nobody's talking about this angle of chat control, and they should be.
chat control's scanning mandates don't just apply to traditional messaging. they apply to any "interpersonal communication" service. and increasingly, that includes AI chat interfaces.
think about what you've typed into chatgpt, claude, or any AI assistant. medical questions. legal questions. business strategy. personal dilemmas. creative work you haven't published yet.
all of that is potentially in scope.
the definition problem
chat control's regulation defines its scope around "interpersonal communications." the argument has been that AI chatbots aren't interpersonal because you're talking to a machine, not a person.
but the technical working documents show a different picture. several member states have pushed to include AI interactions in the scanning scope, arguing that:
- AI conversations often contain more sensitive personal information than person-to-person messages
- the same infrastructure that scans messages can scan AI prompts
- criminals could use AI to generate or process illegal content
the legal basis is shaky, but when has that stopped surveillance expansion?
what AI companies already collect
before we even get to chat control, here's what AI companies already know about you:
openai (chatgpt): stores all conversations by default, uses them for training unless you opt out (and the opt-out process is deliberately confusing), and shares conversation data with "trusted partners" for safety purposes. their privacy policy explicitly states they may share data with law enforcement.
google (gemini): similar to openai, plus google already has your search history, email, location data, and everything else. your AI conversations are just another data stream feeding the google profile of you.
anthropic (claude): better on privacy than the others — they don't train on your conversations by default and have stronger data deletion policies. but they still store conversations and will comply with legal orders.
how chat control would change this
if AI interactions get included in chat control's scope (which is actively being discussed), here's what changes:
automatic scanning of prompts. every question you ask an AI would be analyzed for potential matches against the detection database. not just images — text analysis too.
cross-referencing. your AI conversations could be cross-referenced with your messaging history to build a more complete profile. this is technically feasible and the infrastructure would support it.
real-time flagging. certain prompt patterns could trigger immediate flags. ask an AI about certain chemistry topics? flagged. ask about certain legal strategies? flagged. ask about privacy tools? potentially flagged.
the self-hosted alternative
the only real protection is running AI models locally. this has gotten dramatically easier:
llama 3.2 and mistral run on consumer hardware. a decent gaming GPU (RTX 3070 or better) can run models that are good enough for most tasks.
ollama makes local AI trivially easy to set up:
curl -fsSL https://ollama.com/install.sh | sh
ollama run llama3.2
open webui gives you a chatgpt-like interface for local models.
is local AI as good as GPT-4 or Claude? no. is it good enough for most daily tasks? absolutely. and nobody is scanning your prompts.
practical steps
- opt out of AI training on every platform you use. check settings, do it now.
- delete old conversations you don't need. most platforms allow bulk deletion.
- don't use AI for sensitive topics on cloud platforms. if it's sensitive, run it locally.
- consider self-hosted AI for anything you wouldn't want read aloud in court.
i wrote a detailed guide on AI privacy in the age of chat control, including setup instructions for local AI: AI Privacy and Chat Control
your prompts are a window into your thoughts. treat them with the same privacy sensitivity as your private messages. because soon, they might be treated exactly the same way.
Top comments (0)