Positive Alignment: Artificial Intelligence for Human Flourishing¶
Source: arXiv:2605.10310 \ Authors: Laukkonen, Krier, Bakalar, Chandaria, Kringelbach, Elwood, Ford, Rosas, Bohacek, Franklin, Tomašev, Chan, Rieser, Patel, Levin, Rao (Oxford, Google DeepMind, OpenAI, Anthropic, Stanford, Tufts, UCLA) \ Date: May 2026
TL;DR¶
A paradigm paper arguing that current AI alignment (focused on safety/harm-avoidance or "Negative Alignment") is necessary but fundamentally incomplete. The authors propose Positive Alignment — building AI systems that actively support human and ecological flourishing while remaining safe. They connect flourishing science to actionable ML targets across the model lifecycle and advocate for a polycentric, decentralised governance model to avoid paternalistic top-down value imposition.
Negative vs. Positive Alignment¶
Negative (Safety) Alignment — The Current Paradigm¶
- Goal: Prevent AI from causing harm. Establish a behavioural floor.
- Methods: RLHF, Constitutional AI, red-teaming, safety benchmarks.
- Achievements: Refusal rates >97% on dangerous requests.
- Limitations:
- Floor without ceiling: A model can satisfy all safety constraints while being sycophantic, mediocre, or subtly harmful over extended use.
- Preference–Wellbeing divergence: RLHF optimises for expressed preferences, not deep wellbeing.
- Hidden value system: Safety framing obscures value judgements; tends toward static, monocultural values.
- Does not scale: Enumerating harms becomes intractable as systems become more autonomous.
Positive Alignment — The Emerging Paradigm¶
- Goal: Optimise toward positive attractors — virtues, growth, meaning, wisdom. Establish a behavioural ceiling.
- Analogy: A physician doesn't just prevent disease; they promote health.
- Dynamical systems view: Safety pushes systems away from negative attractors; positive alignment pulls systems toward flourishing regimes.
"Positive Alignment is the development of AI systems that (i) actively support human and ecological flourishing in a pluralistic, polycentric, context-sensitive, and user-authored way while (ii) remaining safe and cooperative."
Technical Lifecycle¶
The paper rejects single-solution fixes, arguing positive alignment must be embedded across every stage:
| Stage | Positive Alignment Approach |
|---|---|
| Data & Filtering | Upsample prosocial discourse, cross-cultural ethics, virtuous interactions |
| Pre-Training | "Alignment pretraining" — bake in prosocial values early to prevent rebound effects |
| Post-Training | Multi-objective reward modelling; adaptive constitutions; longitudinal interaction data |
| Memory & Inference | Track user goals over time; distinguish first-order preferences from second-order desires |
| Agents & Multi-Agent | Optimise for process ethics, cooperative equilibria, de-escalation; avoid zero-sum dynamics |
| Forward-Looking | Epistemic humility architectures; mechanistic interpretability; interface alignment |
Governance¶
The paper advocates for polycentric, decentralised governance — avoiding both top-down global imposition and pure free-market laissez-faire. Different communities should be able to configure AI systems to reflect their conceptions of flourishing within universal safety constraints, using user-authored "flourishing profiles" rather than one-size-fits-all alignment.
Significance¶
This paper represents a shift in the alignment conversation from a purely defensive posture (don't let AI cause harm) to an aspirational one (make AI actively good). The involvement of researchers from across the major labs (OpenAI, DeepMind, Anthropic) and multiple academic institutions suggests the idea is gaining mainstream traction.