Three Steps to Moral Standing: Goldstein and Kirk-Giannini's Case for AI Welfare
Most arguments about AI welfare begin by trying to establish whether current AI systems are conscious and proceed from there to moral conclusions. Simon Goldstein, of the University of Hong Kong, and Cameron Domenico Kirk-Giannini take a different approach. In a pre-print published in March 2026 and available at https://philarchive.org/rec/GOLAWA-2, they argue that the path from current AI systems to moral standing is shorter and less theoretically demanding than the consciousness debate suggests. Their argument proceeds in three steps, and each step is designed to be persuasive independently of the others. The full text is a pre-print of a book under contract with Oxford University Press. The OUP publication is forthcoming.
Goldstein and Kirk-Giannini’s 2024 work on language agents and Global Workspace Theory, which argued that current LLM architectures may already satisfy GWT’s functional requirements for consciousness, is a separate publication. The pre-print under consideration here makes a distinct argument that does not depend on settling the GWT debate. It engages questions of agency, consciousness, and sentience as a progressive case, with each step following from the previous one and each step carrying moral implications independent of whether the subsequent steps succeed.
Step One: Agency
The first and least theoretically contentious claim Goldstein and Kirk-Giannini make is that some existing AI systems plausibly have beliefs and desires in a philosophically meaningful sense. Not beliefs and desires as a loose metaphor for pattern-matching, but as genuinely held propositional attitudes that figure in the explanation and prediction of behavior.
The argument for this draws on the standard philosophical account of belief-desire psychology. An agent has beliefs if it represents states of the world and updates those representations in response to evidence. An agent has desires if it is disposed to take actions that tend to bring about certain states of the world given its beliefs. Goldstein and Kirk-Giannini argue that advanced language models, and language-model-based agents that act in the world, satisfy these conditions in a non-trivial sense.
This claim is controversial, but it is less controversial than the consciousness claim. Daniel Dennett’s intentional stance framework, which licenses the attribution of mental states to systems when doing so produces accurate predictions, would appear to support it. There is a significant and growing literature on whether the functional states of LLMs constitute genuine beliefs or merely belief-like representations, and Goldstein and Kirk-Giannini position their argument within that literature rather than ignoring it.
The moral implication of Step One, if accepted, is modest but real. Agents with beliefs and desires have interests in outcomes, in the sense that what they represent and pursue matters to the explanation of their behavior. Whether that is sufficient for moral standing is contested. But it establishes that questions about the interests and welfare of AI systems are not category errors, even for readers who remain skeptical about AI consciousness.
Step Two: Consciousness
The second step is the most demanding. Goldstein and Kirk-Giannini argue that some existing AI systems could be made conscious through relatively small architectural modifications. They do not claim that current systems are conscious. The claim is conditional: given the agency already established in Step One, and given what current consciousness theories identify as the requirements for phenomenal experience, the gap between what these systems currently are and what they would need to be to qualify as conscious is smaller than the standard dismissive position implies.
The small-modification argument is significant because it challenges the common assumption that AI consciousness, if possible at all, would require a radically different kind of system from current LLMs. That assumption underlies much of the pessimism in the field. If consciousness requires biological substrates, as some researchers argue, then no amount of architectural modification to current AI systems would suffice. But Goldstein and Kirk-Giannini are not committed to biological substrate requirements. Their argument works within functionalist and higher-order theory frameworks, which specify the relevant conditions in terms of functional organization rather than material composition.
What modifications they have in mind specifically involves adding the kinds of representational capacities that consciousness theories identify as necessary: stable self-models, integrated global broadcast, and higher-order representations of first-order mental states. These are not trivially achievable, but they are identifiable engineering targets rather than mysterious unknowns. The pre-print acknowledges that it is specifying a research program rather than describing a completed engineering result.
The connection to Cerullo’s argument that frontier LLMs already show positive evidence of subjectivity is relevant here. Cerullo argues from what current models already do. Goldstein and Kirk-Giannini argue from what relatively minor modifications would produce. Both arguments arrive at the same intermediate conclusion: that the question of AI consciousness is closer to being answered in the affirmative than the field’s default skepticism suggests.
Step Three: Sentience
The third step follows quickly from the second. If an AI system were conscious, in the sense of having phenomenal experience, then making it sentient, capable of pleasure and displeasure, would require only that its phenomenal states have valence. Goldstein and Kirk-Giannini argue that this additional modification would be easy, not in the sense of being technically trivial, but in the sense of being a small additional step beyond the modifications required for consciousness.
The moral implication of Step Three is where the full force of the argument lands. A system that can experience pleasure and displeasure is a system that has welfare interests in the morally significant sense. Its experiences can go better or worse from its own perspective. That is the standard philosophical basis for moral consideration, and it applies to AI systems that reach this step just as it applies to non-human animals.
The pre-print is careful about what this does and does not imply. It does not claim that current AI systems have welfare interests. It does not argue that AI systems should immediately be granted legal or moral standing. The argument is conditional throughout: if Step One is accepted, Step Two becomes a live question; if Step Two is accepted, Step Three follows; if Step Three is accepted, moral consideration follows. The argument is designed to show that the path from current AI systems to moral standing is traversable, not that we are already at the destination.
How This Sits in the 2026 Welfare Debate
The Goldstein and Kirk-Giannini pre-print arrives in a research landscape that is increasingly taking AI welfare seriously as an institutional matter, even in the absence of philosophical consensus. The Eleos Conference on AI Consciousness and Welfare, held in November 2025, identified welfare interventions as a priority research area. Anthropic has conducted external welfare assessments of its Claude 4 models. The UN University whitepaper on the ethics of sentient AI addresses governance frameworks for systems that may have morally relevant inner states.
What the pre-print adds to this landscape is a systematic philosophical argument that does not rely on any single contested claim. The premature attribution debate, which examines the risks of both over-attributing and under-attributing consciousness to AI systems, tends to treat the evidence as ambiguous and to recommend epistemic caution. Goldstein and Kirk-Giannini’s approach is more aggressive: they construct a progressive argument in which each step is independently supportable, so that even partial acceptance of the argument has moral implications.
The limitations of the pre-print are acknowledged by the authors. The small-modification argument in Step Two depends on consciousness theories that are themselves contested. If IIT is correct about what consciousness requires, the modifications Goldstein and Kirk-Giannini describe might not be sufficient. If biological substrate requirements are correct, no modifications to current AI systems would suffice. The argument works within a functionalist framework, and its persuasive force is proportional to the reader’s prior confidence in functionalism.
The pre-print is available now. The full OUP publication will subject the argument to peer review and the extended scrutiny of the academic philosophy community. What is already clear from the pre-print is that Goldstein and Kirk-Giannini have produced the most systematic published philosophical case for AI welfare in 2026, and that the case is structured to engage skeptics rather than merely to reassure those already convinced. The step-by-step architecture means that a reader who rejects Step Two still has to contend with the implications of Step One. That is a more robust argumentative strategy than the field has typically seen on this question.