Stephen M. Downes

This paper presents the preliminary results of a survey-based study investigating geneticists’ understanding of genetic causation. We are motivated by an underlying assumption that to improve interpretation and implementation of behavioral genetics research, and furthermore for that research to have a positive societal impact, it is important to reveal how geneticists themselves understand their research. Results of sixty-two survey responses from clinical and research geneticists at the University of Utah are discussed and assessed to fulfill these aims and survey the landscape of the field. Analysis of our survey results revealed some variation in responses to questions about genetic causation from geneticists across different research areas.

I assess the potential for human behavior genetics research in advancing our understanding of human evolution. I argue that, despite such work in genetics producing alternate measures of heritability, it is not set up to provide evolutionary answers to questions about human variation or diversity. I pursue this by comparing and contrasting the way in which human behavior geneticists, on the one hand, and human evolutionary geneticists, on the other, pursue genetics research. Each area is guided by different research questions, different theoretical assumptions, and different uses of or understandings of key theoretical concepts. In Sect. 11.2, I outline the type of research included under the heading “human behavior genetics” and consider some examples of work on human height and educational attainment. In Sect. 11.3, I characterize human evolutionary genetics via a brief introduction to the shared theoretical assumptions and key concepts in the field and introduce a number of examples of human evolutionary genetics research. In Sect. 11.4, I compare and contrast the deployment of evolutionary concepts in human evolutionary genetics and human behavior genetics, focusing on several key concepts: trait and phenotype; genetic load; and additivity. I argue that in each of these cases, human behavior geneticists adapt evolutionary concepts to the context of their research, which is the establishment of relations between genotype and phenotype. I conclude that this move undercuts any potential for GWAS-based genetics to contribute to our understanding of human evolution.

We propose an approach to the evolution of joint agency and cooperative behavior that contrasts with views that take joint agency to be a uniquely human trait. We argue that there is huge variation in cooperative behavior and that while much human cooperative behavior may be explained by invoking cognitively rich capacities, there is cooperative behavior that does not require such explanation. On both comparative and theoretical grounds, complex cognition is not necessary for forms of joint action or the evolution of cooperation. As a result, promising evolutionary approaches to cooperative behavior should explain how it arises across many contexts.

It is commonly acknowledged in science that model construction is one of the most important components of theorizing. Philosophers of science are gradually coming to acknowledge this situation, spurred on by holders of the semantic view of theories. In this paper I wish to defend a very deflationary version of the semantic view of theories, which is more or less a re-statement of the above commonplace. I reject the view encapsulated in the identity statement “scientific theories are families of models,” although acknowledging the useful insights into science that holders of this strong position have given us. My position derives from a critique of various of the semantic views of theories, and further from a guiding presupposition that rather than providing necessary and sufficient conditions for what a theory is, philosophers should focus on the nature of scientific theorizing. Theorizing is carried out by practicing scientists, and we cannot say what scientific theories are unless we appreciate the myriad ways they are used and developed in all of the sciences.

We propose an approach to the evolution of joint agency and cooperative behavior that contrasts with views that take joint agency to be a uniquely human trait. We argue that there is huge variation in cooperative behavior and that while much human cooperative behavior may be explained by invoking cognitively rich capacities, there is cooperative behavior that does not require such explanation. On both comparative and theoretical grounds, complex cognition is not necessary for forms of joint action, or the evolution of cooperation. As a result, promising evolutionary approaches to cooperative behavior should explain how it arises across many contexts.

LLMs are statistical models of language learning through stochastic gradient descent with a next token prediction objective. Prompting a popular view among AI modelers: LLMs are just next token predictors. While LLMs are engineered using next token prediction, and trained based on their success at this task, our view is that a reduction to just next token predictor sells LLMs short. Moreover, there are important explanations of LLM behavior and capabilities that are lost when we engage in this kind of reduction. In order to draw this out, we will make an analogy with a once prominent research program in biology explaining evolution and development from the genes eye view.

In lieu of an abstract, here is a brief excerpt of the content:The Uselessness of Polygenic Scores for Addressing Campus DrinkingBennett Knox (bio), Hannah Allen (bio), and Stephen M. Downes, PhD (bio)Here we articulate a negative answer to Turkheimer and Greer’s question: “Is it possible to envision a genetically informed program that ethically intervenes on campus drinking?” (Turkheimer & Greer, 2024). However, first, we note that the authors cover an immense amount of ground in their paper. They lend insight into how psychiatric genetics, at its very core, is conducted through their detailed examination of a large body of work in one specific area of this large field. A main result of this is to explain the gulf between results and conclusions, work that provides an invaluable service not just to various areas in philosophy and bioethics, but that are deserving of readership by a very wide audience. This gulf is explained via a careful accounting of the statistical measures used to assess the impact of genes on alcohol-related behavior in the college students sampled and comparisons between the strength of these measures and the strength and scope of conclusions drawn by the Spit for Science (S4S) researchers.A brief mention of the statistical measures in question helps set up our main focus. As Turkheimer and Greer note, before advances in sequencing technology, human behavior geneticists generated heritability coefficients for traits of interest. These coefficients were standardly produced from twin study work, which historically was the best way of getting a sense of the relative contribution of genes to a trait of interest as we cannot, or rather should not, conduct breeding experiments on humans. Current descendants of heritability coefficients come from genome-wide association studies (GWAS), which can be conducted on large groups of unrelated people. One key measure of interest arising from GWAS is a polygenic score (PGS), which “can be thought of as an observed manifestation at the individual level of the abstract variance ratio estimated by a heritability” (Turkheimer & Greer, 2024). Turkheimer and Greer focus on R2 values, which roughly “describe the magnitude of the relationship between the polygenic score and the phenotypes [of interest]” [End Page 437] (Turkheimer & Greer, 2024). They point out that the largest R2 reported in an S4S study is 6.5% (0.065). R2 are taken to be small at 0.01, medium at 0.09, and large at 0.25 in many areas of social science research but there are fields, for example, drug assay work, where only much higher R2 values are considered indication of a large effect. By this assessment, most of the R2 values reported by the S4S researchers, or calculated by Turkheimer and Greer when not reported, are very, very small, sitting well below the level of small effect of 0.01.Given these very low R2 measures, we agree with Turkheimer and Greer that S4S researchers did not uncover genetic information that could be informative, let alone actionable, for administrators concerned with campus drinking. However, it is worth asking what it would have looked like if S4S had produced such information. Suppose that S4S had been wildly successful, producing the Gattaca-like genetic information of “polygenic scores making reliable predictions that individuals had a 35% chance of developing a psychiatric disorder [here, alcoholism]” (Turkheimer & Greer, 2024). If a student was given their PGS, how are they to interpret it and act on it? How would this information be put to use by an administrator to address the problem of campus drinking?For the information to be leveraged at all, it would require consent from a significant number of students to have their PGS calculated and made available to themselves or administrators. Were scores to be calculated without active consent from the student (perhaps by using the sample provided to S4S), this would be a serious ethical violation. But even if a sufficient number of students did consent to having their PGS for alcoholism calculated, it is still not clear how this information could be used in an ethical manner to address the problem of campus drinking.If this student was already aware of a family history of alcoholism, then the PGS might do little more than quantify...

Biologists, climate scientists, and economists all rely on models to move their work forward. In this book, I explore the use of models in these and other fields to introduce readers to the various philosophical issues that arise in scientific modeling. I show that paying attention to models plays a crucial role in appraising scientific work. After surveying a wide range of models from a number of different scientific disciplines, I demonstrate how focusing on models sheds light on many perennial issues in philosophy of science and in philosophy in general. For example, reviewing the range of views on how models represent their targets introduces readers to the key issues in debates on representation, not only in science but in the arts as well. Also, standard epistemological questions are cast in new and interesting ways when we confront the question, "What makes for a good (or bad) model?"

LLMs are statistical models of language learning through stochastic gradient descent with a next token prediction objective. Prompting a popular view among AI modelers: LLMs are just next token predictors. While LLMs are engineered using next token prediction, and trained based on their success at this task, our view is that a reduction to just next token predictor sells LLMs short. Moreover, there are important explanations of LLM behavior and capabilities that are lost when we engage in this kind of reduction. In order to draw this out, we will make an analogy with a once prominent research program in biology explaining evolution and development from the gene's eye view.

Cognitive science of science attempts to explain a range of phenomena familiar to philosophers of science, such as theory choice and scientific discovery. The appeal to cognitive science may be seen as an attempt to naturalize the philosophy of science. I examine and criticize several of the most important contributions to this new field. I argue that an unrecognized common defect of this work is its reliance on an explanatory approach that takes individuals' cognitive capacities as its units of analysis. I introduce the term "cognitive individualism" to identify this position, and conclude by examining the position in detail and sketching alternative approaches to naturalizing philosophy stressing the social dimensions of science. ;In Chapter 1 I briefly describe the field of cognitive science, and outline the empirical resources it can provide a philosopher of science. I then outline key themes of current cognitive science of science. In the next four chapters I critically examine the work of four prominent cognitive scientists of science: Herbert Simon, Paul Thagard, Ronald Giere, and Paul Churchland. All share the same goals of naturalizing philosophy of science by using the empirical resources of cognitive science. I show that all four accounts ignore the important social nature of science, and share an adherence to cognitive individualism. ;In the final chapter I develop the notion of cognitive individualism in detail. I show that relying on empirical evidence from work in cognitive psychology on human judgment, and work in the sociology of science is more fruitful for explaining science than the current cognitive individualist approach. I conclude with several theses that act as guidelines for future research in naturalized philosophy of science.

Herbert Simon’s work on scientific discovery deserves serious attention by philosophers of science for several reasons. First, Simon was an early advocate of rational scientific discovery, contra Popper and logical empiricist philosophers of science (Simon 1966). This proposal spurred on investigation of scientific discovery in philosophy of science, as philosophers used and developed Simon’s notions of “problem solving” and “heuristics” in attempts to provide rational accounts of scientific discovery (See Nickles 1980a, Wimsatt 1980). Second, Simon promoted and developed many of the crucial techniques and methods used in cognitive science. One is the use of computers to model internal psychological processes, a technique central to his account of scientific discovery. Another is protocol analysis, the use of the verbal reports of experimental subjects in psychology to construct accounts of their psychological processes. Protocol analysis is given a detailed formulation by Simon (Simon and Ericsson 1984), and is modified for use in the study of scientific cognition in the paper on Krebs (Kulkarni and Simon 1988).

There are a number of competing hypotheses about human evolution. For example, Homo habilis and Homo erectus could have existed together, or one could have evolved from the other, and paleontological evidence may allow us to decide between these two hypotheses (see, e.g., Spoor et al., 2007). For most who work on the biology of human behavior, there is no question that human behavior is in some large part a product of evolution. But, there are competing hypotheses in this area as well. Some claim that human behavior is produced by a collection of psychological mechanisms, for the most part, and that these mechanisms are adaptations that arose in the Pleistocene Epoch (e.g., Tooby & Cosmides, 1992; Buss, 2007). The claim is important and testable (although, more difficult to test than the above mentioned hypotheses about origins); but importantly, it is only one among many hypotheses about the evolutionary origins of human behavior. While I think that there may be components of our behavior that are best explained by appealing to processes or mechanisms that arose in the Pleistocene, I think that human behavior is a result of evolutionary processes both much older and more recent than the Pleistocene. I also maintain that much of human behavior, and the mechanisms underlying it, could still be subject to evolutionary..

Fisher’s 1918 paper accomplished two distinct goals: unifying discrete Mendelian genetics with continuous biometric phenotypes and quantifying the variance components of variation in complex human characteristics. The former contributed to the foundation of modern quantitative genetics; the latter was adopted by social scientists interested in the pursuit of Galtonian nature-nurture questions about the biological and social origins of human behavior, especially human intelligence. This historical divergence has produced competing notions of the estimation of variance ratios referred to as heritability. Jay Lush showed that they could be applied to selective breeding on the farm, while the early twin geneticists used them as a descriptive statistic to describe the degree of genetic determination in complex human traits. Here we trace the early history of the heritability coefficient now used by social scientists.

In this paper I briefly introduce work on ancient-DNA and give some examples of the impact this work has had on responses to questions in archaeology. Next, I spell out David Reich’s reasons for his optimism about the contribution aDNA research makes to archaeology. I then use Robert Chapman and Alison Wylie’s framework to offer an alternative to Reich’s view of relations between aDNA research and archaeology. Finally, I develop Steven Mithen’s point about the different questions archaeologists and geneticists ask, arguing that different disciplinary perspectives color researchers’ perceptions of “the most important questions” or the “central topics” in a field. I conclude that evidence from aDNA research cannot solve archaeological disputes without closer, mutually respectful collaboration between aDNA researchers and archaeologists. Ancient DNA data, like radiocarbon data, is not a silver bullet for problems in archaeology.

In this paper I examine Paul Thagard's computational approach to studying science, which is a contribution to the cognitive science of science. I present several criticisms of Thagard's approach and use them to motivate some suggestions for alternative approaches in cognitive science of science. I first argue that Thagard does not clearly establish the units of analysis of his study. Second, I argue that Thagard mistakenly applies the same model to both individual and group decision making. Finally, I argue that in attempting to account for psychological and social processes as well as providing a philosophical model of successful reasoning Thagard attempts to explain too much with one model, thus straining the plausibility of his model.