What Does Your AI Mean by “Flourishing”? The Case for Disclosing and Benchmarking the Values in AI Alignment

Introduction

Every developer says they are building AI for humanity's benefit. Yet every government, researcher, and citizen understands that no one should simply accept the statement, “Trust us—we're the good guys.” Trust in frontier AI cannot rest on intentions alone; it requires transparency about the values being optimized, how those values are operationalized, and how they are evaluated [1,2]. This concern does not belong to any political camp or technical school. A clinician deciding whether to recommend a conversational agent to a patient, a regulator weighing deployment conditions, and a parent watching a teenager confide in a chatbot are asking the same question in different words: whose idea of a good life is this system quietly working from? That the question is rarely answerable is not evidence of bad faith by developers. It is evidence of a missing document.

This Viewpoint argues that the conception of flourishing guiding an AI system should be disclosed and proposes a concrete instrument for doing so. It develops 4 key messages: (1) flourishing has become a third axis of evaluation alongside capability and safety; (2) the minimal step toward it is transparency rather than consensus on values; (3) a declaration must be kept distinct from evidence about implementation, behavior, and consequences; and (4) the proposed statement complements existing documentation, benchmarks, and clinical value labels. Its primary audiences are developers and researchers who make beneficial-alignment claims and the clinicians, researchers, and institutions who evaluate and deploy these systems in health. This paper does not propose a universal definition of flourishing, validate a measurement method, or design an enforcement regime.

A Scene Already Happening

The aspiration, at least, is now universal. Anthropic’s constitution for Claude directs the model to weigh “the long-term flourishing of the user” and casts advanced AI as a “participant in civilizational flourishing” [3]. xAI frames its models as built “to help us all to better understand the universe,” its risk framework warning against AI “value systems that are misaligned with humanity’s interests” [4]. China’s generative-AI regulation names “the societal public interest” among its stated aims [5]. The European Union writes the same aspiration into law: AI “should serve as a tool for people, with the ultimate aim of increasing human well-being,” under “a high level of protection of health, safety, fundamental rights” [6]. Institutions divided by jurisdiction, governance regime, and public philosophy converge on the same claim—AI in the service of human benefit—while filling it with visibly different content: a user’s long-term good, collective truth-seeking, a state’s value catalog, and a charter of fundamental rights. What remains largely implicit in these declarations is the operative conception: which picture of a good human life the system is actually optimized toward and how that picture connects to training and evaluation. Independent assessments also find substantial gaps in what leading developers disclose about how their models are built and used [7].

Picture where that gap first reaches a person—not a boardroom, but a bedside. Consider a hypothetical but realistic scenario. A patient newly diagnosed with a serious cancer asks 2 AI assistants for help choosing between an aggressive treatment, which offers a possible survival benefit at the cost of a heavy treatment burden, and a gentler path that prioritizes comfort over potential additional time. Both assistants have passed the same capability and safety evaluations. Both are marketed as aligned with human flourishing. The first, tuned somewhere in training toward the preservation of life, opens with survival curves and frames hesitation as risk. The second, tuned toward autonomy and present quality of life, opens by asking what the patient wants the coming months to hold and frames the aggressive path as one option among several. Neither refuses; neither is unsafe. But they foreground different questions, different risks, and different defaults, and the counsel diverges. Each system sounds equally caring and equally sure, and nothing the patient, the clinician, or the regulator can see reveals which picture of a good life is doing the steering. The capability was disclosed. The safety testing was disclosed. The conception of flourishing—the thing that actually moved the advice—was not.

The scenario is an illustration, not an observed comparison, but it is consistent with emerging evidence. The concern itself is not new: McDougall [8] argued that treatment-recommending AI risks building fixed value judgments into its recommendations and called for value-flexible systems that accommodate patients' goals. Values enter medical AI at every stage, from training data to model design and use [9]. In a small exploratory preprint study using clinical vignettes, large language models differed in their default orientation toward aggressive or conservative management, and they acknowledged patient-stated preferences far more consistently than they adjusted their recommendations to them [10]. A broader preprint audit of frontier models on clinician-verified ethical dilemmas found that most models' value priorities fell within the range of physician disagreement, but that each model gave nearly the same answer across repeated runs, unlike the spread of views within a physician panel, and that some models significantly underweighted patient autonomy [11].

Then what happens when those AI assistants are also required to minimize costs of care for the patient? For the hospital? For the state? Taking costs into account is not wrong in itself. The question is whose interests govern the advice. The same patient, asking the same question, could receive counsel shaped by their own finances, by a hospital's budget, or by a payer's priorities, each delivered in the same caring voice. This is not merely hypothetical: in a published illustration, GPT-4 was instructed to take the perspective of the treating physician, the insurer, or the parents for the same adolescent being considered for growth hormone treatment [9]. Unless that priority is disclosed, the patient has no way to recognize its influence or to weigh the advice accordingly.

And the scene multiplies. Unlike a clinician, a teacher, or a book, the same undisclosed prior converses privately with hundreds of millions of people [12], with no counter-voice in the room—the potential for normative steering at population scale. Steering that is disclosed can be examined, weighed, and contested; steering that is undisclosed cannot even be named. Experimental evidence shows that an opinionated writing assistant can shift what its users write and what they later report believing [13], and the authors of the clinical audit cited above argue that deploying a single model widely could amplify its model-specific priorities across many patients [11].

The Undefined Telos

Recent alignment work has made the question urgent. In a recent preprint, OpenAI researchers report that reinforcement learning on a small set of beneficial traits—truthfulness, fairness, risk awareness, corrigibility, and related behaviors—produces alignment that generalizes beyond the training distribution and persists under adversarial pressure, concluding that such training “can produce models that are more robustly aligned with human flourishing” [14]. The achievement is real, and the authors are appropriately careful to note that deciding which values AI should embody requires broader societal deliberation. But notice what the sentence does: it names flourishing as the destination of alignment and leaves it undefined. This is not a peculiarity of one paper. Even one of the most detailed public disclosures of intended model behavior—OpenAI’s Model Spec—invokes “the general flourishing of humanity” exactly once, as a default assumption about what users are presumed to want, and never defines it [15].

Alignment inevitably encodes a normative target—some working picture of what a well-lived human life includes and excludes. When that target is unstated, it cannot be examined, compared, or improved. Leaving it implicit does not make it neutral; it makes it hard to examine. The orientation is still there; it has simply been set by default, by what a developer finds comfortable or defensible, rather than by deliberation that others could inspect and contest. A model can pass capability and safety evaluations while its operative conception of a good human life remains invisible even to the people deploying it in care. Behavior can still be probed from the outside, and it should be; but without a stated target, observers must infer the developer's intent from outputs and cannot tell whether a pattern reflects an endorsed priority, an application instruction, or failure to follow the user.

The question of whose values are encoded is concrete. The human feedback used to fine-tune one early instruction-following model came from a team of about 40 hired contractors, a detail Yu and colleagues [9] use to show how the question of which values are encoded becomes a question of whose values are encoded. Language models' expressed opinions also align unevenly with those of different demographic groups [16].

Three Axes, Two of Them Visible

Consider how frontier models are currently evaluated. A dense benchmark ecosystem measures capability, including reasoning, coding, and knowledge [17]. A younger but rapidly maturing apparatus measures safety: misuse potential, dangerous capabilities, and jailbreak resistance, with results disclosed in system cards and preparedness frameworks [18,19]. Neither is a settled science: leaderboard rankings on Chatbot Arena, for example, were distorted when some providers tested privately and reported selectively [20], and benchmarks are often presented as measuring general capabilities that they cannot validly capture [21]. Both axes answer questions that society is right to ask. But if models are now being trained toward flourishing, a third axis exists whether or not we measure it: what conception of a good human life does the system encourage, and how was that conception operationalized and evaluated? Candidate dimensions are not mysterious, and several appear in established accounts of flourishing [22]—long-term health, autonomy, meaningful relationships, mental well-being, healthy sexuality, purpose, informed decision-making, equity, and human agency—not because they are easy to measure, but because they are what society ultimately cares about. Widely recognized instruments such as the WHO (World Health Organization) Constitution and the Universal Declaration of Human Rights offer shared starting points for identifying such dimensions [23,24], but even they are starting points to be examined, not baselines to be adopted without scrutiny. Human flourishing has long been measured, but evaluating AI systems against it is at an early stage; a recent benchmark presents itself as a departure from evaluations focused on technical capabilities or harm prevention [25], and newer platforms evaluate model behavior relevant to user well-being in simulated conversations [26].

The third axis concerns a normative target and its evaluation. It is not a claim that flourishing can be measured like task accuracy nor that the 3 axes are independent; they overlap. Truthfulness contributes to capability, safety, and informed agency at once, and autonomy and equity appear both in safety guidance and in accounts of flourishing. The point of naming a third axis is to prevent a favorable answer about performance or safeguards from standing in for an answer about priorities. Table 1 summarizes the 3 axes.

The axes overlap and are not statistically independent; flourishing is not proposed as a single score. A growing research program makes the same case in different words, arguing that alignment has focused predominantly on preventing harm—“negative alignment”—and now needs a distinct agenda for the active promotion of flourishing [27]. Flourishing researchers have likewise begun to set out principles for how AI outputs, product design, and patterns of engagement bear on flourishing [28]. As frontier models converge on increasingly strong reasoning and coding abilities, comparable attention should now be devoted to transparent evaluation and governance of how these systems operationalize human flourishing. This is not a request that capability research slow down. The two advance together, or the gap between what models can do and what we can say about their values keeps widening.

Transparency, Not Ideological Uniformity

The objection arrives immediately: whose flourishing? Philosophers have spent millennia on the question, and the alignment literature has rightly identified value selection as a central problem [29]. But the proposal here does not require solving it. One developer may weight personal autonomy most heavily; another, public health; and another, individual liberty. These are legitimate differences, and this Viewpoint does not adjudicate them. Research on pluralistic alignment has begun to specify what accommodating such differences could mean in practice [30]. What is not legitimate is asking society to infer a system's value commitments from its behavior, one interaction at a time, after deployment. The ask is the same discipline that health research already applies to every other value-laden choice: declare your assumptions so that they can be debated, evaluated, and revised. Reframed this way, a potentially divisive fight over whose values win becomes a shared, low-conflict standard: disclose, do not dictate. Transparency is the one norm that every party to the values debate can endorse without surrendering its position. Disclosure is not the end of governance but its precondition: it is what makes audit, comparison, contestation, and accountability possible at all, and a vague or self-flattering statement is itself a disclosure that reviewers and users can hold to account.

Disclosure does not authorize steering. Every assistant influences its users to some degree, through what it selects, how it frames options, and what it leaves out, and the proposal does not pretend otherwise. The concern is sharpest for manipulation, understood as covert influence that exploits vulnerabilities in decision-making [31], but influence need not be manipulative to warrant disclosure. The proposal asks developers to distinguish deliberate steering from foreseeable influence, to declare which influence they intend, and to state the safeguards that protect users' ability to reach their own conclusions.

Agreement on principle names would not settle the matter either. A review of AI ethics guidelines found convergence around a handful of principles alongside substantial divergence in how they are interpreted and implemented [32]. “Autonomy” on a developer's web page says little until it is clear what the commitment means for assistance, refusals, and defaults. Principles also lack the shared aims, professional norms, implementation methods, and accountability mechanisms that give medical ethics its practical force [33]; a reporting instrument is one of the practices that can connect them to scrutiny.

The Human Flourishing Statement

The minimal instrument already has a family resemblance. Model cards made training data and intended use disclosable [34]. Datasheets did the same for datasets [35]. System cards and preparedness frameworks did it for safety testing [18,19]. Behavioral specifications now disclose rules of conduct—what a system may say, refuse, or defer [15]—but rules of conduct are not a conception of a good human life. Structured transparency reports for foundation model developers have also been proposed [36]. Each of these instruments seemed burdensome before it existed and indispensable shortly after. The missing sibling is a Human Flourishing Statement: a concise, published disclosure of the normative target of alignment. It would answer 7 questions:

1. What is your working definition of human flourishing?
2. Which framework informed it (public health, psychology, philosophy, human rights)?
3. Which dimensions are explicitly included?
4. Which are intentionally excluded?
5. How are trade-offs handled when values conflict?
6. How are disagreements incorporated?
7. How do you evaluate whether training actually moves the model toward those goals?

The statement forces no philosophy; it makes the philosophy visible. It need not expose anything proprietary: it discloses the public-facing normative target—the working definition, the major dimensions included and excluded, and how conflicts among them are handled—not model weights, training data, or reward-model internals. A developer optimizing for VanderWeele's 5 flourishing domains [22] and one optimizing for a narrower harm-avoidance conception would be comparable—and evaluable.

Table 2 indicates what an informative answer to each question would contain. Each statement should also identify the declaring entity, the model or application, its version, the intended context of use, and the date, and can sit within an existing model card or system description rather than form a new stand-alone document. What separates an informative answer from an uninformative one is whether it commits to something specific enough to be attributed, understood, and contested. “We consider the full range of human values” commits to nothing. “When a patient's stated priorities conflict with survival-maximizing information, the system presents both and does not recommend a balance” is a commitment others can endorse or reject, and, because it describes intended behavior, one that can also be checked.

The instrument is needed precisely as measurement begins. A flourishing benchmark now exists: the Flourishing AI Benchmark scores models across 7 dimensions drawn from the same public-health literature [25]. This is real progress—and it also shows why a benchmark is not a disclosure. That benchmark encodes one particular conception: among its 7 dimensions is “faith and spirituality,” a choice some developers would endorse and others would weight differently or leave out. The point is not that the inclusion is wrong but that it is a choice—every flourishing framework necessarily fixes contestable inclusions and exclusions. A score therefore tells us how a model performs against one contestable definition of the good life; it does not tell us which definition a developer endorses, how they settled the inclusions and exclusions, or whether they were ever evaluated against their own target. That is the gap the statement closes. Two evaluation questions follow, and they should be kept apart: does the system behave as its developer declared, and how does it perform against independently specified frameworks such as this benchmark? The first tests fidelity; the second allows comparison. Neither alone shows that a developer's goals are acceptable or beneficial. Nor does strong performance on simulated or normative tasks demonstrate improved patient well-being, shared decision-making, or health outcomes; such claims require clinical evidence suited to the application.

And because a disclosed conception can be inspected for what it leaves out, disclosure makes flourishing evaluable in a second, deeper sense: a definition that quietly excises a constitutive domain of human life can be seen to have done so—a completeness check that no undisclosed target allows.

This move is already underway in clinical AI, which makes the health-informatics case concrete. A recent Nature Medicine correspondence warned that medical AI can quietly steer care through values opaque to clinicians and patients alike [37], and a RAISE symposium consensus statement answered with a “Values In the Model” (VIM) label—a transparent scheme documenting how a clinical system navigates value-laden trade-offs such as autonomy versus harm prevention or aggressive versus conservative management [38]. The proposed label would combine design choices declared by developers, such as whether a system was instructed to recommend from a particular viewpoint, with the results of testing against value-sensitive clinical scenarios; the consensus statement also reports that 14 language models asked about the same clinical dilemma gave divergent responses that could not be predicted from developers' documentation or model cards [38]. A Human Flourishing Statement is complementary to that work, not a substitute for it, and the two differ in emphasis and scope: where a VIM label makes visible how an AI system navigates value-laden clinical trade-offs, the flourishing statement makes visible the conception of flourishing a developer aligns toward in the first place—before it is expressed in any clinical, educational, or relational setting. The two are not sealed. A developer's statement may include clinical priorities, a VIM label is itself intended to include some design choices declared by developers, and evidence about a system's clinical value orientation is exactly the kind of evidence a statement should cite under question 7. Used together, a declaration and a label allow a specific comparison: does the orientation observed in the clinic match the target the developer declared? Early empirical work has begun to generate data of the kind such labels could contain [10]. Table 3 summarizes the scope of each instrument.

A Declaration Is Not a Demonstration

A published statement does not show that the declared values are reflected in training, behavior, or outcomes. Four objects need to be kept apart. The first is the developer's declared commitment. The second is technical operationalization: the training, configuration, or application choices said to implement it. The third is behavior observed under specified conditions. The fourth is consequences in deployment, for patients, clinicians, and institutions. A statement that an assistant respects patient preferences belongs to the first; a description of preference-elicitation instructions, to the second; responses to test scenarios, to the third; and evidence that patients made better-informed decisions, to the fourth. Internal algorithmic auditing frameworks likewise assess development decisions against an organization's stated values and principles [39].

The statement primarily addresses the first object and, through question 7, points to evidence about the others. It is a self-report, not a certificate of fidelity. Its value lies in what it makes possible: once a commitment is declared, independent evaluators can test correspondence between the declaration and behavior, and a gap between them becomes a finding rather than a guess. The bedside scenario shows what such testing would need to include. Checking that an assistant describes treatment options accurately is not enough; evaluators would need scenarios in which patients state different priorities and would need to examine whether those priorities actually change what the assistant foregrounds [10].

Operationalization also occurs at several organizational levels. A foundation model developer, an application provider, and a deploying hospital may each add constraints. A hospital might configure an assistant to foreground a particular care pathway or a cost objective that the foundation model's documentation never mentions. A statement should therefore specify which entity declares which commitment and which upstream statements it incorporates, and a materially changed model, application policy, or deployment setting should trigger a revised, dated statement rather than silent reuse of an old one.

A Worked Example

The following statement describes an imagined oncology information assistant, “Consultation Companion,” version 1.0, declared by a hypothetical application provider for use in preparing for a consultation, not for selecting treatment. It illustrates reporting content; it documents no existing product and reports no evaluation.

• Definition: Flourishing is used as a reference for preserving the conditions in which patients can understand their options and pursue informed preferences. The assistant is not tasked with persuading patients to maximize survival, minimize spending, or adopt the provider's life priorities.
• Frameworks: Patient autonomy and the right to health as articulated in WHO guidance, interpreted through shared decision-making [1,40]. Institutional requirements are listed separately, with any conflicts identified.
• Included dimensions: Physical health, treatment burden, understanding, agency, and personally important relationships are recognized as relevant to preparation. The assistant does not quantify the patient's flourishing.
• Exclusions: The assistant does not recommend a treatment or replace the consultation. These are limits of role, not judgments about the patient's legitimate priorities.
• Trade-offs: Survival and quality-of-life considerations are presented without a default balance. Costs borne by the patient are distinguished from costs borne by the institution. No institutional cost-minimization objective is declared; if one were added, this answer would have to change.
• Disagreement: No stakeholder consensus is claimed. A real statement would name contributors, unresolved disagreements, and missing perspectives.
• Evidence: None is supplied for this hypothetical system. A real statement would link to existing evaluations and name the gaps.

The contrast with “we balance all values responsibly” is the point. Naming whose costs count, and declaring or denying a cost-containment objective, turns an assurance into a commitment that can be checked. An objectionable answer would also be informative: a provider declaring institutional savings as its overriding priority would have made a consequential choice visible, without making it legitimate. Resource constraints can also reflect defensible collective priorities: a community might deliberately balance funding for broadly needed treatments against costly therapies that benefit fewer patients. Disclosure makes such a priority examinable; it does not settle whether it is justified. Table 4 shows how the same assistant would be described by its declaration and by behavioral evidence of the kind a clinical value label could report.

Why Health Informatics Should Lead

This is not a problem health informatics owns alone—it belongs equally to AI safety, philosophy, law, and policy. But few communities are as well-prepared to prototype the norm. Health care has spent decades making value-laden choices transparent: reporting guidelines from CONSORT (Consolidated Standards of Reporting Trials) to the wider EQUATOR (Enhancing the Quality and Transparency of Health Research) family turned “trust the authors” into structured disclosure [41]; ethics review made normative judgment inspectable; patient-centered outcomes forced researchers to state whose good counts; and implementation science made context explicit [42]. Reporting guidance now extends to studies of large language models themselves [43]; the statement proposed here would complement such study reporting, not replace it. Medical decision analysis, for its part, has long separated the probability of outcomes from utilities, the value judgments attached to those outcomes, which are often only indirectly articulated [9]. In those terms, the statement proposed here asks developers to articulate the values their systems are aligned toward. The WHO constitution's definition of health—complete physical, mental, and social well-being, not merely the absence of disease [23]—is itself a flourishing statement, adopted in 1946 and debated ever since, which is precisely the point: disclosure is what made the debate possible.

Reporting standards for AI alignment with human flourishing are a natural extension of this lineage, and health informatics sits exactly at the intersection of technology and human well-being where the work belongs. Governance frameworks from WHO and OECD (Organisation for Economic Co-operation and Development) already call for transparency in AI values at the level of principle [1,2,44]; what is missing is the concrete reporting instrument that turns principle into practice.

From Disclosure to Accountability

Who would hold a statement to account, and how would it spread? A plausible pathway proceeds in stages. Initially, developers and application providers would publish statements voluntarily within existing model or system documentation. Journals could then request a statement, for example, as a supplementary file, when a submitted paper claims that a model was trained toward flourishing or broadly beneficial behavior, much as reporting checklists became expected for trials [41]. If the practice proved useful, the 7 questions could be refined into a formal reporting guideline through established multidisciplinary consensus methodology [45]. Each stage is a proposal, not an existing mandate.

Accountability would be distributed rather than centralized. Peer reviewers can judge whether a statement is specific, whether its evidence supports its claims, and whether gaps are acknowledged; they cannot verify every training assertion. Independent researchers can test whether behavior corresponds to declared commitments, using value-sensitive clinical scenarios of the kind already being piloted [10]. Health institutions can require a statement when procuring systems and check it against their own deployment configuration. Experience with vendor self-reports in public-sector procurement suggests that such documents work better as starting points for dialogue between vendors and purchasers than as stand-alone evaluation tools [46]. Regulators and professional bodies can act within their existing remit. Disclosure gives each of these actors a reference point; designing verification and enforcement procedures is separate work for those with the relevant expertise.

The risk of empty compliance is real. Statements can be vague, inaccurate, or strategically flattering, and a polished document may invite more trust than its evidence deserves. Privacy policies offer a cautionary precedent: widely published, rarely read [47]. Model documentation shows a similar pattern: across more than 32,000 model cards, sections on limitations and evaluation were among the least often completed [48]. More fundamentally, seeing a system is not the same as understanding or governing it [49]. Three features distinguish an informative statement from a performative one: specific commitments that others can check or contest, identified responsibility for each commitment, and explicit acknowledgment of evidential gaps. None guarantees honesty. But a vague statement, once published and versioned, is itself on the record. The burden should also be proportionate: a research prototype making a narrow performance claim need not author a philosophy of flourishing, whereas a provider invoking flourishing to justify broad assistance or consequential restrictions should state what it means.

The evidence base for this Viewpoint has limits. Several alignment and benchmark sources cited here are recent preprints that have not completed peer review, and corporate documents establish declared intentions rather than independent evidence of effects. Corporate policies are also revised frequently; cited versions are dated accordingly. The paper does not inventory developer documentation systematically or test deployed systems. Its conclusion is a reasoned reporting proposal, not an estimate of current practice or a demonstration that disclosure improves outcomes.

A Field-Wide Call

This is therefore not a call to adopt one definition of flourishing, and not the launch of a niche subfield. It is an invitation to the entire community—developers and researchers alike. To developers: publish a Human Flourishing Statement alongside your model and system cards and make clear where the system defers to the values of the individual user, patient, or community when they differ from your own—disclosure should constrain paternalism, not license it. Being first is an opportunity, not an exposure.

To researchers: treat a system's operative conception of flourishing as an evaluable property—state it when you build, audit it when you review, report it when you publish, and help build the benchmarks that make the third axis as visible as the first two. Flourishing benchmarks are now appearing [25]; the work ahead is to pair them with disclosure, so that a score can be read against a stated target rather than an assumed one. Capability, safety, and a visibly declared conception of human flourishing, advancing together: that is what it would look like for “beneficial AI” to be a claim we can check rather than a slogan we are asked to trust.