Artificial Intelligence, Values and Reflective Processes
The science of human behavior is afflicted by a replication crisis. By some estimates, over half of the empirical literature does not replicate. A significant cause of this problem is undertheorization. Without a cumulative theoretical framework from which to work, researchers often lack meaningful hypotheses to test, and so instead default to their personal, often culturally biased folk intuitions. Their resulting interpretations of studies’ data thus frequently fail to replicate and generalize (See the seminal paper of Michael Muthukrishna and my advisor Joe Henrich.)
Finding the correct causal microfoundations for behavioral science can provide a deeper understanding of precisely when we can extrapolate empirical findings out-of-sample. This could be especially helpful for making externally valid predictions in historically unprecedented situations (e.g., regarding emergent technologies or anthropogenic catastrophic/existential risks), for which much of the relevant data required for empirically estimating policy counterfactuals may not yet exist.
One area where the correct causal theory of descriptive human behavior would be particularly helpful is correctly understanding and solving the AI-human alignment problem.
Some approaches include the provision of fellowships, grants, and collaborative opportunities to researchers, as well as teaching/mentoring/incentivizing of undergraduate students to help them become researchers or practitioners of plausible causal theories of behavioral science. (e.g., cultural evolutionary theory; see The Secret of our Success by Joe Henrich)
Causal microfoundations for behavioral science
Artificial Intelligence, Values and Reflective Processes
The science of human behavior is afflicted by a replication crisis. By some estimates, over half of the empirical literature does not replicate. A significant cause of this problem is undertheorization. Without a cumulative theoretical framework from which to work, researchers often lack meaningful hypotheses to test, and so instead default to their personal, often culturally biased folk intuitions. Their resulting interpretations of studies’ data thus frequently fail to replicate and generalize (See the seminal paper of Michael Muthukrishna and my advisor Joe Henrich.)
Finding the correct causal microfoundations for behavioral science can provide a deeper understanding of precisely when we can extrapolate empirical findings out-of-sample. This could be especially helpful for making externally valid predictions in historically unprecedented situations (e.g., regarding emergent technologies or anthropogenic catastrophic/existential risks), for which much of the relevant data required for empirically estimating policy counterfactuals may not yet exist.
One area where the correct causal theory of descriptive human behavior would be particularly helpful is correctly understanding and solving the AI-human alignment problem.
Some approaches include the provision of fellowships, grants, and collaborative opportunities to researchers, as well as teaching/mentoring/incentivizing of undergraduate students to help them become researchers or practitioners of plausible causal theories of behavioral science. (e.g., cultural evolutionary theory; see The Secret of our Success by Joe Henrich)