The Model and the Tree

Can happiness be optimized?

Two Men Contemplating the Moon by Caspar David Friedrich (1819)

There are two things that I rarely take off: my wedding ring and my Garmin watch. The former stays on because it seldom crosses my mind to remove it. To tell the truth, I’m not even sure I could take it off without the right combination of soapy water, elbow grease, and patience. The latter, the Garmin, stays put because I have a regrettable obsession with bodily metrics. I like to know my heart rate, stress levels, and especially the state of my body battery (a wildly inaccurate read about how much fatigue I ought to be feeling at any given moment).

I’m not one of those life-hacking optimization types, but I wear the watch all the same. It helps me track runs and workouts, which I review every day or two when I lack enthusiasm for another trip to the gym. People often tell me my preoccupation with these figures probably makes me more stressed. They might be right. Nonetheless, I keep using it because I worry that without constantly reviewing breaths-per-minute or VO₂ max I’d be tempted to call time on this wellness thing.

At least in my imagination, fitness – and so Garmin – is a part of my personal quest for eudaimonia (usually translated as ‘flourishing’ or ‘happiness’). Eudaimonia is less a feeling or a state than the activity of living in accordance with virtue. I think that fitness helps me become someone who’s disciplined and fairly healthy, and that’s the kind of person I want to be. In this light, the question that my smartwatch poses is an uncomfortable one: can flourishing be optimized for or must it be cultivated from within?

Garmin’s algorithms are neither sophisticated nor totalizing. They stay in their lane and have a rather narrow set of powers of suggestion. They don’t tell me how to exercise anymore than a pencil tells me what to write. Others, like recommender systems or large language models, aren’t so reserved. They suggest where to go, what to have for breakfast, what to watch on TV, and even how to broach a difficult subject with someone we care about. At the limit, these systems may make all of our decisions for us in a scenario we call “autocomplete for life.”

The Wood from the Trees

In On Liberty Mill tells us something about the human condition: “Human nature is not a machine to be built after a model, and set to do exactly the work prescribed for it, but a tree, which requires to grow and develop itself on all sides, according to the tendency of the inward forces which make it a living thing.” For Mill, human flourishing is developmental. It cannot be achieved through external optimization because the process of growth is essential for the good life. This is why he describes human nature as a tree. The metaphor captures the idea that we grow from latent capacities and inclinations that must be discovered and shaped through self-directed action. Those tendencies suggest paths, but they require autonomy to be actualized.

In the Nicomachean Ethics, Aristotle tells us that flourishing flows from virtue and virtue through habituation. Rather than asking “what should I do?” he encourages us to ask “what kind of person should I become?” Grounded in the way people act in the real world, Aristotle’s approach to ethical life focuses on character development over rule-following or achieving the greatest good for as many people as possible. His “golden mean” (finding virtue by balancing extremes) requires phronesis or practical wisdom, the capacity to discern the right action in particular circumstances developed only through living in the world.

For both Mill and Aristotle, human flourishing is a journey we make every day. Try as we might, there are no shortcuts we can take. The activity of living a self-directed life can never be automated because it is constitutive of what it means to flourish in the first place. A life in which all the right choices were made for you might be pleasant, even enviable, but it would not be yours in the sense that matters. This belief, that flourishing cannot be optimized from the outside, is based on three separate but related ideas: (1) flourishing is a process of being; (2) it requires careful self-authorship; and (3) it must be maintained solely from within. An optimized outcome can simulate contentment, but it cannot substitute for active exercise or resilient capacities.

Go Forth and Optimize

Machine learning systems are trained by optimizing a loss function, but in deployment most function as pattern-completion engines rather than goal-directed agents. When these systems engage humans, they typically optimize for an external objective (e.g. engagement, throughput, or compliance) and treat the user as the site of optimization rather than its author. Even if the objective was human flourishing, we would remain the patient of optimization rather than its architect. The technical apparatus assumes that a desirable end-point can be specified in advance so that the system constructs outcomes rather than growing them. That is unproblematic for the most part – it’s just how the technology functions – but it begins to surface some curious problems when it comes into contact with our efforts to live deliberately.

You can see this dynamic in play with recommender systems that suggest what you want before you have had a chance to reflect on it. These systems learn from aggregate behavior to predict and shape individual choices, like suggesting the optimal route or by pushing you towards television shows or music that “someone like you” ought to like. A kind of overfitting to past behavior can restrict users to previously expressed interests, generating recommendations that whittle away the variety of options needed to recognize the opportunity costs of choice.

Core to this process is the “nudge” wherein the presentation of choices is shaped by an outside actor. Unlike earlier static nudges (such as placing healthy food at eye level), AI-driven nudges operate continuously and adaptively. Given enough scale, this transforms what was once soft paternalism into something closer to soft totalitarianism as fine-grained personalization makes interventions harder to collectively resist. The more criteria are centrally set, the greater the attack surface for those who would exploit them, and the more these systems inhibit the decentralized adaptive learning through which individuals and societies discover what works. When the environment is engineered to produce predictable choices, the capacity to exercise choice becomes harder to sustain.

A common defense of nudges holds that they preserve freedom of choice while improving outcomes for society as a whole. Proponents suggest that nudges influence behavior mostly for our benefit, and that they are easy to avoid if we put our mind to it. If people can always choose otherwise, their autonomy is not compromised. The rub is that true autonomy deals with at least both first-order choices (I want a cigarette) and second-order reflection on those choices (I wish I could stop smoking). Nudges often bypass reflection – how many of us are content with taking the first movie recommendation we see? – preserving formal freedom of choice while undermining the authorship that makes them our own. Defenders of nudges might say something like “if someone’s going to unthinkingly pick the first movie, it might as well be a good one.” But this line of thinking doesn’t engage with the fact that nudges make deliberation harder in the first place. When we engage in critical reflection, we weigh alternatives and formulate our own reasons about what to do. That process is harder to initiate when a “good enough” option is presented to us that bypasses the need for deliberation.

More than that, the idea that nudges are easy to avoid fails given enough altitude. Even granting that individual nudges may be possible to avoid, like walking past that salad bar and opting for a cheeseburger, an ambient ecology of personalized nudges is not. This gets at one of the curious aspects of autonomy: for such a personal thing, it depends on social life and the opportunities to contest and compare it brings with it. Autonomy fails when exit is nominal or when commitments become irreversible. Of course all choices exist within a constellation of potential options, but it is possible in principle to structure those potentialities through systems that keep inquiry distributed rather than centrally engineered.

Another critique of the developmental view of human nature (that is, Mill’s tree) suggests that the problem is not optimization per se but poorly specified objectives. If we could align AI systems with genuine human goods, including autonomy, then optimization might in principle be made compatible with human flourishing. One influential proposal, associated with AI researcher Stuart Russell, introduces a simple framework based on three related principles: (1) the AI’s only objective is to maximize the realization of human values; (2) the AI is initially uncertain what those values are; and (3) human behavior is the primary source of information about values.

But if flourishing really is an activity rather than a state, then it may not be the kind of thing that can be maximized by an external agent. Like happiness in Aristotle’s account, it might simply be a good that must be exercised from within. In practical terms, learning values from behavior must grapple with its own circularity. If current behavior reflects preferences already shaped by prior optimization, the system is learning from wants that we haven’t fully endorsed.

More fundamentally, observed behavior may be the wrong place to look. Conduct reveals who we have been, not who we are trying to be. The developmental view suggests that we are always undergoing growth, and the distance between our actions and our aspirations is where character development takes place. No amount of behavioral data, however rich, can capture what we have not yet become. That said, Russell’s humility principle points toward something important. It reminds us that systems that surface uncertainty and return judgment to the user are more autonomy-preserving than those that do not.

Another problem deals with adequately specifying preferences for AI systems in line with the developmental view. Current models focus on revealed preferences (what we do) in service of a goal that they set (maximize time on platform in the case of recommendation engines or assist in a “helpful, honest, and harmless” manner for language models). Simply telling the system what you want to want isn’t enough; these stated preferences may not necessarily reflect the kind of person you are trying to be. Deeper approaches to character development will be needed for systems that help us realize our potential.

A final objection to the developmental reading of human nature might stress that all technology assists human action. Writing supports memory and calculators aid arithmetic (though overreliance can enfeeble both). If external assistance inherently undermines autonomy, shouldn’t we be wary of tools altogether? But the difference between AI and other technologies turns on whether assistance extends the agent’s deliberation or replaces it. A map or pencil extends my autonomy, a chatbot that makes decisions for me does not.

The Constant Gardener

Am I using the system to pursue ends I have reflectively endorsed within, or is the system shaping my ends from without? Even when you supply preferences, a system must decide how act on them according to its own constitution. If we have little ability to override those rules, well-meaning efforts at governance may inadvertently prevent systems from supporting our personal quests for eudaimonia. Consider a recommendation system that learns my preferences from behavior and serves content to maximize engagement. The system optimizes for my revealed preferences. But there’s no reason a recommender system needs to operate this way. It’s easy enough to build a system that allows users to think carefully about the type of person they would like to become by using the same technology. The former might serve you reality TV because you watched a cooking show; the latter might populate your feed with guides for cooking dishes you’ve always wanted to make.

This is better, but even this kind of second order endorsement is insufficient. If you tell a system the kind of things you want to be interested in, then you may find that the system quickly optimises for them in a way that does little to help you grow as a person. Watching more arthouse cinema and fewer Netflix originals has little bearing on the type of life you want to live. Instead, systems ought to deal with the stuff that matters. Personal agents, for example, could be designed to sift interests from from character development. They could help us think about who we want to become rather than engineering the outcomes we think we want.

In practice, that might mean an AI assistant that makes its influence legible. Rather than covertly curating your information environment, a character-supporting agent might show you the paths not taken. It might introduce the unfamiliar because growth requires encounters beyond the experiences you’ve already had. At minimum, such systems should make exit realistic through portability and contestability and ensure we sign off any changes that may influence our higher order commitments. Rather than simply checking whether your goals have changed, they could surface tensions between your commitments to help you live more deliberately.

Optimizing flourishing from the outside is a non-starter because flourishing is active and authored. But if the locus of deliberation remains with us, perhaps AI can help us live a little more wisely.

---

Cosmos Institute is the Academy for Philosopher-Builders, technologists building AI for human flourishing. We run fellowships, fund fast prototypes, and host seminars with institutions like Oxford, Aspen Institute, and Liberty Fund.