Agency, goals, and perspective: how do natural or artificial agents understand the world?
Schwartz Reisman Research Lead Denis Walsh recently spoke at the SRI weekly seminar on normative phenomena and the interaction between an agent and its setting. Guest blogger Cory Travers Lewis reflects here on Walsh’s talk.
When we say that something is good or bad, is that a claim about objective facts in the world, or something dependent on our perspective?
A lot seems to hang on this question. If values are purely subjective, that would put them beyond the reach of systematic study. For the Schwartz Reisman Institute, which operates at the intersection of values, policy, and technology, that might seem like a problem. But values seem quintessentially subjective, meaningless, or ephemeral without some subject who values things. So which are they: objective, or subject-relative?
Schwartz Reisman Research Lead Denis Walsh suggests: they’re both.
Walsh is a professor in the Department of Philosophy and the Institute for the History and Philosophy of Science and Technology at the University of Toronto. His research in the philosophy of biology focuses on organisms as agents. He argues that we can’t understand living things unless we recognize that we have goals, which we pursue using flexible repertoires of behaviour. But talking about goals strikes some people, including many biologists, as at best short-hand for some much more complex story about molecular mechanisms, and at worst a throwback to pre-scientific ways of thinking. A goal implies success and failure, good and bad outcomes. But is this kind of normativity out there in the world?
In a recent SRI seminar series talk, Walsh presented a way of thinking about norms which treats them as both objective facts and as dependent on the perspective of particular living things.
➦ Missed Walsh’s talk? Watch the video here.
If this sounds contradictory, consider maps. A map can be correct or incorrect, of course. If you buy a map of Canada that puts Manitoba east of Quebec, you should get your money back. But there are a great many different ways of constructing a map, and for good reason. In his 1946 short story “On Exactitude in Science,” Jorge Luis Borges imagines a map so precise that it is in one-to-one scale with the kingdom it maps. Such a map would cover the entire territory it represented, and would obviously be useless.
To make a useful map, there has to be a selective representation of the territory. And the way that selection process happens will depend on what we’re interested in. For example: a subway map will pick out different features of a city than a bike map. And yet, either type of map could be right or wrong, objectively. If your bike map says there is no hill where there is one, again, you’ll want to get a refund. But hills don’t appear on subway maps at all, because those maps focus on a different slice of reality. This means that while mapping is perspectival, it is nonetheless subject to the kind of correct/incorrect judgements that are the hallmark of objectivity.
Perspectivism in the philosophy of science is a kind of pluralistic picture of knowledge developed by thinkers like Ron Giere. There may be many different ways of carving up the same territory, some of which are compatible, and many of which can’t be reconciled. But sometimes incompatible perspectives can each contribute something to our understanding. For example, according to philosopher of science Wendy Parker, our understanding of our climate is built on a variety of models which work on different basic assumptions, but each helps reveal different aspects of the whole.
Walsh suggested in his SRI seminar that this sort of perspectivism can help us understand how normativity is both objective and a matter of perspective.
Consider this: it is simply a fact that chocolate is bad for dogs. But that fact doesn’t hang around in the universe independently of the existence of dogs. Rather, it’s a fact about the relationship between the particular makeup of dog physiology, and the chemical composition of chocolate. For me, chocolate is not just good but the food of the gods. And that, likewise, is a fact which resides not just in me, or in chocolate, but in the relationship between me and chocolate. Despite the relativity of these facts to particular living things, they are far from subjective. You can’t just declare that the goodness of chocolate is subjective and feed a bar to your dog. When dogs came into the world with the type of physiology that they have, a new fact appeared in the world—that chocolate is poisonous to them.
This kind of agent relativity is familiar to ecologists. The way living things experience their environment, what it affords them in terms of both benefits and threats, depends on not just the environment, but on the agent-environment relationship. Tiny bugs experience water as a surface stable enough to walk on. We humans sink into it, and have to work hard to swim through it. But a blue whale experiences water as a relatively thin medium due to its enormous size. It’s the same water in every case, but what it affords each type of organism can be radically different, because of the unique organism-environment relationship.
Walsh argues that normativity is a lot like that. There aren’t normative facts free-floating in the world, independent of living things with goals, just as there is no simple fact about whether water is suitable for walking, swimming, or gliding easily. But there are nonetheless objective facts about normativity, as objective as whether a person can walk on water, when we consider the agent-world relationship.
This shows that norms are fully natural—in the sense that there is nothing spooky or otherworldly about them. They are not mere frameworks we impose on the world, in the same way that chocolate is not poisonous to dogs just because we imagine it to be. This makes norms the kind of thing that science can study. It just needs to be sensitive to the fact that norms exist as relationships between organisms and the world. The laws of physics were around before the first single-celled organisms emerged, but norms were not. We bring objective facts into being by being agents with goals, in our complex and unique interactions with each other and the world.
This opens up any number of questions, including questions about what the values of artificial general intelligences (AGI) may look like. If they are truly agents with perspectives, then it seems they will have their own objectively subjective values as well. Walsh didn’t say how we should approach those questions, being happy to have set them on a clearer foundation for further research.
Want to learn more?
Learn more about Denis Walsh’s research and publications.
Listen to Denis Walsh speak about his work on organisms, agency, and their interactions with their environments on the Big Biology podcast.
Read about SRI Director Gillian Hadfield’s research on normativity and aligning AI with human values.
About the author
Cory Travers Lewis is an assistant professor, teaching stream at the University of Toronto, where he teaches at the Institute for the History and Philosophy of Science and Technology (IHPST). He received his PhD in 2017 from the IHPST. His research focuses mainly on generality and contingency in scientific theories. Learn more about Cory on his website.
