Privacy when the AI stays in the house
This page may contain Amazon affiliate links. A purchase after a click can earn the site a commission, at no extra cost to you. The advice stays independent. How affiliation works.
The GDPR (General Data Protection Regulation, European regulation 2016/679) sets rules for how personal data is collected, used, and kept. Personal data is anything that can identify a person, directly or indirectly. Ollama runs large language models on your machine. The prompts and files you give it stay, in principle, on a local disk, with no required call to OpenAI or Google. This guide explains what that changes for a household in October 2026. A small business and a hospital are outside it.
This guide is not legal advice. It is practical household guidance. It explains common questions (photos, voice, children, health) so you can ask better ones. For a professional activity, large-scale processing, or a doubt about a legal basis, talk to a lawyer or a DPO (data protection officer, the person designated to watch over data protection).
Five useful moves this week
Before the longer sections, a concrete plan for a household starting with Ollama:
- Try it before you buy. A first local AI in 15 minutes on the PC you already have.
- Check what the network can reach. Ollama on
127.0.0.1, or on a closed home network. Leave port 11434 unpublished on the internet (see Open WebUI for beginners). - Set sensitive folders aside. A working volume outside iCloud, OneDrive, and Google Photos for the PDFs and photos you analyze.
- Encrypt the disk of the PC or mini PC that stores health, identity, or school documents (BitLocker, FileVault, or LUKS, depending on the OS).
- Write down who can open the family chat. Separate accounts, or Open WebUI logins. Read the private family assistant if several people use the living-room machine.
What actually changes between local and cloud
With ChatGPT or another cloud assistant, your text passes through the provider’s servers. The contract, the subprocessors, the country of processing, and how long they keep data all come into play. With a local stack (install Ollama, Open WebUI, Whisper), the work happens on your CPU or GPU. You take on the physical and technical security of the box, in place of negotiating the online service’s clauses.
| Cloud (online software, SaaS, software as a service) | Local Ollama | |
|---|---|---|
| Prompts crossing the internet | Yes (TLS, Transport Layer Security: an encrypted connection between your device and the provider’s server) | No, once the model is downloaded and you stay offline |
| US or EU subprocessors | The vendor’s list | You, plus the publishers of the OS and the software |
| Access and deletion | You ask the provider | You delete the files and the logs |
| A leak from a bad setup | A hacked account on the vendor’s side | An exposed port, an open NAS, a cloud backup |
Local is not a law-free zone. If you process data about other people (neighbors, informal clients, an association), the duties can move closer to those of a controller, the person who decides why and how personal data is used. LocAIhome aims at household use. Past that, this page is the wrong frame.
Money, next door: cost over 3 years · kilowatt-hour budget.
WAN and LAN
Two abbreviations prevent a lot of false confidence:
- LAN (local area network): the network in the house. The router, Wi-Fi, and Ethernet between the mini PC, the phone, and a NAS. While Ollama listens only on
127.0.0.1or a home-network address, and guests are kept off that same segment, the requests stay physically in the house. - WAN (wide area network): the public internet. As soon as you forward Ollama’s port, use a tunnel such as a badly set Tailscale, or sync a folder to the cloud, the data can leave the home network.
A picture to keep in mind:
Phone ──Wi-Fi──► Router ──home network──► Mini PC running Ollama ✅ ordinary household use
Mini PC running Ollama ──port forward──► Internet (WAN) ⚠️ a way in from outside
Prompt folder ──sync──► Google Drive / iCloud ⚠️ a second copy in the cloud
A Docker install of Ollama and Open WebUI on a headless server should stay on the home network or on localhost for a household. A public server is a different scenario.
Outside a beginner setup: publishing on the internet. Listening on every interface (
0.0.0.0, port 11434) without a reverse proxy (an intermediate server that ends TLS and filters access), a certificate, and strong authentication exposes the Ollama API on the WAN. Skip that as a way to “test from outside.” Prefer a home VPN (WireGuard) or the security documentation for your router. References: the Ollama FAQ and the documentation of whatever proxy you choose.
Risks that remain when everything is “100% local”
Even if your PDFs never go to Microsoft, several leaks are still plausible:
The network is open. A NAT rule (the router forwards a door on the internet to a machine on the home network), a malicious plugin, or a Home Assistant left too open can make the Ollama API reachable from the WAN. Bots scan ports. A language model with no login becomes a free proxy for strangers.
A device is stolen or sent for repair. A laptop with disk encryption turned off (BitLocker, FileVault, LUKS) holds chat history, embeddings, and Whisper’s temporary files.
Backups and the operating system. OneDrive, iCloud, Google Photos, or a synchronized NAS copy what you believed was local only. Screenshots and a cloud clipboard follow the same path.
Other people’s data. A family assistant that remembers the children’s schedule or health contacts processes personal data about other people. Stay open with them about that, even at home.
The model’s quality. A local language model can hallucinate a diagnosis or invent a promise. That is not a GDPR leak. It is a risk for the person concerned if you act on the generated text.
Logs and metadata. Ollama, proxies, and n8n sometimes log IP addresses, timestamps, and excerpts of prompts. Those files are data to protect as well.
Updates and the supply chain. Downloading a model from a public registry is an act of trust. Prefer known sources, and check checksums when you can.
Concrete measures for a household
Keep the posture in proportion. You are not a bank. School photos still deserve the same care you would give them on a closed social network.
Network and exposure
- Ollama and the web interfaces on localhost, or on a VLAN (a separate logical network, for example guest Wi-Fi split from the main home network).
- No port forward to Ollama’s default port (11434), or to Open WebUI, unless you already have security experience for that job.
- The operating-system firewall stays on. Automatic updates stay on for a dedicated mini server.
- If you need remote access: a personal VPN (WireGuard) rather than a raw exposure. That is past a beginner setup. See your router’s documentation.
Data and how long it lives
- A working folder outside cloud sync, with explicit exclusions in iCloud and Drive.
- Disk encryption on the PC that stores health, identity, and children’s documents.
- A retention habit: purge logs and sensitive conversations after you use them (a dedicated SSD versus the system disk).
- Leave bank details and passwords out of the chat, even a local one. Malware exists off the cloud as well.
How the family runs it
- Tell teenagers that a model in the living room can read what they send if the account is shared.
- Separate OS accounts, or an Open WebUI instance with a login per person.
- For Whisper on class recordings or meetings: tell the participants when the law requires it (work or school context).
Hardware choices that line up with privacy
A dedicated AI mini PC limits the mix with a work computer. See choosing hardware, VRAM, and the mini PC comparison. Enough RAM avoids swapping fragments of prompts onto an unencrypted disk.
A sensible order for Amazon searches (partner links, October 2026; check the current price; see the affiliate disclosure):
- A quiet dedicated mini PC, so the AI server stays off the work PC.
- A self-encrypting SSD, a volume for sensitive documents.
- A router with guest Wi-Fi, so guests and the main network stay apart.
- A USB flash drive for offline backup, spare copies that skip automatic cloud sync.
Reminder: LocAIhome speaks to households, not to small businesses. The moment you sell a service that leans on Ollama (coaching, writing for clients), you enter a contractual frame and a record-keeping frame. This site does not cover that case.
Common jobs at home
Paperwork and identity documents
Analyzing a bill or a lease locally avoids uploading a scan to a public chatbot. Store the PDFs on an encrypted volume. Afterward, delete the temporary copies some interfaces create in /tmp or %TEMP%.
Family photos and video
Images without a cloud and ComfyUI can train on or retouch faces. Even with no upload, the people in the photos still have a say in how their image is used. For children, the strictest care applies to sharing, and storage needs that same care.
Voice and transcription
Offline Whisper transcribes meetings or voice memos. The audio file is personal data. Cut the segments you do not need before you process it, and avoid leaving recordings on a NAS the whole family can open with no control.
Code and a work project from home
Continue with a local model limits the leak of work secrets toward the cloud, if the git repository and the API keys are not synced somewhere else. Keep production keys and the development environment apart.
GDPR and vendors: what still leaves the house
Even at “100% local,” you still deal with:
- the maker of the operating system (Windows, macOS, or Linux telemetry, which is worth a look);
- the Ollama model registry when you
pull; - possible CDNs for Docker updates.
Those are not processing of your family prompts in the strict sense. They are flows worth knowing. Turning off OS telemetry, when that is acceptable to you, reduces the noise.
A mixed cloud strategy
Many households mix things: cloud email, iCloud photos, and a local language model. GDPR is judged per purpose. Local protects the “sensitive writing” piece. It does not automatically downgrade the rest of what you still keep elsewhere. A map of where the copies live (phone, NAS, USB stick) is worth more than a “100% local” slogan.
A safer way to start
- A first local AI in 15 minutes on the current PC, so you can try it before a purchase.
- Choose a first model that fits the RAM you have.
- Install Ollama on Windows, or headless Linux on a dedicated machine.
- Move to Open WebUI only after the network listening is locked down.
- Cross-check the hardware selector if you isolate the work on a mini PC.
For noise and power of that dedicated box: a mini PC left on and the Mac mini M4.
Keep going
- Cost over 3 yearsA cloud subscription beside the up-front hardware purchase.
- Electricity budgetKilowatt-hours of a mini server left on for privacy.
- Family assistantAccounts and rules on the home network.
- Choosing hardwareIsolate a box that does the local processing.
In short: local AI brings the processing into the living room and moves the big cloud companies away from the content of the prompts. GDPR and ordinary care still ask you to control the network, the backups, the people involved, and the devices. Treat Ollama like a digital safe. The safe moved. Closing the door is still your job.
Common questions
Does a fully local AI take me outside GDPR?
Not automatically. GDPR covers the processing of personal data as soon as there is collection, storage, or structured use, including for a private person in some cases (neighbors, an association, professional activity). Strictly domestic use for yourself is different from a service you provide to clients, or from processing someone else’s data without a legal basis. Ask a lawyer about a borderline case.
Can I have Ollama look at photos of children, at home?
Locally, you avoid sending them to a US or Asian cloud, which shrinks the leak on the online provider’s side. Household risks remain: a shared Windows account, the phone’s cloud backup, an Ollama port left open. Keep copies few, encrypt sensitive disks, and leave the interface off the public internet.
Do I have to keep a record of processing, the way a company does?
The formal duties (a record, a data protection officer, impact assessments) mainly target organized controllers and higher-risk activities. For a household trying Ollama on its own files, everyday habits are often enough. Writing things down stays useful if you host data belonging to relatives or a small group.
Is Open WebUI or Ollama on the home network risky?
On a home network that is not exposed to the internet, the risk is mainly guest devices, or an accidental port forward toward Ollama (port 11434). Bind listening to localhost, or keep guest Wi-Fi on its own segment, require a login on the interface, and check the router’s rules. See the WAN and LAN section above.
Do iCloud, Google Photos, or OneDrive undo “local”?
Yes, for the copy that syncs. A photo you analyzed locally can still leave for an operating-system cloud if the folder is synchronized. Turn sync off on the Ollama working folders, or use a dedicated volume that is not backed up online, if you want fewer third-party copies.
Is it sensible to run health notes through a local model?
Local shrinks the surface on the hosting company’s side, and health data stays sensitive: encrypt the disk, use a session password, keep the model unshared with other people, and treat generated summaries with care because they can be wrong. Do not use them as a substitute for medical advice. Professional health use sits in a different legal frame.
Sources
Narrow it down with the hardware selector · French version of this section