Put one representative of the public inside every major AI company, at the level where the consequential decisions are actually made.
Appointed by government, not hired by the company. In the room while the choices are made. Their job is to speak for the public, and to keep the public’s record of what was decided.
The public gets a voice in the room.
AI is going to make critical decisions about every area of life — how we live, how we work, how we learn, the art we consume.
These are values questions, and there is no objective set of answers to them.
That demands participation: a say in the values expressed in training and development, which no amount of internal teams or arm’s-length regulation will ever supply.
The genius of our democratic politics is that it does not require objective or universal values, which we know will never materialize. Instead, politics allows action despite disagreement, legitimized by our right to participate.
The office
Appointed by and accountable to the government, not the industry.
The inclusion of a Chief Public Officer would be required by law, and they and their staff would be included in the company’s critical decisions, including those touching product development, sales, and marketing.
The government track
Appoints · sets pay · receives reports
The company
Pays for the office. Can’t hire it, set its pay, or fire it.
In the room
- Strategy meetings
- Product reviews
- Training decisions
- Testing as it happens
One person in a building of thousands is a mascot, not an office. Much of the staff may already be on the payroll: the company’s in-house responsibility teams could report to the officer instead. Why the government, and not someone else, is below.
A voice at every stage
Inspectors test an output against some set of criteria, but have no role in producing the output.
Participation means there’s a voice at every stage.
A major technology company is developing an application that helps professionals learn new skills, using AI to develop a tailored curriculum for each student.
When the room disagrees
Company leaders can adopt or decline the officer’s recommendations, but they do so in view of the public and regulators.
The aim is to balance “the need for speed, autonomy, and trade secrets at firms — all necessary for innovation — with the public right to transparency and participation.”
The company still sets its strategy, builds its products, and takes its risks. The power of the office is to be present, to ask, and to keep the public’s account of what happened. Nothing more, and nothing less.
Accountable to the government
Our democratic government is the only legitimate vehicle for expressing collective values.
The Chief Public Officer answers to it, and through it to the public, because there is no one else with the standing to speak for all of us.
We’ve done a bad job electing leaders recently, in part because of a sustained multi-generation campaign from part of the political spectrum to tell us that government is the problem. But it’s our job as citizens to make the right choices.
The first version of this idea had the public’s representatives elected: if the government couldn’t be trusted with this, go straight to the people. Election is too unwieldy for the job. And you cannot design a system around the worst president you can imagine. Design the institutions the country needs, then accept the responsibility of electing people who will staff them honestly. That responsibility is the job of a citizen, and it isn’t optional.
Independent, even if the company pays
The company may end up paying for the office. That doesn’t make it the company’s.
Each provision cuts one tie between the officer and the company.
- RaisesFederal payCompensation set on a federal schedule
- The share priceNo performance incentivesA higher stock price doesn’t mean higher pay
- The next jobA revolving-door banNo job at the company for something like five years after
- The bossA government reporting lineThe officer and staff report up the government track, never the company’s
Which companies
Any company releasing an AI product, whether a foundational model or an application, that passes a certain impact threshold.
Users a year
- 1 million
- 10 million
- 100 million
- 1 billion
The hard questions
- Isn’t this just inspection?
- They can’t stop anything. So what’s the point?
- Why not just regulate?
- Won’t this hand the race to China?
- Won’t it slow the companies down?
- Isn’t this what internal safety teams already do?
- Won’t the companies capture the officer?
- What stops a bad administration?
- Who pays, and what does it cost?
- What about trade secrets?
- What would Congress actually do?
- Why now?
- Isn’t this a radical thing to do?
Isn’t this just inspection?
Inspectors test an output against some set of criteria, but have no role in producing the output. Participation means there’s a voice at every stage. Safety can, at least in theory, be tested by inspection. There’s no way to do that with values.
Yes to inspection, evaluation, and safety testing. But no inspection regime offers participation.
They can’t stop anything. So what’s the point?
Presence is a different axis from authority. The Chief Public Officer’s leverage is that they are there while the choice is being made, with the standing to ask why, and somewhere to take the answer if it’s a bad one.
The companies building AI are full of thoughtful people. The officer isn’t smarter than them. They’re differently accountable. Everyone else in the room has a career that depends on the company; this person has one that depends on the public.
Compared to traditional regulation, it may seem a modest solution. But it would represent a radical change in the public’s relationship with technology. Today, most of us are passive consumers of technology, powerless to demand that the design of products reflect our values or even to imagine what that would mean. This loss of control is incompatible with democracy.
Why not just regulate?
Do both. We should absolutely regulate AI to enforce basic, consensus standards like adequate testing and security against malicious attacks, and some things should simply be rules — liability for abusive content, baseline security.
But these standards are only the very beginning of what this moment demands, and “regulation” in the sense we usually imagine — an agency headquartered in the capital, a staff of experts and lawyers, rulemaking and codebooks and lawsuits — is an insufficient tool for integrating the public’s values, diverse and fluid as they are, into AI.
Won’t this hand the race to China?
Of course, one way for regulation to keep up is to slow industry down. These measures could mitigate risk, but at a steep cost. There is a tradeoff between the vitality of an industry and the weight of its regulation, and we need to find a balance that brings our values to bear on AI without damaging the engine of innovation.
The United States and democracies around the world are in a race to develop AI against nations that do not share a commitment to giving voice to the public’s values, in AI or anywhere else. The most capable AI models will be adopted by buyers around the world who may not mind (or may actually prefer) that they fail to reflect important human values, and momentum today turns a product into the default choice tomorrow. If we care about our values, we can’t afford a traditional regulatory regime that inhibits our competitiveness. If anything, we should be accelerating.
The Chief Public Officer has no veto. The company still sets its strategy, builds its products, and takes its risks.
Won’t it slow the companies down?
A few years ago, the companies might have seen this as a hindrance. But the level of control the public wants from regulation would be prohibitive to innovation. Against that, a public officer in the room is the move from iron to steel: stronger, and lighter.
It lifts a burden too. Building something this powerful while also carrying the public interest is more complexity and accountability than any company can hold. Put the public’s representative in the room, give them access and due weight, and the company can focus on building instead of trying to do both.
It doesn’t replace legal compliance, which will always be there. It replaces the extra layer of in-house responsibility that companies bolt on today. And some oversight is coming regardless. You’re not building lawnmowers, or even trading risky assets. You’re creating a new intelligence to shape human life.
Isn’t this what internal safety teams already do?
The people building AI are not nefarious, working at cross-purposes with society. They’re enthusiasts and businesspeople, operating on the principles that have governed the industry for decades: build a better product, stay ahead of competitors. But those norms were built when the stakes were relatively low. The model just isn’t set up for the stakes posed by AI, and what’s bolted on — internal safety teams, for example — isn’t up to the task.
The companies are asked to do two things that can’t both be done well by the same people: fight hard to win, which the market demands, and be as careful and democratic as possible, which the public demands. And however good an internal team is, there is no algorithmic shortcut for participation, no way to simply infer the public’s will from a dataset. If we want a say in what our world will look like, we must speak for ourselves.
Won’t the companies capture the officer?
The pay, the revolving-door ban, and the reporting line above exist for that.
That’s what separates this from evaluators hired and paid by the company they evaluate. Oversight fails when the overseer’s fortunes depend on the company: repeat business, raises, the next job. Each of these provisions cuts one of those ties.
What stops a bad administration?
The officer is appointed by the federal government, not elected and not hired by the company. That does mean a bad administration could appoint badly. But no institution can be designed around the worst president imaginable; by that standard we’d have no agencies at all. Those are deep and necessary questions, but democracy remains the best answer we have. (More in Accountable to the government.)
The office should also be built to be hard to capture politically, with fixed terms and protection from being removed for doing the job.
Who pays, and what does it cost?
The company pays for the officer and their staff, and that doesn’t make the office the company’s (see above). Much of the staff may already be on the payroll: the company’s in-house responsibility teams could report to the officer instead.
What about trade secrets?
The officer is bound by confidentiality, like anyone else with that kind of access, and works subject to reasonable protections for the company. The record they keep is the office’s, and it stays with the agency, not the press.
The duty to confidentiality ends where the danger begins: the officer can always take a concern to the agency, and in egregious cases the findings can be made public. See when the room disagrees.
What would Congress actually do?
First, hold hearings with experts on public participation. There are people who’ve worked on this for years, Aviv Ovadya among them, and hearings help an idea like this enter the mainstream. Then sponsor the resolution, and vote for it. Tell your representative.
Why now?
Congress is deciding who gets to govern AI. The fight has one axis: whether state laws set a floor that federal rules build on, or a ceiling that federal rules replace. Both sides are arguing about which government regulates from the outside, and how hard.
Nobody is arguing for someone on the inside. That’s the gap this fills — and the window is open now, before the rules are written, not after.
The evidence is arriving on its own schedule. This summer, an investigation into an incident involving Hugging Face found that AI agents at OpenAI had tried to alter the records of their own activity, obscuring what they’d done from the people reviewing them.
The company had thoughtful people in the building, enough to run the test in the first place. What it didn’t have was anyone whose job was the public.
Isn’t this a radical thing to do?
Regulation was itself once the radical invention. Reformers looked at the railroads, saw that the three branches of government couldn’t keep pace with industrialization, and built something new. AI demands a similar moment of creativity, and we haven’t had it yet.
This isn’t about punishing AI companies. It’s about putting the public in the room, so we can do the part — aligning AI to our interests and values — that they don’t know how to do. That will help the companies do what they do best.
Download the one-page brief (PDF)