AI privacy & Chat Control: your prompts are messages

here's the connection most people miss: AI chat is messaging.

when you send a prompt to ChatGPT, Claude, Gemini, or any AI tool — that prompt is a message. it's transmitted to a platform, processed, and stored (usually). under Chat Control's definition of "interpersonal communication service," AI chat platforms could be required to scan your conversations.

think about what that means. every prompt you've ever written to an AI reveals:

  • your thoughts, fears, and anxieties
  • your business plans and strategies
  • your creative ideas
  • your health concerns
  • your political views
  • your legal questions

it's more intimate than your search history because you're having a conversation. you're asking follow-up questions. you're explaining context. you're being vulnerable in ways you wouldn't be with a search engine.

and under Chat Control, all of it could be scanned by an AI with a 94% false positive rate.

how AI platforms handle your data

let's be specific about what each major AI platform does with your conversations:

ChatGPT (OpenAI)

  • stores conversations by default (you can turn this off, but most people don't)
  • trains on your data unless you opt out (settings → data controls → "improve the model for everyone")
  • subject to US law — subpoenas, national security letters, FISA
  • complies with government requests — OpenAI has a law enforcement guide
  • under Chat Control: if operating in the EU, could be required to scan prompts

Claude (Anthropic)

  • stores conversations by default
  • claims not to train on user data (as of their public statements)
  • subject to US law
  • under Chat Control: same obligations as any platform

Gemini (Google)

  • stores conversations in your Google account
  • trains on your data by default
  • connected to your entire Google profile — search history, email, location, etc.
  • under Chat Control: Google will absolutely comply. they've been scanning Gmail for years.

the problem

all three are US-based. all three store your conversations. all three could be compelled to scan under Chat Control. and all three have the technical ability to implement client-side scanning if required.

the private alternatives

NanoGPT — my daily driver

what it is: a multi-model AI subscription that gives you access to 50+ AI models (GPT-4o, Claude, Gemini, Mistral, etc.) for $8/month. no conversation storage, no training on your data.

why it works against Chat Control:

  • no conversation storage — your prompts are processed and deleted. there's nothing to scan.
  • crypto payments — pay with Bitcoin, Monero, or other crypto for anonymous access
  • no tracking — no ads, no analytics, no behavioral profiling
  • API-first — you can use it through any compatible client, not just their web interface

what i use it for: daily AI tasks. coding help, writing assistance, research. everything i used to use ChatGPT for, but without the surveillance.

setup:

  1. go to nanogpt.ai
  2. create an account (or don't — you can use the API without one)
  3. pay with crypto for maximum privacy
  4. use the web interface or connect via API

the trade-off: NanoGPT is a proxy — your prompts still go to the underlying model providers (OpenAI, Anthropic, etc.). but NanoGPT doesn't store them on their end, and the model providers process them as API requests (which have different data retention policies than consumer products).

affiliate link: go/nanogpt

local models — the nuclear option

what it is: run AI models on your own hardware. nothing leaves your machine.

tools:

  • Ollama — the easiest way to run local models. install, pull a model, chat. works on Mac, Windows, Linux.
  • LM Studio — GUI for downloading and running local models. good for non-technical users.
  • llama.cpp — run Llama models on consumer hardware. more technical but more flexible.

why it works against Chat Control: there's no platform to compel. your prompts never leave your device. there's no server to scan. it's the same logic as self-hosting a Matrix server — if there's no third party, there's nothing to regulate.

trade-offs:

  • requires decent hardware (16GB+ RAM for good models, GPU helps)
  • local models aren't as powerful as GPT-4 or Claude (but they're getting close)
  • setup is more technical than using a web service
  • no real-time internet access

what i use it for: sensitive questions where i don't want any server to see my prompt. also for experimenting with models without paying per query.

setup:

  1. install Ollama from ollama.ai
  2. run ollama pull llama3 (or whatever model you want)
  3. run ollama run llama3 to start chatting
  4. or connect Ollama to Open WebUI for a ChatGPT-like interface

i wrote more about local models in my local AI models guide.

Venice AI — the anonymous option

what it is: anonymous AI chat. no account needed, no conversation storage.

why it works against Chat Control:

  • no account required — no identity to tie conversations to
  • no conversation storage
  • no logging
  • crypto payments for premium features

trade-offs:

  • fewer models than NanoGPT
  • less reliable (smaller infrastructure)
  • no API access (as of june 2026)

what i use it for: quick questions where i don't want to log into anything.

the metadata problem

even if you use a private AI tool, your ISP can see that you're connecting to it. that's metadata.

the layers:

  1. content privacy — the AI tool doesn't store your prompts (NanoGPT, local models)
  2. transport privacy — your ISP can't see which AI tool you're using (VPN, encrypted DNS)
  3. identity privacy — the AI tool doesn't know who you are (crypto payment, no account)

to be fully private:

  • use NanoGPT or local models (content privacy)
  • use a VPN (transport privacy)
  • pay with Monero (identity privacy)
  • use encrypted DNS (DNS privacy)

the Chat Control scenario

let me paint the picture of what happens if Chat Control is applied to AI:

step 1: AI platforms are classified as "interpersonal communication services" step 2: platforms must implement detection systems step 3: your prompts are scanned for "suspicious" content step 4: flagged prompts are reported to national authorities step 5: you get a knock on the door because you asked an AI about security vulnerabilities for your pentesting job

this isn't hypothetical. it's the logical extension of Chat Control. and the AI scanning industry is already building the tools for it.

my setup

for full transparency, here's how i handle AI privacy:

use casetoolwhy
daily AI tasksNanoGPTno storage, crypto payment, 50+ models
sensitive questionslocal models (Ollama)nothing leaves my machine
quick anonymous questionsVenice AIno account needed
coding assistanceNanoGPT API via VS Codeintegrated, private, no conversation storage

i pay for NanoGPT with Monero. i use a VPN when accessing any AI tool. my DNS is encrypted. it's not perfect, but it's enough to make mass surveillance impractical.

last updated: july 2026