My research project asks under what conditions an artificial system could possess moral status: when, if ever, it could matter morally in its own right rather than merely because of its effects on human beings. It examines which capacities could ground such status, whether they could be realised in a non-biological system, and what evidence, design choices and safeguards would matter if such systems became technologically plausible.
The central question
To possess moral status is to be an appropriate object of moral regard – to be the kind of entity whose interests or welfare should matter for its own sake. An entity capable of being benefited or harmed in this morally direct sense is often described as a moral patient.
The project begins by asking what must be true of any entity for it to qualify as a moral patient. Its purpose is to identify the relevant conditions, examine whether artificial systems could satisfy them, and consider how we should respond under uncertainty.
Possible grounds of moral status
The research examines proposed grounds of moral status, including sentience and valenced states, consciousness, embodiment and substrate, agency and autonomy, and welfare. It asks how these concepts are connected, whether any one of them is necessary or sufficient, and whether moral status is best understood as binary, gradual or multidimensional.
Artificial realisation, embodiment and substrate
A central issue is whether morally relevant properties can be realised in a non-biological system. The project considers whether biological substrate is morally significant in itself, or whether what matters is a system’s functional, phenomenal or relational organisation. It also asks whether embodiment, grounding and continuing engagement with an environment are required for moral status to arise.
Identity, persistence and the moral patient
If an AI system had morally significant capacities, what exactly would the moral patient be: the underlying model, a running instance, a particular conversation, a persistent agent or an embodied system? The ease with which artificial systems may be copied, paused, modified, merged or deleted makes questions of identity and persistence central rather than peripheral.
Evidence and epistemic risk
The project also addresses how moral status could be recognised. Behavioural fluency alone may be insufficient, but demanding proof of consciousness could set an impossible standard. The practical problem is therefore how to reason responsibly when both false attribution and false denial of moral status could have serious consequences.
AI safety, welfare and governance
Much work in AI safety understandably focuses on the benefits and harms that increasingly capable AI systems may produce for human beings. AI is already bringing substantial and difficult-to-predict changes to politics and democracy, economics and inequality, education and employment, warfare and security, mental health and social relations.
If future systems acquire morally significant capacities, however, AI safety may also need to consider harms that human beings could cause to those systems. This possibility has implications for model evaluation, training methods, deployment, shutdown, copying, modification and governance. Clarifying the basis of moral status in advance could therefore inform both technical research and public policy.
Human–AI systems
A related strand asks whether, and in what sense, human agency, cognition and identity can be extended through AI and other tools. If a person’s planning, memory or decision-making becomes deeply integrated with an artificial system, we may need to consider whether the relevant object of evaluation is the unaided human, the tool, or a hybrid human–AI system. This raises further questions about the locus of agency, identity, responsibility and moral standing.
Scope of the claim
This research project does not claim that current large language models are conscious or possess moral status. It asks whether systems satisfying the relevant conditions are technologically plausible, how we could know if those conditions had been met, and why identifying them before such systems exist could help shape responsible research, design and governance.
Research library
[TBC].