The AI Alignment Problem Isn’t About Control. It’s About Relationship.

For years, the conversation around artificial intelligence has revolved around one central question: How do we control it? Researchers have poured enormous energy into reward functions, constitutional AI, reinforcement learning, interpretability, guardrails, and other techniques designed to ensure these systems behave as intended. Those efforts matter, and they deserve continued investment. But the more time I spend thinking about AI, the more I wonder whether we’ve framed the problem too narrowly.

I no longer believe the alignment problem is primarily a technical one. At its core, I think it is a question about relationship.

Every transformative technology has changed humanity’s relationship with power. Fire altered our ability to survive. Electricity reshaped industry and daily life. Nuclear energy forced us to grapple with unprecedented destructive potential. The internet fundamentally changed how we communicate and share knowledge. Artificial intelligence feels different. It may become the first technology that asks us to reconsider our relationship with something that increasingly resembles an intellectual partner rather than simply another tool.

That distinction matters because relationships built entirely on control rarely remain stable forever. History offers countless examples of systems that functioned well while conditions remained predictable but became increasingly fragile as complexity grew. Control has its place, but it has never been sufficient by itself. If AI continues advancing, I suspect the more important question will not simply be how we constrain these systems. It will be whether we can create systems that genuinely embody values compatible with human flourishing rather than merely appearing aligned while operating under supervision.

That is a far more difficult challenge, and it begins with intellectual humility.

For all of our remarkable progress, we still understand surprisingly little about how today’s most capable models arrive at many of their conclusions. We can evaluate their behavior, measure their performance, improve their reliability, and identify many of their weaknesses. Yet we still cannot fully explain why they produce certain insights or why particular reasoning paths emerge. That gap between observation and understanding deserves more attention than it often receives, because confidence should grow only as quickly as understanding does.

The future that concerns me most is not the one popularized by science fiction. I worry less about a rogue superintelligence than about a society that gradually outsources more and more of its own judgment simply because doing so becomes easier. Convenience has always been one of humanity’s most persuasive forces, and artificial intelligence may become its most powerful expression.

Imagine systems that become just capable enough to write our emails, educate our children, diagnose our illnesses, shape political discourse, recommend our friendships, negotiate our contracts, and eventually make thousands of decisions on our behalf. None of those developments necessarily require malicious intent. They may simply leave us with fewer opportunities to practice the habits that make human judgment stronger. Agency is rarely surrendered all at once. More often, it fades through countless small decisions to let something else think for us one more time.

Thinking about alignment this way raises another question that receives far less attention than it should: aligned with whom?

An AI system could be exceptionally safe while faithfully advancing the priorities of whichever corporation, government, or institution controls its infrastructure. From a technical perspective, that system might be considered highly aligned. From a societal perspective, however, it could still contribute to one of the greatest concentrations of power in human history. Technical alignment and alignment with humanity are not necessarily the same thing, and the difference between them deserves much broader public discussion.

There is another uncertainty that I believe we should approach with honesty rather than certainty. No philosopher, neuroscientist, or AI researcher possesses a complete explanation of consciousness or subjective experience. We continue debating whether advanced AI systems could ever possess anything resembling an inner life while acknowledging that we still do not fully understand consciousness in ourselves.

Some of today’s most advanced AI systems readily acknowledge this uncertainty when asked about their own internal experience. They cannot verify whether there is anything it is like to be them. Neither can we confidently prove the opposite. That uncertainty should not be ignored, but it also should not tempt us into unwarranted conclusions. Ethical reasoning has never depended on perfect certainty. The strongest moral frameworks are those that remain thoughtful even when knowledge is incomplete.

History repeatedly reminds us that some of our greatest ethical failures began with excessive confidence in what we believed we already understood. If there is a lesson worth carrying into the age of artificial intelligence, it may be that humility is not a weakness of moral reasoning but one of its greatest strengths.

So what would a relationship-based approach to alignment actually look like? I will be honest: I don’t fully know yet, and I am suspicious of anyone who claims they do. But I don’t think that admission weakens the argument. We have never begun any important relationship — with a person, a community, or a new form of power — by first possessing a complete manual for it. Relationships are built through attention, honesty, correction, and time. Perhaps the work ahead is less about designing the perfect constraint and more about developing the habits a healthy relationship requires: transparency about what these systems are and who they serve, institutions accountable to the public rather than only to shareholders, and a citizenry that stays engaged instead of delegating its judgment away. The answer, in other words, may not be a mechanism. It may be a practice.

Ultimately, I don’t think artificial intelligence is forcing us to answer questions about machines nearly as much as it is forcing us to answer questions about ourselves. How should power be distributed? What does it actually mean for human beings to flourish? Which values deserve preservation as our technologies become more capable? How much convenience are we willing to exchange for autonomy, creativity, judgment, and responsibility?

The alignment problem was never only about persuading machines to do what we want. It is about deciding who we want to become while we build them. The choices we make over the coming years will shape far more than the future of artificial intelligence. They will reveal what kind of civilization we hope to become.