Most AI governance frameworks focus on preventing harm. You document risks, set thresholds, monitor for drift, and escalate when things go wrong. It's defensive by design, and that defensiveness shows up in how we talk about beneficial AI: vague commitments to "human-centric" systems that sound good in a policy document but mean nothing when you're writing acceptance criteria.
Meanwhile, wellbeing science, a multidisciplinary field spanning positive psychology, economics, and philosophy, offers concrete frameworks for what makes human lives actually go well. Yet governance teams rarely engage with it. Why? Because several persistent myths make the integration seem impractical, unmeasurable, or outside the scope of model risk management.
Let's address them directly.
Myth 1: "We can't define wellbeing, so we can't govern for it"
Reality: You don't need philosophical consensus to operationalize wellbeing metrics.
Yes, philosophers debate whether wellbeing comes from pleasure, desire satisfaction, or living by your values. Economists measure it differently than psychologists. There's no single theory everyone agrees on.
But that hasn't stopped humanity from building institutions around wellbeing for centuries. Education systems, healthcare infrastructure, labor protections, all exist without a unified theory of human flourishing.
The wellbeing literature provides workable frameworks you can actually measure. Martin Seligman's PERMA model breaks wellbeing into positive emotions, engagement, relationships, meaning, and achievement. Self-determination theory focuses on autonomy, competence, and relatedness. These aren't perfect maps of human experience, but they're measurable proxies that beat "don't be evil" as a design constraint.
Your AI Management System can incorporate wellbeing indicators the same way it incorporates fairness metrics: imperfectly, iteratively, with acknowledgment that the measure isn't the territory.
Myth 2: "Wellbeing is a product feature, not a governance concern"
Reality: If your AI system affects how people spend their time, it's affecting their wellbeing, and that's a governance issue.
Consider what's already happened. Social media launched twenty years ago. The platforms optimized for engagement. Now we have measurable declines in reported loneliness, trust in institutions, and shared sense of reality. Those aren't product failures; they're governance failures. The systems did exactly what their objective functions specified.
Your model inventory likely includes recommendation engines, content moderation systems, or decision aids that shape user behavior. If you're only governing for accuracy and fairness, you're missing the second-order effects: Does this system help users pursue meaningful goals, or does it optimize for compulsive checking? Does it support relationship quality or substitute for it?
ISO/IEC 42001's Annex A controls require you to identify AI system impacts. Wellbeing impacts, on attention, autonomy, social connection, belong in that analysis alongside discrimination risks and data privacy concerns.
Myth 3: "We can't measure wellbeing outcomes from AI systems"
Reality: You can measure wellbeing proxies the same way you measure model performance.
You already track precision, recall, and calibration. You monitor for drift. You collect user feedback on model outputs. Wellbeing measurement isn't fundamentally different, it just requires different instruments.
Subjective wellbeing research has developed validated survey instruments for life satisfaction, emotional experience, and psychological needs. You can deploy these in user research, just as you'd deploy usability testing. You can track behavioral proxies: time spent on meaningful tasks versus addictive loops, depth of social interactions, progression toward stated goals.
The challenge isn't measurability; it's deciding what to measure and building feedback loops between those measures and your model development process. That's a governance design problem, not a scientific impossibility.
Myth 4: "Optimizing for wellbeing means building 'nanny AI' that decides what's good for users"
Reality: Wellbeing-informed design supports user autonomy; it doesn't replace it.
The concern here is real: you don't want your AI system to override user preferences based on some paternalistic view of their "true" interests. But wellbeing science actually centers autonomy as a core component of flourishing.
The distinction is between supporting informed choice and manipulating behavior. A wellbeing-informed recommendation system might surface content that aligns with a user's stated long-term goals, not just their moment-to-moment impulses. It might make attention costs transparent. It might help users reflect on whether their usage patterns serve them.
This is different from dark patterns that exploit cognitive biases or engagement optimization that treats user time as an extractable resource. Your Technical Documentation (Annex IV) should specify how the system respects user agency while supporting their wellbeing. If you can't articulate that distinction, you've found a design gap.
Myth 5: "This is too abstract for risk tiering and model validation"
Reality: Wellbeing impacts map directly to existing risk frameworks.
The NIST AI RMF already asks you to map impacts on individuals and society. The EU AI Act's high-risk categories include systems that affect access to education, employment, and essential services, all of which directly influence wellbeing. SR 11-7 requires you to assess materiality, and wellbeing degradation at scale is material.
When you're conducting an AI System Impact Assessment, you can evaluate wellbeing effects alongside other impact categories: Does this system support or undermine users' sense of competence? Does it facilitate meaningful relationships or displace them? Does it help users pursue their values or distract them?
These questions inform risk tiering. A model that shapes how millions of people spend their attention deserves more rigorous validation than one with narrow technical scope, even if both have similar accuracy requirements.
What to do instead
Start by acknowledging that your AI systems already affect wellbeing, you're just not measuring it. Then build it into your governance process:
In requirements definition: Specify wellbeing outcomes you're designing for, not just harms you're avoiding. Make them concrete enough to test.
In impact assessment: Add wellbeing indicators to your ISO/IEC 42005 methodology. Use validated instruments where they exist; develop proxies where they don't.
In model validation: Evaluate whether deployed systems support the wellbeing outcomes specified in requirements. Treat misalignment as a validation finding.
In Post-Market Monitoring: Track wellbeing proxies over time. If you see degradation, that's drift, even if technical performance metrics remain stable.
The science of wellbeing gives you a practical foundation for what "beneficial AI" actually means. You don't need to solve philosophy first. You need to decide that human flourishing is a first-class concern in your AI Management System, then build the measurement and feedback loops to support it.
