Private-by-default AI features
Why it matters If you build websites or apps, someone will soon ask you to "add AI". The choices you make in the first week decide whether your users' private text quietly flows to a third party forever, or stays with them. Private-by-default is mostly about choosing well, not about clever code.
The packed lunch
Packing a lunch from home is cheaper and you know exactly what is in it. Buying lunch is sometimes worth it. A good product team picks deliberately each day, instead of always buying because that is what everyone does. AI features are the same: local when you can, cloud when it is really worth it, and never by accident.
Private by default: the setting a user gets without touching anything is the private one. Sharing more is something they choose, never something they have to switch off.
A ladder of choices
For each AI feature, start at the top and only step down when the step above cannot do the job.
- No AI. Could a simple rule, a template or a good form do it? A rule never leaks, never invents facts and works offline.
- On-device model. Small jobs like rewording, sorting or short summaries often fit a small local model. Nothing leaves the device.
- Cloud model, minimised. Send only what the task needs: redact first, send one paragraph instead of the whole document, drop old chat history.
- Cloud model, full context. Only when the task truly needs it, and only with the user's clear agreement each time.
Floati's CV Builder sits on the top two rungs: a rule-based helper that always works, plus an optional on-device polish. The polish runs in your browser. The rule-based helper runs on Floati's own server, so your CV text does travel there, but it is never sent to an AI company. Being able to say exactly that, in one sentence, is part of private by default.
Questions to ask before shipping
| Question | Private-by-default answer |
|---|---|
| Is the feature on or off by default? | Off, or local-only, until the user turns it on |
| Can the user see what will be sent? | Yes: a preview of the exact text before it goes |
| Does the provider keep prompts or train on them? | Pick settings or a plan where it does not, and say so plainly |
| Do our logs store prompts? | No. Log that a request happened, not what it said |
| Can the AI call tools or read other data? | Only what this feature needs, nothing more |
| What happens if the AI part fails? | A non-AI fallback still works |
Your own logs are part of it
Teams often pick a careful AI provider and then save every prompt in their own error logs "for debugging". That undoes the whole effort. Log counts, durations and error codes; if you must keep text to fix a bug, redact it and delete it on a short schedule.
A common misconception
"Asking for consent fixes everything." A pop-up that says "AI may process your data, OK?" does not protect anyone if the button is the only way forward. Consent matters, but sending less in the first place matters more. The best consent screen is one you do not need because nothing personal leaves the device.
Try it Pick one AI feature in an app you use (a smart reply, a writing helper, a photo search). Walk it down the ladder: which rung is it on? Could it have been one rung higher? Write one sentence about what you would change.
Recap
- Private by default means the setting users get without acting is the private one.
- Climb down the ladder only as far as you must: no AI, then on-device, then minimised cloud, then full cloud.
- Show users what will be sent, keep prompts out of your own logs, and always keep a non-AI fallback.
- Sending less beats asking for consent to send more.
This lesson is locked
Lessons open one at a time. Finish the previous lesson to unlock this one.