Alignment Needs a Better Public Conversation

When an AI researcher says there is no plan to solve alignment for superintelligence, the public is entitled to pay attention.

In a recent post, Anthropic researcher Evan Hubinger expressed that concern while also saying he believes the company is trying its best. Those positions can coexist. An organization can take a problem seriously, invest in it, and still lack a convincing path to solving it. His statement deserves a substantive response.

But “we have not solved this” and “we are doing nothing about this” are different claims. Anthropic publishes a Responsible Scaling Policy and risk reports. OpenAI publishes a Preparedness Framework. There is work to examine.

My concern is the distance between that work and the public’s ability to understand it.

I think OpenAI and Anthropic need to put more effort into listening to people and explaining alignment and safety. Their work has implications far beyond their customers. Communicating with the people affected is part of the responsibility that comes with it.

The existence of reports does not finish that job.

I use responsibility in a particular sense: response-ability, the capacity to perceive what is happening, judge it, and respond while a response can still matter. Accountability is distinct. It is the obligation to answer for decisions and their consequences. You can hold someone accountable. Their capacity to respond has to exist before that reckoning.

Applied to AI development, responsibility raises practical questions. What can the lab detect? What can it interrupt? What happens when a system behaves outside expectations? Where does its capacity to intervene end?

Those questions do not have to produce perfect answers to be worth asking. Nobody can foresee every circumstance. A safeguard can be useful without being sufficient for every future system. A plan can represent serious progress while containing unresolved problems.

But the limits belong in the explanation too.

When the consequences extend to the public, response-ability acquires a public dimension. People raise concerns. They ask what is being done in their name, or at their expense, or with risks they cannot choose to avoid. An organization operating at that scale should be able to hear those questions and answer them intelligibly.

That takes more than publishing what the organization has decided to say. It requires discovering what people are actually asking.

Someone asking whether an AI system could escape control may not be asking for an explanation of a benchmark. They may want to know who could intervene, what would trigger that intervention, and whether it would still be possible in time. Someone asking who agreed to the risk is asking a different question again. More technical detail may leave that question entirely unanswered.

Listening is how you find the distinction.

Insider criticism makes this especially important. A researcher speaking personally is not automatically speaking for the whole corporation. But neither does that make their assessment disposable. They may have relevant knowledge, understand the plans perfectly well, and still judge them inadequate.

I cannot infer from that disagreement alone that internal communication has failed. I can say that it creates a public question the organization should address. Where does it agree with the criticism? Where does it disagree? What evidence supports its position? What remains unresolved?

A clear answer might leave people more concerned. That does not make the answer a failure.

The purpose of communication cannot be to eliminate every fear. No set of reports will satisfy everybody. Some disagreements concern evidence; others concern values, acceptable exposure, or who should get to decide. Those disagreements will not disappear because a communications team finds better wording.

This is where legitimacy enters.

I have written about legitimacy as a matter of recognized procedure: how a decision acquires authority beyond the confidence of the person making it. Public acceptance is related, but it is not the same thing. Popularity cannot authorize every risk, and a good explanation cannot substitute for a legitimate decision process.

Still, explaining yourself is part of maintaining a credible relationship with the people your decisions affect. If you want them to recognize your judgment as serious, you must make that judgment open to examination. They need to understand what you believe, why you believe it, and how their concerns can receive a response.

That relationship will require continuing work.

There is no reason to expect a single alignment breakthrough to settle every question about every future system, use, and circumstance. Nor should we imagine one giant red button that resolves the entire social and technical problem. Particular systems can have controls. Particular risks can have solutions. The broader task will continue to change.

The relationship analogy helps here. You do not solve a relationship by getting engaged. You do not reach “happily ever after” and retire from the work of understanding each other. Circumstances change. Disagreements emerge. You have to keep listening, explaining, setting boundaries, and deciding what happens next.

That does not mean accepting every development or staying with an unsafe arrangement. Boundaries and refusal are part of the work.

Our growing involvement with AI makes sustained communication more important, not less. Neither assurances of perfect safety nor certainty of catastrophe should relieve anyone of the need to explain their reasoning. Uncertainty leaves us with work to do.

My request to OpenAI and Anthropic is therefore modest in scope, though substantial in effort: make alignment and safety a better public conversation.

Explain the progress and the gaps together. Answer the difficult question before directing people to a long report. Give informed disagreement a serious hearing. Show what you heard and what, if anything, changed because of it.

Call that better PR if you like. I do. But the public-relations work I mean begins with the relationship.

When your work affects the public, being able to respond includes being able to respond to us.