Is Your AI Really Local?
Is Your AI Really Local?
You download an AI app. It runs on your computer. Its website mentions open source, local models, and privacy.
So your files stay on your computer. Right?
Not necessarily.
These terms describe different things. Where an application runs, where its AI processes your documents, and what happens to your conversations are separate questions.
Customers should not need to investigate source code or network connections to understand the difference. But sometimes, that is exactly what it takes.
Does open source mean an AI app is local and private?
No. Open source lets people inspect, modify, and redistribute software under its license. It does not guarantee that the software processes your data locally or avoids sending it to cloud services.
Transparency is valuable. Being able to inspect an application is better than having no visibility into how it works.
But an open-source application can still send your documents to a remote AI model. And a downloadable model can run on someone else’s server rather than your laptop.
The question is not just whether the code is available.
Where does your data actually go?
Does supporting local models mean everything runs locally?
No. Supporting local models means an application offers that capability. It does not establish that local processing is the default or that every background feature uses your chosen local model.
Even “runs on your computer” can describe the application rather than the AI doing the work. A desktop interface can send your questions and file contents to a cloud provider.
OpenCode: a local model, but ?
In January 2026, an OpenCode user reported an unexpected discovery. They had configured a locally hosted model, but session titles were being generated outside their network. Blocking outgoing connections stopped that feature from working.
A project collaborator confirmed that title generation could fall back to a cloud model. In October, a maintainer said that default behavior had been removed in March. Read the GitHub discussion.
The lesson remains important: selecting a local model did not automatically make the entire workflow local, your conversation can be sent to AI cloud provider to generate a title.
For someone handling confidential documents, even a small background feature matters. Their privacy expectations do not stop at the main chat window.
Perplexity: whose “Computer” is it?
The word “computer” suggests something familiar. Your machine. Your files. Something you control.
But Perplexity’s own documentation explains that Perplexity Computer runs tasks in the cloud, while Personal Computer adds access to your local files, applications, and desktop. Access to your computer is not the same as processing everything on it. Perplexity’s explanation.
Perplexity also now offers Portable Computer, which runs models locally and asks before escalating tasks to cloud models. That is an important distinction. Portable Computer.
Our criticism is the ambiguity: customers should not have to decode a family of product names to understand where their information is processed.
Do “no training” and “no retention” mean your data never leaves your computer?
No. “No training” concerns how a provider uses your data. “No retention” concerns what it stores and for how long, subject to the policy’s scope and exceptions. Neither means your data stays on your computer.
A cloud service can process your document without using it to train a model. It can also receive and process a conversation before deleting it.
Whether people can review content, whether safety systems analyze it, and whether derived records are retained are additional questions. A “no training” statement does not answer them.
For example, Anthropic’s zero-data-retention documentation says that it still retains safety-classifier results and describes exceptions involving legal obligations, misuse, or harm. Anthropic’s policy.
These protections can be useful. But limiting what happens after your data reaches a provider is not the same as keeping it on your device.
Can Fenn work without cloud AI?
Yes. Fenn uses local AI by default. After setup, with the required models and files available on your Mac, local search and AI processing works 100% offline. Using an external cloud model requires an explicit choice from you.
That word matters: explicit.
You should know when you are choosing to send questions and relevant file context to another provider. Its data policies then apply. Internet access is also needed for downloads, license checks, updates, and optional online services. How Fenn handles local and cloud AI.
You can test the local workflow yourself: download the models, disconnect your Mac from the internet, and keep working.
Supporting local AI is an option. Keeping your data local is a product decision.