Pamela Robinson

How should we design AI systems that make moral decisions that affect us? When there is disagreement about which moral decisions should be made and which methods would produce them, we should avoid arbitrary design choices. However, I show that this leads to a regress problem similar to the one metanormativists face involving higher orders of uncertainty. I argue that existing strategies for handling this parallel problem give verdicts about where to stop in the regress that are either too arbitrary or too difficult to implement. I propose a new strategy for AI designers that is better than these alternatives.

We propose developing AIs whose only final goal is being shut down. We argue that this approach to AI safety has three benefits: (i) it could potentially be implemented in reinforcement learning, (ii) it avoids some dangerous instrumental convergence dynamics, and (iii) it creates trip wires for monitoring dangerous capabilities. We also argue that the proposal can overcome a key challenge raised by Soares et al. (2015), that shutdown-seeking AIs will manipulate humans into shutting them down. We conclude by comparing our approach with Soares et al.'s corrigibility framework.

Artificially intelligent systems will be used to make increasingly important decisions about us. Many of these decisions will have to be made without universal agreement about the relevant moral facts. For other kinds of disagreement, it is at least usually obvious what kind of solution is called for. What makes moral disagreement especially challenging is that there are three different ways of handling it. _Moral solutions_ apply a moral theory or related principles and largely ignore the details of the disagreement. _Compromise solutions_ apply a method of finding a compromise and taking information about the disagreement as input. _Epistemic solutions_ apply an evidential rule that treats the details of the disagreement as evidence of moral truth. Proposals for all three kinds of solutions can be found in the AI ethics and value alignment literature, but little has been said to justify choosing one over the other. I argue that the choice is best framed in terms of _moral risk_.

Generative artificial intelligence systems based on transformers, including both text-generators like GPT-4 and image generators like DALL-E 3, have recently entered the popular consciousness. These tools, while impressive, are liable to reproduce, exacerbate, and reinforce extant human social biases, such as gender and racial biases. In this paper, we systematically review the extent to which DALL-E Mini suffers from this problem. In line with the Model Card published alongside DALL-E Mini by its creators, we find that the images it produces tend to represent dozens of different occupations as populated either solely by men (e.g., pilot, builder, plumber) or solely by women (e.g., hairdresser, receptionist, dietitian). In addition, the images DALL-E Mini produces tend to represent most occupations as populated primarily or solely by White people (e.g., farmer, painter, prison officer, software engineer) and very few by non-White people (e.g., pastor, rapper). These findings suggest that exciting new AI technologies should be critically scrutinized and perhaps regulated before they are unleashed on society.

We sometimes seem to face fundamental moral uncertainty, i.e., uncertainty about what is morally good or morally right that cannot be reduced to ordinary descriptive uncertainty. This phenomenon raises a puzzle for noncognitivism, according to which moral judgments are desire-like attitudes as opposed to belief-like attitudes. Can a state of moral uncertainty really be a noncognitive state? So far, noncognitivists have not been able to offer a completely satisfactory account. Here, we argue that noncognitivists should exploit the formal analogy between moral uncertainty and moral multi-objective decision problems. Our ‘multi-objective story’ enables noncognitivists to save our moral uncertainty thought and talk by explaining how the underlying phenomenon could be noncognitive.

According to ‘Excluders’, descriptive uncertainty – but not normative uncertainty – matters to what we ought to do. Recently, several authors have argued that those wishing to treat normative uncertainty differently from descriptive uncertainty face a dependence problem because one's descriptive uncertainty can depend on one's normative uncertainty. The aim of this paper is to determine whether the phenomenon of dependence poses a decisive problem for Excluders. I argue that existing arguments fail to show this, and that, while stronger ones can be found, Excluders can escape them.

Kevin McCain’s Evidentialism and Epistemic Justification is the most thorough defense of evidentialism to date. In this work, McCain proposes insightful new theses to fill in underdeveloped parts of evidentialism. One of these new theses is an explanationist account of evidential fit that appeals to dispositional properties. We argue that this explanationist account faces counterexamples, and that, more generally, explanationists should not understand evidential fit in terms of dispositional properties.