Drafts

Comments welcome! For up-to-date drafts, pointing out mistakes or if you'd like to collaborate, email me.

Formal Ethics

Intention, Alignment, and Risk in AI Agents

I make precise a notion of intent-alignment in Markovian agents using ideas from social choice theory, and I define a quantitative notion of risk for such agents.

Existential Catastrophic Risk with Non-Standard Measures

We propose to use hyperreal numbers to model existential catastrophic risks.

Modeling Value-Based Disagreement via Imprecise Measure Theory

I provide a model of value-based peer disagreement via Imprecise Measure Theory, which generalizes imprecise probability.

Philosophy, Law and Ethics of AI

Intentions for AI Agents

I distinguish between a descriptive question (can AI agents have intentions?) and a prescriptive question (should AI agents (be designed to) have intentions?).

Within the descriptive question, I discuss some older and newer proposals in the literature.

Within the prescriptive question I explore two aims of a positive answer: a retrospective aim (e.g. to ascribe backward-looking responsibility) and a

prospective aim (e.g. as an alignment strategy).

On Law-Following AI

Can an artificial agent follow the law?

I will argue against an anthropocentric view, i.e. that in order to be able to follow the law, an agent needs to have human abilities. When applied to artificial agents, this would amount to require them to be strong AI in Searle’s sense (which I argue against).
Instead, I focus on the reason-responsiveness theory, and investigate whether artificial agents (in particular based on reinforcement learning techniques) can be said to respond to and have reasons, and if so, whether our ideas of law-following (and therefore, or law) should be broadened.

Reasons-based AI Ethics

This paper introduces and defends a reasons-based AI ethics, focusing on  RL agents.

Normative Risk

This paper introduces and defends the concept of normative risk.

It discusses three notions of normative risk: (i) normative risk as a probable normative or moral harm, (ii) normative risk as normative underdeterminacy, and (iii) normative risk as norm-related existential risk. It uses AI as a case study.

Logic

The Content of Generics

I argue for a novel hyperintensional understanding of generic sentences based on arbitrary truthmaker semantics and subtraction.

Bilateral truthmaker Semantics for Would and Might Counterfactuals and Counterpossibles

We explore different options to give a bilateral truthmaker semantics to would and might counterfactuals and counterpossibles (joint work with Mattia Vargas).

Truthmaker Semantics for Justification Logic

We explore different options to give a truthmaker semantics for several justification logics.

First-order Hyperintensional Deontic Logic

I present a version of a first-order hyperintensional deontic logic.

Metaphysics (no longer in progress)

Identity in Higher-Order Logic and Metaphysics

I investigate the question of the identity of higher-order entities via model-theoretic notions.

First, I propose suitable notions of indiscernibility for higher-order entities and discuss preliminary issues about the adequacy of formal languages. Second, I formulate two conjectures [edit Sept. 2016: one is solved] about the relations between indiscernibility, automorphisms and higher-order equivalence, one of which depends on assumptions about large cardinals.

I then discuss their relevance for higher-order metaphysics, structuralism in the philosophy of mathematics, and the philosophy of physics.

Higher-Order Supervenience and Nonnaturalism

I argue that a higher-order supervenience principle, together with a hyperintensional account of properties, deflect Jackson-style arguments against nonreductive normative nonnaturalism.

I also prove that usual definitions of supervenience cashing out indiscernibility in terms of isomorphisms are model-theoretically inadequate.