I’ve been thinking a lot lately about what it means to share sensitive material with AI systems like ChatGPT. On the one hand, these tools are powerful collaborators — whether you’re working through health challenges, brainstorming a story, or hashing out a screenplay draft. On the other hand, the moment you type something in, you have to trust that it won’t end up somewhere you don’t want it.
For many of us, the stakes are different. Some of it touches on health information (where HIPAA comes to mind). Some of it touches on creative and professional work — intellectual property, future contracts, and the legal landmines of collaboration.
So let’s unpack what really happens when you share something with an AI today, and what an ideal model might look like in the future.
How It Works Today: Thin Client, Thick Server
Right now, services like ChatGPT are built on what’s essentially a thin client / thick server architecture.
-
Your browser or app is just the window.
-
OpenAI’s servers hold your history, memory, and the text of your sessions.
-
You can toggle “Chat History & Training” on (so chats are saved and may be used in training) or off (so they’re deleted after the session and not used in training).
This system is convenient: you get to scroll back through your old conversations in the sidebar, and the model can pick up threads across multiple days.
But it also means the persistence of your data lives in the cloud, not on your machine. If you want maximum privacy — for health notes or sensitive IP — you have to turn off history, which kills continuity.
The Ideal Future: Thick Client, Thin Server
What would be better?
A thick client / thin server setup.
-
The server would only do the heavy lifting: run the model, generate responses, and send them back.
-
The client (your computer) would store all your data locally — your history, your drafts, your “memory.”
-
Each time you opened a new session, your client would send only the relevant context to the server. When you closed it, nothing would persist beyond your machine.
This is the best of both worlds: continuity without ceding control. Your health reflections never leave your computer. Your creative drafts remain your intellectual property, with zero risk of leakage. The AI still works, but the model never “keeps” your work behind the scenes.
Where Things Stand for Creatives and Health Concerns
Until that architecture exists in mainstream AI products, here’s what we can do:
-
Health information: You can share personal reflections safely for your own use, but don’t put in identifiable patient data if you’re a clinician. The current platforms aren’t HIPAA-compliant.
-
Creative work: You keep full IP ownership of everything you write here, but if you don’t want drafts to linger on servers, flip off “Chat History & Training” and save the transcripts locally.
-
Hybrid use: Keep two modes — “history on” for casual, low-stakes work and “history off” for sensitive or professional projects.
For the ultra-private route, you can also run local LLMs (like GPT4All, LM Studio, or small LLaMA builds) directly on your computer. These don’t touch the cloud at all. Many professionals are starting to draft locally, then bring only what they’re comfortable sharing into ChatGPT for refinement.
Why This Matters
If AI is going to become a true collaborator — in health, in writing, in film, in law — then the foundation has to be trust.
I don’t want my medical reflections slipping into a training set. I don’t want my unpublished screenplay ideas floating around in some server. And I don’t want to give up continuity with my AI partner just because I choose privacy.
That’s why the next leap in AI design isn’t just about bigger models. It’s about architecture: giving users the ability to keep their data local, private, and theirs, while still tapping into the model’s power.
Because without that, we’re always weighing usefulness against risk. And for many of us — in healthcare, in the arts, in professional collaborations — that’s a trade-off we shouldn’t have to make.
Cheers! Sláinte! Na zdravie!
Compiled with aid of ChatGPT
No comments:
Post a Comment