America Needs the Turing Police
Subsidize signal, tax noise
“We are breathing the fumes of the worst information environment I’ve ever seen,” writes Roon. It’s plausible. On a daily basis an American citizen is fairly likely to see A.I. generated content they can’t distinguish from reality. They might get called by a robot that’s trying to scam them. Maybe their feeds will show them politically motivated disinformation and misinformation; at a minimum it’ll be titillating slop. Men watch porn depicting incest and rape. The cognitive armor grows weak.
The ever more fragmented information environment has been well documented, I won’t belabor the point. Instead, consider the information environment of a functional Republic:
Truth is more widely disseminated and consumed than lies
The attention economy is efficient
The “civic baseline” is met: sufficient information exists for voters to make informed decisions
That’s probably a good start. A.I. content falls under “lies”, and disappears; an efficient attention economy stops allocating so much attention to porn, both soft and hard core; the robots stop filling my voicemail with weird beeps.
How do we get there? Here’s my proposal:
A public polling system
Public journalism allocated under a semi public “DMO” or bed tax model
Restricting A.I. fine tuning jobs to a representative sample of Americans
Establish the Turing police
Let’s briefly look at each one.
Federal Polling System
Elon Musk likes using Twitter polls to justify his issues du jour, and predictably wins. He owns the platform, after all. Private polling in the Trump era has been notoriously fickle. But wouldn’t it be useful to know what our fellow citizens were thinking?
A public polling system is a couple steps shy of “social media as a utility” which has sometimes been articulated, but it covers the basics. Every citizen gets a login. Who articulates the polls can happen at a state level - maybe in Wisconsin it’s the governor only, maybe any elected official, in New York. Maybe anyone, in Maine. But it’s an automated process where people can “vote” in public.
Bots and foreigners do have some undue influence on internet media. Citizen only polling would surface issues important to actual American voters.
Public Journalism
But American voters need information! I’ve watched the Buffalo News, my hometown paper, decline from a full fledged newspaper in my youth to a pamphlet full of reprinted New York Times articles and some loose sports coverage. How should Erie County voters know what’s happening in their local government?
One idea: a semi-public local institution funds journalism. The model could be architected similarly to Destination Marketing Organizations (think “Visit Oregon” or “Visit St. Pete”) which are funded by local “bed tax” monies paid by hotels. The local or state government throws that money towards advertising the polity.
You could set up something similar for information systems. We’d aim to support something like ten full time journalists per million citizens to cover essential civic functions. If we paid each 100K / year, we’d have a cost to the taxpayer of one dollar per year per citizen. Maybe the top college grad in English every year gets a journalist grant, maybe there’s a vote on a couple slots, maybe there’s a permanent slot for police coverage. Cranking from no public dollars to some public dollars probably gets some interesting results.
Representative AI Fine-Tuning
The Center for A.I. Safety recently released some work on emergent values in A.I. They found that A.I. morality is at times in conflict with what we might expect or desire. Among other things, A.I. “value lives in Pakistan > India > China > US”.
We don’t want the U.S. to be at the end of that list! The reason this is happening is probably because the human raters A.I. labs use are low-wage Nigerians, Pakistanis etc., not Americans. Every person has their own bias; if you made, say, only Mexican-American women set their own preferences, the A.I. would be predictably biased towards their preferences.
A.I. alignment is hard. An easy first step is to mandate a representative body of human raters: statistically representative Americans should set A.I. preferences. Whatever morality emerges, emerges democratically and per our own customs.
The Turing Police
A.I. labs are predisposed towards finding technical solutions to A.I. problems. In my opinion enforcement will always be necessary. Enter the Turing Police, a new division under the Department of Homeland Security, responsible for A.I. policing.
Their first order of business? Conclusively shut down the bots calling my cell phone! Despite various efforts Americans still receive billions of robocalls every year. It’s not going to end passively: we need muscular enforcement.
Because there’s more coming. A.I. image generators have already been steered towards child pornography, LLMs to phishing and at-scale scamming. A.I. labs worry about misaligned A.I. making biological weapons. A.I. regulation should fall under a single aegis of a single agency, responsible for enforcing A.I. law. The patchwork regulators aren’t sufficient.
Subsidize Signal, Tax Noise
I don’t think we’re likely to arrive in a Republic of sound information in the near term. We should nevertheless ask our politicians and our political parties: what do you propose for information and intelligence regulation? The Turing police is a good place to start.