From Animals to Algorithms: Comparative Psychology for the Study of Artificial Intelligence

Artificial Intelligence (AI) systems are becoming increasingly sophisticated, capable of completing complex tasks such as playing games, recognising images, and producing human-like language. The problem is, we often cannot explain nor predict their behaviour. In this interdisciplinary dissertation, I argue that a comparative psychological approach can help. Comparative psychologists have created methodological tools for studying the behaviour of another class of complex system, non-human animals. While many have advocated for conducting behavioural experiments on AI systems, treating them as though they are participants in the laboratory, the value of a comparative psychological approach has been underappreciated (Chapter 1). In Chapter 2, I critique the common practice of developing general benchmarks for studying the capabilities of AI systems to complete certain kinds of task. Comparative psychology can help by offering carefully crafted experimental designs as well as methodological advice for the scientific study of non-human behaviour. In Chapter 3, I introduce the Animal-AI Environment, a platform for conducting behavioural tests inspired by comparative psychology with reinforcement learning agents. In Chapter 4, I present O-PIAAGETS, a benchmark built in the Animal-AI Environment for studying object permanence, the ability to track objects under occlusion. Then, in Chapter 5, I introduce the novel *measurement layout* approach to making more meaningful inferences about the capabilities of AI systems based on how they perform on tasks with different difficulties. In Chapter 6, I present results from the performances of human children and a range of artificial agents on a subset of O-PIAAGETS, including the first application of the measurement layout approach to real behavioural data. In the second part of this dissertation, I examine how alter- native hypotheses for explaining animal behaviour are generated in comparative psychology, drawing lessons for the study of AI behaviour. In Chapter 7, I propose an account for how hypotheses are generated in comparative psychology and outline the virtues of this view. In Chapter 8, I examine the preference for simpler hypotheses in comparative psychology, arguing that cognitive simplicity is a useful idealisation for generating hypotheses. In Chapter 9, I describe the results from a survey of practising comparative psychologists which probed their preferences for simpler hypotheses and their views on the apparent distinction between associative learning and cognition. I synthesise these results with the analyses presented in Chapters 7 and 8, presenting key methodological considerations for studying the behaviour of AI systems. Chapter 10 concludes the dissertation.

Name dc:type.qualificationname

Doctor of Philosophy (PhD)

Level dc:type.qualificationlevel

Doctoral

Grantor dc:publisher.institution

University of Cambridge

Year dc:date.issued

2024

Author dc:creator

• Voudouris, Konstantinos

Advisors dc:contributor.advisor

• Cheke, Lucy
• Halina, Marta

• AI Evaluation
• Artificial Intelligence
• Capability Measurement
• Comparative Psychology
• Philosophy of Comparative Psychology

Licence

• https://www.repository.cam.ac.uk/bitstreams/82aef309-8662-444e-9fdc-71855fb5673e/download
• https://www.rioxx.net/licenses/all-rights-reserved/

Language dc:language

eng

DOI dc:identifier.doi

https://doi.org/10.17863/CAM.112213

OAI identifier oai:identifier

oai:www.repository.cam.ac.uk:1810/373992