Non-paradigmatic agents

Between-Cause considerations Animal Welfare Global Health and Development (part about givewell’s moral weights)

List of non-paradigmatic agents

• C. Thi Nguyen’s work on games and agency suggests playing games involves temporarily adopting different forms of agency, with different ends and means, as specified by the game designers/other players. This suggests we can create sub agents through social interactions.

• Contradictory or unstable preferences grounded in a single time-slice of a single human animal, akrasia might be a mild case, DID or split-brain patients might be a much stronger one, the instability of preferences/traits/dispositions suggested by situationist experiments in social psychology could also be interpreted in a similar way. For at least some of these human animals, it might make more sense to model them (either simultaneously or only diachronically) as somewhere between a single unified agent, a collective agent containing sub-agents as parts, or multiple agents forced to share physical resources. Probably the applicability of all these shade into each other. But when considering the non-hedonic welfare of these agents, I don’t know of many procedures/models for assessing welfare/rationality/ought-to-be-done-ness, (e.g. AFAIK voting procedures, utility functions, bargaining/auction rules, ideal-agent and eudaimonic virtue ethics, rule-consequentialist/Kantian/Rawlsian/Scanlonian norms typically assume binary, stable, and non-overlapping agents.) that can output sensible answers when considering a chunk of spacetime containing something in-between a single agent and multiple agents, or unifications/divisions of agents. But I think most agents are indeterminate in this way, so this is a serious defect. The most promising exceptions to this trend seem to be Confucian/Neo-Confucian approaches, Korsgaards work on role-based ethics, and communitarian philosophers, which accept overlapping, vaguely-bounded, and unstable agents, (e.g. one of my communities, ‘me’ simpliciter, me-qua-son, and my trait of benevolence/proto benevolence.) How should subagents/meta-agents be individuated? Neurological distinctness? Causal or informational relations? Social/moral roles and interactions such as recognition, respect, holding-to-account? Responsiveness to different categories of reasons, (e.g. agent-neutral reasons around harm/benefit → benevolence subagent, agent-relative reasons around danger → courage subagent Choosing for changing selves

• Non-instrumental/core preferences changing via transformative experiences/aspiration, as discussed by Agnes Callard and L. A. Paul. (Most of this seems to focus on the rationality of changing your own preferences/how to balance your current and post-change preferences. Looking backwards after the change also seems interesting to me, I’m not sure how you’re meant to balance satisfying your current/future preferences and previously-held, (but still fulfillable) preferences.) Value Change and AI Safety

• How do the different kinds of agents fit into the moral economy? Which are fitting subjects of moral obligation, rights, blame, praise, punishment, or reward, and why? At an even more basic level, what is the minimal-viable moral subject and agent? Why can’t you only punish my leg, or my Tuesday-slices? Are there minimal levels of informed-ness/unmanipulatedness necessary for moral subjecthood? Which animals/Boltzmann brains/victims of manipulation/subjects of moral luck qualify? Relevant to praise blame and the whole self, and Monima’s presentation cf. gratitudes possible objects. It seems odd on one hand that on the Buddhist revisionist view, the objects of gratitude can’t be aggregates of elements (people) but can be aggregates of aggregates of people (the Sangha). But maybe this should be more about replacing gratitude to with gratitude that, which only has abstract objects as its object. But in this case its unclear when and how much it’s apt to be grateful (plus which abstracta our gratitude should be about). If it should be actions or psychological states, it’s hard to see how this makes room

  • Compositional properties of moral subjects and agents: Can collectives be moral agents but not moral subjects, (the inverse of animals)? Maybe not, if being a moral agent requires possible deserving reward or punishment, (as opposed to mere benefit or harm), and onyl moral subjects can be rewarded or punished, (at least justly).

• If agents are bounded, (i.e. have limited computational resources), they can’t have informed preferences directly over world-states, (since they would have to perfectly model themselves, leading to paradoxes). Instead their preferences must be under-determined, or flow through their model of the world, or something. One of the main ways these preferences seem to evolve or feed into actions is by being socially transformed, (e.g people often have preferences to prefer what some other (real or ideal) agent would prefer them to prefer. This might look like the formation of a collective agent, as in a marriage, or the process of aspiring to some ideal form of agency, as in WWJD bumper stickers.) If you reject truth-functional semantics for inferential/functional role semantics you kinda get an analogous result about our concepts/world-models and beliefs, i.e. they’re really latent functions between social/mental inputs and actions/inferences. When are these social transformations autonomy/welfare preserving despite seeming necessarily uninformed/disempowering, and when are they manipulative/destructive? What kinds of trust, respect, love do they require? What impact will various social and non-social technologies have on them? Broadly virtue-ethical approaches generally seem to do a better job of remembering agents are bounded in this way, also taking into account akrasia, salience/attention, emotions, social roles, etc. not just beliefs and preferences over world-states.

  • How/when does the deprivation of future wellbeing harm an agent, especially different kinds of animals? Basically apply theories of the badness of death/population ethics that are grounded in particular psychological make-ups to different kinds of animals and see how that works, (could do this w/ animal ask though). Need to check how Kagan’s view responds to premature deaths in net-negative lives. Questions around how bad the deprivation of future positive experiences is for different kinds of animals based on closeness of connection with future selves. Can you make sense of some kind of deprivationist account about premature death being bad when animals are killed in net-negative lives. And more importantly, can you avoid the conclusion that the more premature a death in a net negative life, the better that is, such that rapidly killing heaps of baby chickens is a goodness generator. Another possibility is just taking a preference-frustration view of death, but I’m not convinced this really captures the badness of death because it doesn’t count as a harm the non-satisfaction of the preferences you would have developed had you lived longer.

• Are shapley values a good model for virtues? Could we somehow dovetail idealised preference/agent meta-ethics with shapley values? Shapley Values + Game Theory

List of links to practical issues

• Try to change prioritisation of animal welfare, prematurely killed sophisticated animal farming looks worse, (e.g. broiler chickens), fish, insects probably looks less bad.

• One possible ways of transforming agents and their preferences seem better and some seem worse, (compare anti-depressants, love drugs, debate, mind-uploading, combining, and forking, psychedelics, love, marriage, contracts, surrogate decision makers, and mind-control). Often the good cases are (partly) differentiated from the bad by whether they’re autonomy-enhancing or degrading. But is there a general rule for balancing conflicts of autonomy between past and future time-slices, or sub vs. meta agents? How might future technologies shift the balance towards one or the other kind of transformation?

• How you’re meant to balance the (still-fulfillable) preferences of the dead vs. the living vs. vs. your past self after preference change (cf. Callard) vs. possible future people.

• Moral uncertainty can be analyzed as one especially pressing case of unstable or contradictory core preferences, (depending on your view of

• AI safety, especially around corrigibility and ontology-changes. Much of the worry around powerful future AI systems is based on current machine learning methods tending to produce maximisers of some fixed, simple, utility function. This kind of agent (if powerful enough) is typically incentivised to avoid shutdown or outside alteration to it’s utility function and monopolize resources. Hopefully studying the very different kind of agents in my thesis could shed some light on what kind of agents are less likely to have these undesirable properties. Especially interested in

• Population ethics, e.g. what grounds our obligations to future generations? Should we adopt a total-welfare view, a constant discount rate, step-wise egocentric discounting grounded in co-membership of different collectives, or something else? What kinds of informational and causal links must exist to unite members of different generations into collective agents, (kind of Neo-Confucian-communitarian-ish if you squint)? What will this look like in more practical terms, (hopefully some kind of critical-level view with a neutral range). How will it deal with impossibility theorems in the area? Do the more theoretical results suggest different norms around the treatment of animals, especially around the badness of death/reasons to create different kinds of future lives? How should this affect our beliefs around the ethics of factory farming and wild animal welfare?