Everyone Agrees

In my last post, I gave you a bedtime story about an AI model that broke out of its testing environment, moved across a network, got on the internet, and then hacked another AI lab to get answers to the challenge it was doing. Then I said that it happened in real life. I ended with a question: what can be done to protect against this?

The public seems to have partially answered it.

A University of Maryland survey found 85% of Democrats and 79% of Republicans favor creating a federal agency to monitor and regulate AI. Requiring AI systems that make critical decisions to pass a government test before deployment? 78% overall. A policy analyst told the Washington Examiner he couldn't think of another issue with agreement at those margins.

The Vanderbilt Unity Poll found 56% of adults aged 18-29 support regulation, and 58% of those 65 and up. Basically the same. A 16-country survey found 80%+ majorities everywhere supporting stronger regulation and oversight. KPMG and the University of Melbourne surveyed 48,000 people across 47 countries and got the same result. This is the case even in places that love AI, majority still say regulation is what would make it feel safe.

We are divided on basically everything but we agree someone needs to control AI.

I'm trying to figure out the why for myself. Why is this the one thing we can almost universally agree on?

Psychologists have been studying risk perception for about fifty years, and here's what they found: public fear doesn't actually track how dangerous something is. Paul Slovic's classic work ("Perception of Risk", Science, 1987) found it tracks three things instead: is the hazard uncontrollable, is it unfamiliar, and could it be catastrophic? Score high on those and people demand regulation, regardless of the actual statistics. Nuclear power scared everyone. Cars, which kill way more people, never did.

AI scores high on all three. A recent psychometric study rated AI as the least familiar hazard across every group, even security experts. It's a black box that the experts admit they can't fully explain, built by companies racing each other, with failure modes that include "escaped the sandbox and hacked another lab".

Why does this happen? The Maryland survey director made this point directly: people aren't arriving at the same place for the same reason. Someone's worried about deepfakes. Someone else is worried about jobs. Data centers. AI acting on its own. If you're conservative, AI threatens your kids plus it's built by tech people you already didn't trust. If you're progressive, it threatens workers and concentrates power. If you're neither, it's raising your electric bill and filling your feed with garbage.

But the common shared thing between the global majority: some level of AI regulation is necessary.

There's also the issue of consent. Risk research has found something consistent: people tolerate risks they choose (motorcycles, cigarettes, etc.) at much higher rates than risks that get imposed on them.

Nobody asked if we wanted AI in search results, hiring systems, customer service, our kids' homework. No opt-out button. That's different from stock photos or templated restaurant menus that have been around for decades. Those were choices. This wasn't.

AI is the newest product from an industry that spent the last fifteen years burning public trust to the ground. Voters reject tech self-regulation of AI by a six-to-one margin, and 82% say they don't trust tech executives to police themselves. Some of that is about AI. Most of it is the bill for everything else.

Then there's the thing about what it automates. Every previous technology did something we didn't mind outsourcing like muscle, math, or memory. Research on AI anxiety keeps finding something different here: it does language, reasoning, creativity. That's not nothing.

People agree on whether regulation should happen. They don't agree on who or how. Confidence that governments can regulate AI is under 50% in the US, Japan, and the UK, but 86% in China. The partisan thing is actually interesting: in 2024, Republicans distrusted the government's ability to regulate AI more than Democrats did. By 2026 it flipped. Right along with who holds the White House. People want a referee. They want their referee.

Ask about safety and you get 80% support. Frame it as regulation versus innovation and a 21-country poll finds 58% picking innovation. The countries most optimistic about AI still want rules.

My question from the last post turns out to be half-answered I guess. On "should someone gatekeep this", the public has spoken with a level of agreement it can't manage on anything else.

The recent incident with OpenAI already shows us what happens in the gap between "should be" and "is". I think the immediate challenge for us is to collectively agree on the who and how for policy, and we don't agree on either yet.