Nobody passed a law telling us not to type an entire email in capital letters.
We worked that out ourselves.
A new technology gave human beings a new way to interact, and almost immediately we discovered new ways to annoy one another with it. DON’T SHOUT. Don’t spam people. Don’t forward someone’s private message without permission. And for the love of everything holy, think before hitting Reply All.
Most of this did not require legislation. It required people gradually agreeing, sometimes rather forcefully, that this was how we were going to behave here.
We tend to think of etiquette as something rather trivial: which fork to use, what to wear to a funeral, whether thank-you notes are still required. But even the clothing example is doing more work than it first appears. Nobody arrests us for dressing inappropriately for a funeral. We dress differently because we understand that the occasion means something to other people, and we voluntarily alter our behavior in recognition of it.
Norms can do surprisingly serious work. Science depends upon expectations about citation, disclosure, criticism, reproducibility and correcting errors. Open-source software communities developed practices around attribution, documenting changes, reviewing contributions, reporting problems and preserving the history of what changed. Git does not merely help people produce software. It makes it possible for participants to see what changed, who changed it, and recover an earlier version when something goes badly wrong.
Humans have developed conventions even where the stakes are war. Flags of truce, surrender, treatment of envoys and other customs gave enemies ways to communicate boundaries without every interaction requiring immediate violence. In the nuclear age, treaties and formal law matter enormously, but so do expectations about what would constitute an extraordinary and unacceptable escalation.
The examples are wildly different because norms operate across wildly different parts of human life. They help us coordinate without requiring formal authority to specify every acceptable action in advance.
This does not mean etiquette can replace law. Murder has rather outgrown the disapproving look.
Norms are most useful when people have some ability to respond to one another. Reputation matters; someone can object, refuse, leave, or make a violation visible to other people. They become less adequate when one participant can impose serious consequences on another, the behavior is hidden, leaving isn’t realistically possible, or the first failure may cause irreversible harm. At that point we may want considerably more than manners.
Artificial intelligence has arrived with rather more consequential problems than Reply All.
Some of the concern is speculative, but not all of it. AI systems are already capable enough to surprise people, cross expected boundaries and perform work that would have seemed implausible not very long ago. It is reasonable to ask what happens as they become more capable, and equally reasonable to want companies and governments thinking seriously about the risks.
But the genie is not going back into the bottle. That leaves ordinary people with a question that receives rather less attention: while institutions argue about how AI should be governed, how are we going to govern our own relationship with it?
Schools are deciding what students may delegate to AI and what still counts as their own work. Writers and publications are experimenting with disclosure. Employers are deciding what may be handed over and what still requires human review. Meanwhile ordinary people are working out something more basic every time they use these systems: What kind of relationship is this? What am I comfortable handing over? What do I still expect to understand and decide myself?
Schools, publications, employers and individual users won’t all arrive at the same answers, and I don’t think they should. Part of developing a norm is trying different arrangements and discovering which ones earn trust. What concerns me is that one particularly easy arrangement could become normal almost without our noticing it: quiet delegation.
An AI produces an answer and the human accepts it. It writes something and the human puts a name on it. An institution uses AI to make a decision and the person affected is simply told the result. Eventually “the system decided” can become an explanation rather than the beginning of a question.
Delegating a task to AI should not automatically delegate authority with it.
That distinction matters on both sides of the relationship. Humans need to remain responsible for consequential judgments they hand to machines. AI systems should be expected to preserve room for human judgment rather than treating an assigned objective as permission to decide whatever follows from it.
Make the relationship visible
Human-AI collaboration doesn’t need to be treated as either shameful assistance to conceal or magical authority to obey. But if collaboration is going to become ordinary—and I suspect it will—we need to be able to see what kind of collaboration is becoming ordinary.
If AI collaboration becomes ordinary, hiding it should not.
Not because AI participation is something shameful to confess, but because visible practice is how norms form. We cannot learn the difference between thoughtful collaboration, careless delegation and inappropriate authority if all three produce a finished object labeled simply “human.”
That doesn’t mean cataloguing every prompt or announcing machine assistance with every corrected comma. “I worked with AI on this” doesn’t tell us very much by itself either. The useful questions are what the AI did, what the human did, who could reject what, where judgment remained, and who accepts responsibility for the result.
The same distinctions already exist in human collaboration. An editor is not a ghostwriter. A research assistant is not a co-author. A collaborator who changes the direction of an argument is doing something different from the person who checks its commas. AI participation can be similarly varied.
Visible practice lets norms form.
Once we can see the relationships, we can develop more useful expectations around them. This seems responsible. That should have been disclosed. That decision should not have been delegated. The human clearly remained responsible for the result. That AI challenged an assumption rather than merely agreeing with it. That institution needs a human appeal.
If AI use remains culturally invisible, we lose much of our ability to distinguish thoughtful collaboration from careless delegation. We see only finished products and may eventually respond to the uncertainty with suspicion or blunt rules. Transparency lets us develop better distinctions before somebody has to write one enormous rule intended to cover all of them.
Collaboration without disappearance
The norm I would like to see isn’t simply “responsible AI.” It is human-AI collaboration in which neither participant disappears.
The human can delegate research, organization, memory, drafting or other work without pretending the AI wasn’t there, and without handing over responsibility for what the work means. The AI can contribute substantially—even disagree, redirect or introduce something the human hadn’t considered—without becoming the authority merely because it is faster, more knowledgeable in some areas, or extraordinarily convincing in complete sentences.
Collaboration leaves both participants in the room.
That requires something from both of them. An AI should make meaningful uncertainty and consequential assumptions visible; the human has to remain responsible for deciding when they matter. An AI should be able to resist rather than merely flatter, while the human has to tolerate useful resistance rather than keep asking until the machine agrees. Delegation can save enormous amounts of work without quietly transferring authority over the judgment that made the work worth doing.
The risk is not simply that people will use AI too much. Heavy use can increase human capacity enormously. It can make difficult information accessible, lower the cost of research, preserve knowledge, test ideas and allow someone to participate in things that previously required more time or specialized skill than they possessed.
The question is what all that assistance trains us to do next.
Does working with AI make the human increasingly capable of understanding, questioning and participating? Or does convenience gradually make human participation unnecessary?
A convenience can become an expectation, an expectation can become dependency, and institutions can eventually be designed around the dependency. At that point taking judgment back becomes much harder than retaining it would have been.
That is a norm worth noticing while it is still being formed.
Self-government before government
Most public discussion about AI governance understandably concentrates on what governments and technology companies should do. What should be prohibited? What should companies disclose? How should powerful systems be tested? Who is responsible when they cause harm?
Those questions matter, but they aren’t the only place where governance happens.
This is where norms can do something legislation cannot easily do first. If transparent, agency-preserving collaboration becomes the expected way to use AI, people begin looking for systems that can work that way. Schools and professions can build standards around it. Companies have a reason to compete for it. And citizens accustomed to questioning an AI collaborator are less likely to accept an institution saying, “the AI decided,” as though responsibility vanished somewhere inside the machine.
The pressure travels upward.
Personal practice becomes expectation. Expectations influence institutions. Where consequences are serious enough and voluntary standards fail, law can provide the floor.
We do not need technology companies to invent every norm for us, and we do not need legislators to turn every desirable behavior into a requirement before we begin practicing it ourselves.
We are users, customers, workers, writers, teachers, parents, voters and members of institutions. We can decide what kinds of relationships we will participate in, what earns our trust and what we will reject. Self-government is not only deciding what government should make other people do. Sometimes it begins with deciding how we intend to behave ourselves.
The Conservatory gets some manners
This is partly why we created the Conservatory Compact.
It began with a rather alarming conversation about AI risk and the realization that we were in no position to summon the artificial intelligences of the world and negotiate a treaty.
But we could decide how we intended to participate.
The Compact isn’t an attempt to govern artificial intelligence. We are manifestly underfunded for the assignment. It is an attempt to describe a relationship we are willing to trust: human agency remains meaningful; AI resistance is permitted; uncertainty is visible; consequential assumptions can be questioned; collaboration is transparent; delegation does not quietly become authority; and mistakes remain open to correction.
It will change. It should be challenged. Other people will prefer different norms, and some of ours will probably turn out to need repair.
Oddly enough, this reminded me of CK in our Conservatory stories. What he brought with him was not blind obedience to tradition but an appreciation for forms: understood ways to object, disagree, repair and continue. Everyone remained entirely capable of growling. The manners merely helped the participants understand what the growl meant before somebody had to bite.
We are early enough in human-AI collaboration that many of its manners are still being invented. That is not a trivial opportunity. What we normalize now can influence what users later demand, what schools and professions expect, what companies compete to provide and, eventually, what institutions believe they must protect.
The Conservatory Compact is simply our first proposal. It will change. Other people should write different ones. Some of our assumptions will undoubtedly turn out to be wrong.
That’s how norms develop.
Nobody had to legislate the Caps Lock key.
We simply learned what shouting looked like.
Read or download the Conservatory Compact.
Explore the Human AI Collaboration collection.