Imagine yourself a decade ago, jumping directly into the present shock of conversing naturally with an encyclopedic AI that crafts images, writes code, and debates philosophy. Won’t this technology almost certainly transform society — and hasn’t AI’s impact on us so far been a mixed-bag? Thus it’s no surprise that so many conversations these days circle around an era-defining question: How do we ensure AI benefits humanity? These conversations often devolve into strident optimism or pessimism about AI, and our earnest aim is to walk a pragmatic middle path, though no doubt we will not perfectly succeed.

While it’s fashionable to handwave towards “beneficial AI,” and many of us want to contribute towards its development — it’s not easy to pin down what beneficial AI concretely means in practice. This essay represents our attempt to demystify beneficial AI, through grounding it in the wellbeing of individuals and the health of society. In doing so, we hope to promote opportunities for AI research and products to benefit our flourishing, and along the way to share ways of thinking about AI’s coming impact that motivate our conclusions.

By trade, we’re closer in background to AI than to the fields where human flourishing is most-discussed, such as wellbeing economics , positive psychology , or philosophy , and in our journey to find productive connections between such fields and the technical world of AI, we found ourselves often confused (what even is human flourishing, or wellbeing, anyways?) and from that confusion, often stuck (maybe there is nothing to be done? — the problem is too multifarious and diffuse). We imagine that others aiming to create prosocial technology might share our experience, and the hope here is to shine a partial path through the confusion to a place where there’s much interesting and useful work to be done. We start with some of our main conclusions, and then dive into more detail in what follows.

One conclusion we came to is that it’s okay that we can’t conclusively define human wellbeing. It’s been debated by philosophers, economists, psychotherapists, psychologists, and religious thinkers, for many years, and there’s no consensus. At the same time, there’s agreement around many concrete factors that make our lives go well, like: supportive intimate relationships, meaningful and engaging work, a sense of growth and achievement, and positive emotional experiences. And there’s clear understanding, too, that beyond momentary wellbeing, we must consider how to secure and improve wellbeing across years and decades — through what we could call societal infrastructure : important institutions such as education, government, the market, and academia.

One benefit of this wellbeing lens is to wake us to an almost-paradoxical fact: While the deep purpose behind nearly everything our species does is wellbeing, we’ve tragically lost sight of it. Both by common measures of individual wellbeing (suicide rate, loneliness, meaningful work) and societal wellbeing (trust in our institutions, shared sense of reality, political divisiveness), we’re not doing well, and our impression is that AI is complicit in that decline. The central benefit of this wellbeing view, however, is the insight that no fundamental obstacle prevents us from synthesizing the science of wellbeing with machine learning to our collective benefit.

This leads to our second conclusion: We need plausible positive visions of a society with capable AI, grounded in wellbeing. Like other previous transformative technologies, AI will shock our societal infrastructure — dramatically altering the character of our daily lives, whether we want it to or not. For example, Facebook launched only twenty years ago, and yet social media’s shockwaves have already upended much in society — subverting news media and our informational commons, addicting us to likes, and displacing meaningful human connection with its shell. We believe capable AI’s impact will exceed that of social media. As a result, it’s vital that we strive to explore, envision, and move towards the AI-infused worlds we’d flourish within — ones perhaps in which it revitalizes our institutions, empowers us to pursue what we find most meaningful, and helps us cultivate our relationships. This is no simple task, requiring imagination, groundedness, and technical plausibility — to somehow dance through the minefields illuminated by previous critiques of technology . Yet now is the time to dream and build if we want to actively shape what is to come.

This segues into our final conclusion: Foundation models and the arc of their future deployment is critical. Even for those of us in the thick of the field, it’s hard to internalize how quickly models have improved, and how capable they might become given several more years. Recall that GPT-2 — barely functional by today’s standards — was released only in 2019 . If future models are much more capable than today’s, and competently engage with more of the world with greater autonomy, we can expect their entanglement with our lives and society to rachet skywards. So, at minimum, we’d like to enable these models to understand our wellbeing and how to support it, potentially through new algorithms, wellbeing-based evaluations of models and wellbeing training data. Of course, we also want to realize human benefit in practice — the last section of this blog post highlights what we believe are strong leverage points towards that end.

The rest of this post describes in more detail (1) what we mean by AI that benefits our wellbeing , (2) the need for positive visions for AI grounded in wellbeing , and (3) concrete leverage points to aid in the development and deployment of AI in service of such positive visions . We’ve designed this essay such that the individual parts are mostly independent, so if you are interested most in concrete research directions, feel free to skip there .

Discussion about AI for human benefit is often high-minded, but not particularly actionable, as in unarguable but content-free phrases like “We should make sure AI is in service of humanity.” But to meaningfully implement such ideas in AI or policy requires enough precision and clarity to translate them into code or law. So we set out to survey what science has discovered about the ground of human benefit, as a step towards being able to measure and support it through AI.

Often, when we think about beneficial impact, we focus on abstract pillars like democracy, education, fairness, or the economy. However important, none of these are valuable intrinsically. We care about them because of how they affect our collective lived experience, over the short and long-term. We care about increasing society’s GDP to the extent it aligns with actual improvement of our lives and future, but when treated as an end in itself, it becomes disconnected from what matters: improving human (and potentially all species’) experience.

In looking for fields that most directly study the root of human flourishing, we found the scientific literature on wellbeing. The literature is vast, spanning many disciplines, each with their own abstractions and theories — and, as you might expect, there’s no true consensus on what wellbeing actually is. In diving into the philosophy of flourishing, wellbeing economics, or psychological theories of human wellbeing, one encounters many interesting, compelling, but seemingly incompatible ideas.

For example, theories of hedonism in philosophy claim that pleasure and the absence of suffering is the core of wellbeing; while desire satisfaction theories instead claim that wellbeing is about the fulfillment of our desires, no matter how we feel emotionally. There’s a wealth of literature on measuring subjective wellbeing (broadly, how we experience and feel about our life), and many different frameworks of what variables characterize flourishing. For example, Martin Seligman’s PERMA framework claims that wellbeing consists of positive emotions, engagement, relationships, meaning, and achievement. There are theories that say that the core of wellbeing is satisfying psychological needs , like the need for autonomy, competence, and relatedness. Other theories claim that wellbeing comes from living by our values . In economics, frameworks rhyme with those in philosophy and psychology, but diverge enough to complicate an exact bridge. For example, the wellbeing economics movement largely focuses on subjective wellbeing and explores many different proxies of it, like income, quality of relationships, job stability, etc.

After the excitement from surveying so many interesting ideas began to fade, perhaps unsurprisingly, we remained fundamentally confused about what “the right theory” was. But, we recognized that in fact this has always been the human situation when it comes to wellbeing , and just as a lack of an incontrovertible theory of flourishing has not prevented humanity from flourishing in the past, it need not stand as a fundamental obstacle for beneficial AI. In other words, our attempts to guide AI to support human flourishing must take this lack of certainty seriously, just as all sophisticated societal efforts to support flourishing must do.

In the end, we came to a simple workable understanding, not far from the view of wellbeing economics: Human benefit ultimately must ground out in the lived experience of humans . We want to live happy, meaningful, healthy, full lives — and it’s not so difficult to imagine ways AI might assist in that aim. For example, the development of low-cost but proficient AI coaches, intelligent journals that help us to self-reflect, or apps that help us to find friends, romantic partners, or to connect with loved ones. We can ground these efforts in imperfect but workable measures of wellbeing from the literature (e.g. PERMA), taking as first-class concerns that the map (wellbeing measurement) is not the territory (actual wellbeing) , and that humanity itself continues to explore and refine its vision of wellbeing.

More broadly our wellbeing relies on a healthy society, and we care not only about our own lives, but also want beautiful lives for our neighbors, community, country, and world, and for our children, and their children as well. The infrastructure of society (institutions like government, art, science, military, education, news, and markets) is what supports this broader, longer-term vision of wellbeing.

Each of these institutions have important roles to play in society, and we can also imagine ways that AI could support or improve them; for example, generative AI may catalyze education through personal tutors that help us develop a richer worldview, may help us to better hold our politicians to account through sifting through what they are actually up to, or accelerate meaningful science through helping researchers make novel connections. Thus in short, beneficial AI would meaningfully support our quest for lives worth living, in both the immediate and long-term sense.

So, from the lofty confusion of conflicting grand theories, we arrive at something sounding more like common sense. Let’s not take this for granted, however — it cuts through the cruft of abstractions to firmly recenter what is ultimately important: the psychological experience of humans. This view points us towards the ingredients of wellbeing that are both well-supported scientifically and could be made measurable and actionable through AI (e.g. there exist instruments to measure many of these ingredients). Further, wellbeing across the short and long-term provides the common currency that bridges divergent approaches to beneficial AI, whether mitigating societal harms like discrimination in the AI ethics community , to attempting to reinvigorate democracy through AI-driven deliberation , to creating a world where humans live more meaningful lives , to creating low-cost emotional support and self-growth tools , to reducing the likelihood of existential risks from AI , to using AI to reinvigorate our institutions — wellbeing is the ultimate ground.

Finally, focusing on wellbeing helps to highlight where we currently fall short. Current AI development is driven by our existing incentive systems: Profit, research novelty, engagement, with little explicit focus on what fundamentally is more important (human flourishing). We need to find tractable ways to shift incentives towards wellbeing-supportive models (something we’ll discuss later), and positive directions to move toward (discussed next).

Technology is a shockingly powerful societal force. While nearly all new technologies bring only limited change, like an improved toothbrush, sometimes they upend the world. Like the proverbial slowly-boiling frog, we forget how in short order the internet and cellphones have overhauled our lived experience: the rise of dating apps, podcasts, social networks, our constant messaging, cross-continental video calls, massive online games, the rise of influencers, on-demand limitless entertainment, etc. Our lives as a whole — our relationships, our leisure, how we work and collaborate, how news and politics work — have dramatically shifted , for both the better and worse.

AI is transformative, and the mixed bag of its impacts are poised to reshape society in mundane and profound ways; we might doubt it, but that was also our naivety at the advent of social media and the cell-phone. We don’t see it coming, and once it’s here we take it for granted. Generative AI translates applications from science fiction into rapid adoption: AI romantic companions ; automated writing and coding assistants ; automatic generation of high-quality images , music , and videos ; low-cost personalized AI tutors ; highly-persuasive personalized ads ; and so on.

In this way, transformative impact is happening now — it does not require AI with superhuman intelligence — see the rise of LLM-based social media bots; ChatGPT as the fastest-adopted consumer app; LLMs requiring fundamental changes to homework in school. Much greater impact will yet come, as the technology (and the business around it) matures, and as AI is integrated more pervasively throughout society.

Our institutions were understandably not designed with this latest wave of AI in mind, and it’s unclear that many of them will adapt quickly enough to keep up with AI's rapid deployment. For example, an important function of news is to keep a democracy’s citizens well-informed, so their vote is meaningful. But news these days spreads through AI-driven algorithms on social media, which amplifies emotional virality and confirmation bias at the expense of meaningful debate. And so, the public square and the sense of a shared reality is being undercut, as AI degrades an important institution devised without foresight of this novel technological development.

Thus in practice, it may not be possible to play defense by simply “mitigating harms” from a technology; often, a new technology demands that we creatively and skillfully apply our existing values to a radically new situation. We don’t want AI to, for example, undermine the livelihood of artists, yet how do we want our relationship to creativity to look like in a world where AI can, easily and cheaply, produce compelling art or write symphonies and novels, in the style of your favorite artist? There’s no easy answer. We need to debate, understand, and capture what we believe is the spirit of our institutions and systems given this new technology.

For example, what’s truly important about education? We can reduce harms that AI imposes on the current education paradigm by banning use of AI in students’ essays, or apply AI in service of existing metrics (for example, to increase high school graduation rates). But the paradigm itself must adapt: The world that schooling currently prepares our children for is not the world they’ll graduate into, nor does it prepare us generally to flourish and find meaning in our lives. We must ask ourselves what we really value in education that we want AI to enable: Perhaps teaching critical thinking, enabling agency, and creating a sense of social belonging and civic responsibility?

To anticipate critique, we agree that there will be no global consensus on what education is for, or on the underlying essence of any particular institution, at root because different communities and societies have distinct values and visions. But that’s okay: Let’s empower communities to fit AI systems to local societal contexts; for example, algorithms like constitutional AI enable creating different constitutions that embody flourishing for different communities . This kind of cheap flexibility is an exciting affordance, meaning we no longer must sacrifice nuance and context-sensitivity for scalability and efficiency, a bitter pill technology often pushes us to swallow.

And while of course we have always wanted education to create critical thinkers, our past metrics (like standardized tests) have been so coarse that scoring high is easily gamed without critical thinking. But generative AI enables new affordances here, too: just as a teacher can socratically question a student to evaluate their independent thought, advances in generative AI open up the door for similarly qualitative and interactive measures, like personalized AI tutors that meaningfully gauge critical thinking.

We hope to tow a delicate line beyond broken dichotomies, whether between naive optimism and pessimism, or idealism and cynicism. Change is coming, and we must channel it towards refined visions of what we want, which is a profound opportunity, rather than to assume that by default technology will deliver us (or doom us), or that we will be able to wholly resist the transformation it brings (or are entirely helpless against it). For example, we must temper naive optimism (“AI will save the world if only we deploy it everywhere!”) by integrating lessons from the long line of work that studies the social drivers and consequences of technology, often from a critical angle. But neither should cynical concerns so paralyze us that we remain only as critics on the sidelines.

The case so far is that we need positive visions for society with capable AI, grounded in individual and societal wellbeing. But what concrete work can actually support this? We propose the following break-down:

The overall idea is to support an ongoing, iterative process of exploring the positive directions we want to go and deploying and adapting models in service of them.

This point follows closely from the need to explore the positive futures we want with AI. What directions of work and research can help us to clarify where is possible to go, and is worth going to, in the age of AI?

For starters, it’s more important now than ever to have productive and grounded discussions about questions like: What makes us human? How do we want to live? What do we want the future to feel like? What values are important to us? What do we want to retain as AI transformations sweep through society? Rather than being centered on the machine learning community, this should be an interdisciplinary, international effort, spanning psychology, philosophy, political science, art, economics, sociology, and neuroscience (and many other fields!), and bridging diverse intra- and international cultures.

Of course, it’s easy to call for such a dialogue, but the real question is how such interdisciplinary discussions can be convened in a meaningful, grounded, and action-guiding way — rather than leading only to cross-field squabbles or agreeable but vacuous aspirations. Perhaps through participatory design that pairs citizens with disciplinary experts to explore these questions, with machine learning experts mainly serving to ground technological plausibility. Perhaps AI itself could be of service: For example, research in AI-driven deliberative democracy and plurality may help involve more people in navigating these questions; as might research into meaning alignment , by helping us describe and aggregate what is meaningful and worth preserving to us. It’s important here to look beyond cynicism or idealism ( suggestive of meta-modern political philosophy ): Yes, mapping exciting positive futures is not a cure-all, as there are powerful societal forces, like regulatory capture, institutional momentum, and the profit motive, that resist their realization, and yet, societal movements all have to start somewhere, and some really do succeed .

Beyond visions for big-picture questions about the future, much work is needed to understand where we want to go in narrower contexts. For example, while it might at first seem trivial, how can we reimagine online dating with capable AI, given that healthy romantic partnership is such an important individual and societal good? Almost certainly, we will look back at swipe-based apps as misguided means for finding long-term partners. And many of our institutions, small and large, can be re-visioned in this way, from tutoring to academic journals to local newspapers. AI will make possible a much richer set of design possibilities, and we can work to identify which of those possibilities are workable and well-represent the desired essence of an institution’s role in our lives and society.

Finally, continued basic and applied research into the factors that contribute and characterize human wellbeing and societal health also are highly important, as these are what ultimately ground our visions. And as the next section explores, having better measures of such factors can help us to change incentives and work towards our desired futures.

For better and worse, we often navigate through what we measure. We’ve seen this play out before: Measure GDP, and nations orient towards increasing it at great expense. Measure clicks and engagement, and we develop platforms that are terrifyingly adept at keeping people hooked. A natural question is, what prevents us from similarly measuring aspects of wellbeing to guide our development and deployment of AI? And if we do develop wellbeing measures, can we avoid the pitfalls that have derailed other well-intended measures, like GDP or engagement?

One central problem for measurement is that wellbeing is more complex and qualitative than GDP or engagement. Time-on-site is a very straightforwardmeasure of engagement. In contrast, properties relevant to wellbeing, like the felt sense of meaning or the quality of healthy relationships, are difficult to pin down quantitatively, especially from the limited viewpoint of how a user interacts with a particular app.

Wellbeing depends on the broader context of a user’s life in messy ways, meaning it’s harder to isolate how any small intervention impacts it. And so, wellbeing measures are more expensive and less standardized to apply, end up less measured, and less guide our development of technology. However, foundation models are beginning to have the exciting ability to work with qualitative aspects of wellbeing. For example, present-day language models can (with caveats) infer emotions from user messages and detect conflict; or conduct qualitative interviews with users about its impact on their experience.

So one promising direction of research, though not easy, is to explore how foundation models themselves can be applied to more reliably measure facets of individual and societal wellbeing, and ideally, help to identify how AI products and services are impacting that wellbeing. The mechanisms of impact are two-fold: One, companies may currently lack means of measuring wellbeing even though all-things-equal they want their products to help humans; two, where the profit motive conflicts with encouraging wellbeing, if a product’s impact can be externally audited and published, it can help hold the company to account by consumers and regulators, shifting corporate incentives towards societal good.

Another powerful way that wellbeing-related measures can have impact is as evaluation benchmarks for foundation models. In machine learning, evaluations are a powerful lever for channeling research effort through competitive pressure. For example, model providers and academics continuously develop new models that perform better and better on benchmarks like TruthfulQA . Once you have legible outcomes, you often spur innovation to improve upon them. We currently have very few benchmarks focused on how AI affects our wellbeing, or how well they can understand our emotions, make wise decisions, or respect our autonomy: We need to develop these benchmarks.

Finally, as mentioned briefly above, metrics can also create accountability and enable regulations. Recent efforts like the Stanford Foundational Model Transparency Index have created public accountability for AI labs, and initiatives like Responsible Scaling Policies are premised on evaluations of model capabilities, as are evaluations by government bodies such as AI safety institutes in both the UK and US . Are there similar metrics and initiatives to encourage accountability around AI’s impact on wellbeing?