“The AI runs on my laptop” sounds reassuring. So does “the meeting is private,” until somebody starts a livestream. For AI, the useful question is specific: which parts of this workflow stay on the computer, and which parts can contact another service?

Running a model locally can keep that model's inference on your device. It does not, by itself, describe the behaviour of the chat app, document importer, search tool, backup service or connector around it. Before trusting a privacy label, map the whole job.

Separate the model, the application and its tools

Consider a hypothetical task: summarising a folder of meeting notes. There are at least four questions. Where does the model generate its answer? Where are the notes processed? Where is the conversation saved? Can a tool send part of the request somewhere else?

Write the answers in plain language. “Model on laptop; documents on laptop; chat history in this folder; web search disabled” is more informative than “private AI.” If you cannot fill in a line, treat it as an unanswered question rather than an implied privacy guarantee.

This follows the distinction behind our older article on AI and public versus private cloud infrastructure. Location is part of the design; the route taken by the actual data is what you need to inspect.

What two local-AI tools say about offline use

Ollama says it does not see prompts or data when models run locally. Its cloud-hosted models instead process prompts and responses to provide the service. Those are different execution paths inside the same product family. Its FAQ documents a local-only configuration, using disable_ollama_cloud or OLLAMA_NO_CLOUD=1, followed by a restart; disabling cloud features also removes its cloud models and web search. These are the vendor's documented behaviours, not results of an independent audit here. Read the Ollama FAQ.

LM Studio says its downloaded models, document chat and local server can operate offline. Its documentation separately identifies operations needing connectivity, including model search and downloads, runtime downloads and update checks. It says document processing stays local. Again, these are statements about the described application features, not a guarantee covering every other program or integration on your computer. LM Studio's offline-operation documentation explains the boundary.

These tools illustrate what to ask; they are not a claim that one is the best choice. Recheck the current documentation for the version and features you use, especially after adding a cloud model, search function or third-party interface.

Make a small data-route table before using real documents

An original worksheet for inspecting your own setup

Part of the task	Question to resolve
Chat model	Is this exact selected model running locally or through a hosted endpoint?
Document import and search	Where are text extraction, indexing and retrieval performed?
Optional tools	Can web search, a connector or an action transmit text outside the app?
Saved history and files	Where are copies stored, and are those folders backed up or synchronised elsewhere?
Model server	Can only this computer reach it, or is it exposed to other devices?

This is an inspection worksheet, not a product-specific finding. Read the app's settings, documentation and the rules for your documents. For workplace material, an acceptable local tool still needs to fit the organisation's approved handling process. Replacing a cloud request with a local request does not answer who is allowed to access the files.

Ollama documents a default binding to 127.0.0.1:11434 and a setting that changes it. That is a useful example of a boundary you should inspect before enabling network access. Do not assume “installed locally” means “reachable only by me.” The network-configuration section of its FAQ describes the defaults.

Try an offline rehearsal with harmless material

Use a made-up document containing a clearly fictional project, such as a plan to deliver 12 umbrellas to a lunar picnic. Download the model through the normal documented process first. Then select the local model, disable optional online features as documented, disconnect networking and try the summary.

If the task fails offline, investigate which dependency it needs. If it succeeds, you have evidence that this particular operation can complete without an active network connection. You have not proved that the application never communicates when connected, that no local copies remain, or that another enabled feature behaves the same way.

Reconnect, review the documented settings and decide whether the result meets your needs. This rehearsal is a proposed check, not a test performed for this article. Use synthetic material until you understand the workflow. “It produced a paragraph” is a functionality observation, not a privacy certification.

Does local AI require an expensive computer?

There is no single hardware answer for every model and task. Hugging Face explains that quantization stores model weights at lower precision to reduce memory requirements while trying to preserve accuracy. Supported methods, hardware and numerical precision differ. Its quantization overview is a useful technical starting point.

Here is simplified arithmetic, not a buying recommendation: eight billion parameters stored at four bits each represent about four billion bytes of raw weight values. That is roughly 4 GB in decimal units. It is not a claim that the complete model fits or runs well on a 4 GB device. Actual formats, metadata, working memory, context and the rest of the application add requirements.

Before buying hardware, choose a representative harmless task and check the exact model's requirements and your runtime's supported hardware. Judge whether the result is accurate enough and whether the wait is acceptable. A model that fits but repeatedly misses the important sentence is not a successful document assistant.

Privacy and accuracy need separate checks

A response can remain local and still be wrong. Use the same claim-checking process we describe for AI answers: inspect the source document, verify the consequential claims and distinguish a quotation from the model's interpretation.

For the meeting-note example, ask for the source passage behind each deadline or commitment, then compare it with the original. Do not let a fluent summary silently turn “we might launch in June” into “launch approved for June.” That hypothetical mistake is about interpretation, and moving the computation onto your laptop would not make it acceptable.

Use local AI when its documented data route, practical performance and handling requirements fit the job. Start with one workflow you can explain from input to saved output. Privacy becomes easier to assess when it is a description of actual steps, rather than a sticker on the download button.

Preparation note — 19 October 2026: AI assisted research organisation and writing. Current primary product documentation was checked; vendor claims are identified as such. No software installation, network audit, performance benchmark or hands-on offline test was performed. The rehearsal, document and memory calculation are illustrative.