.

This isn’t a scandal — it’s just how cloud LLMs work. Models that big don’t run on phones yet, so apps that use them have to send your data out.

How GoSosh works instead

Apple released the Foundation Models framework in iOS 18.1. It’s a small language model that runs entirely on the device — fast enough for live use, and never sends data anywhere. We built every AI feature in GoSosh on top of it.

What that means concretely:

• No OpenAI, Anthropic, or Google API calls. We don’t have accounts with them. There’s no key in the app. No cloud LLM is ever invoked.
• No analytics on tracking behavior. We don’t ship which topics you watch, which stocks you follow, or which conversations engage you most.
• No accounts. You don’t sign up. You don’t give us an email. There is no profile.
• No personalization server. The ranking that surfaces what’s being talked about is computed on your device from sources you’ve chosen.
• No “anonymized telemetry.” A pattern of tracking interests is identifying even without your name attached to it. We don’t collect the pattern.

The trade is real: we don’t get to know what features users like, what topics trended, or whether anyone clicks our buttons. We accept that because the alternative is the architecture every other AI app uses, and that architecture doesn’t fit our brief.

“Doesn’t on-device AI suck?”

It’s a fair question. The headline numbers go to the giant cloud models — GPT-5, Claude, Gemini Ultra — and they do things smaller on-device models can’t.

What we found building GoSosh is that the specific jobs we need an AI to do — recognizing when two sources are talking about the same event, extracting which people and companies are being discussed, finding semantic similarity between items — fit comfortably within what Apple’s on-device model can handle. We’re not asking it to write essays. We’re asking it to cluster and classify.

For that work, on-device is good enough, and the privacy delta is huge.

One thing that does leave the device

There is one place where GoSosh does talk to a server: when an article in your feed doesn’t come with a cover image attached, we send the article’s URL to a small service we run ourselves (api.donbon.com — a tiny Cloudflare worker) that finds the cover image and sends it back. This is what gives the magazine-style cards in the Expanded view their hero images.

We want to be specific about what this service sees and what it doesn’t:

• It sees the URL of the article being looked up. That’s it. No account identifier (there are no accounts). No device fingerprint. No correlation to your other tracking activity. No “user N opened article X” log line. The service is not architected to know who you are; it cannot know.
• It caches the lookup result globally. Once any GoSosh user, ever, has looked up a given article URL, every subsequent lookup of that URL returns instantly from the cache without any scraping. That cache is the only persistent state the service has.
• We built it ourselves. The alternative paths were a cloud LLM (sends URLs to OpenAI / Anthropic), a third-party scraping service (sends URLs to a company optimizing for their incentives, not ours), or a “free” image-preview API (same problem). We built our own service so that the URL lookup is governed by the same person who built the app and writes its privacy page.
• You can turn it off. Open Settings → Image lookup and flip the toggle off. You’ll see fewer hero images on cards from publishers whose RSS feeds don’t include them — but no off-device requests will be made.

This is the smallest compromise we could find that still produced a feed worth reading. We are spelling it out here rather than burying it in a legal page because it is a real architectural choice and we would rather you make it knowingly.

What we do collect

Honesty section. The app does ship some data — but the list is short and conventional:

• App Store anonymous purchase records so your Plus purchase works. This is Apple’s infrastructure, not ours, and it doesn’t carry tracking data.
• Crash logs, if you opt in to send them, via Apple’s standard crash reporting. These contain no content from inside the app.
• Source requests — when you select a topic, GoSosh fetches RSS feeds and APIs from that topic’s sources. Those sources see a generic request from your phone (same as if you opened their site in Safari). They don’t see who you are or what else you’re tracking.
• Article URLs for image lookup — when a feed item arrives without a cover image, the article URL is sent to our own service (api.donbon.com) for a hero-image lookup. The service sees the URL, nothing else. Full architecture is in the “One thing that does leave the device” section above. You can turn this off in Settings → Image lookup.

That’s it.

The principle

There is a class of apps that asks you to trust their data handling. There is a smaller class of apps that tries to make trust unnecessary by not having your data in the first place.

GoSosh is in the second class. That’s not a marketing position. It’s a design choice that runs through how the app was built, why it’s iOS-only, and why we use on-device AI even when cloud AI would be flashier.

If you’ve been waiting for an app that tracks what you care about without tracking you, this is it.