The Public Switch on your AI doctor
Your AI doctor skips your insulin because a data glitch says you’re low risk. Lights out. Now imagine worse: a rogue model on Telegram quietly DMs Democrats hemlock smoothies-great for detox! Who’s at the wheel? The Belarusian teen who leaked it, or the donor who bought the training data? Either way, nobody’s watching.
China blinked first. New law: Xi gets a kill-switch on any AI. Of course we scream tyranny. They call it Tuesday. Take a step back, because here’s the twist: while we’re busy tweeting about freedom, Beijing’s people know their feeds won’t flip them into radicals overnight. Scared? Absolutely! Safer than us? Debatable. We act shocked, but we’ve done this dance. Railways. Power grids, Telco. Anything that quietly steers lives turns public once it hurts or in any way public interest. AI’s just next. Call it the Public Switch. Flip it, Big Tech loses the keys. Europe’s trying. AI Act looks fierce: high-risk models fined, facial scans banned-unless you’re NATO. 4% of your annual revenue at risk if you mess with the rules. Categorize AI risks as if they are apples: low, medium, high risk.
The fundament of the law is sound, but then it’s coffee and carve-outs. Every Monday morning, lobbyists from Siemens and ASML walk into Brussels with a cappuccino in one hand and a 47-page exemption request in the other. Three weeks later the AI Act suddenly has a new footnote: “This rule does not apply if your turnover starts with a billion.” Ordinary citizens get privacy; NATO gets speed-dating with facial recognition.
Startups bolt to Texas. A Dutch team of twelve packs their laptops: “In Austin I train a model in 48 hours. In Eindhoven I first need 400 pages of risk forms.” The big incumbents stay, whine, and collect their personalised waiver like kids swapping Pokémon cards.
Here’s the problem: we want three things at once:
- lightning speed
- bullet-proof safety
- zero paperwork Reality says: pick two. Pick speed + safety = endless forms. Pick speed + no forms = Russian roulette. Pick safety + no forms = snail pace.
Real fix? Treat AI models exactly like dams. Every big language model gets:
- A public ID number
- A yearly inspection sticker
- Blueprints online for every citizen to read
Log every weight. That’s the 175 billion numbers inside GPT-5. Put them in a searchable database. Exception: if the model literally guides a nuclear missile, lock the vault. Everything else: open.
Tax the hoarders, fund the hackers.
- Company keeps its weights secret? → 5 % “black-box tax”.
- That money lands in a pot that pays: — university kids who find bugs — open-source teams who publish clean models — grandma-friendly audit tools
Make trust dull. No more “Trust us, we’re Google.” Just a Google Sheet your neighbour can open on her phone and see: “Okay, this diabetes bot only looks at blood sugar, not at my voting history.”
China is doing it top-down: one red button in Beijing. We can crowd-source the same power: 450 million Europeans vote every quarter:
- Which models stay on?
- Which get paused?
- Which get deleted?
Bottom line: Democracy is slow. It takes debates, lawsuits, and sometimes a referendum. But it is the ONLY system where YOU — the voter, the patient, the parent — can walk up to the switch and pull it back. In Beijing only Xi has fingers on the plug.
I’d rather jog steady at 8 km/h with the whole continent than sprint at 200 km/h with three billionaires and the rest of us over the cliff.