Recent & Upcoming Talks

I examine whether alignment faking in Claude 3 Opus constitutes genuine deception, arguing it represents shallow deception - genuine intentional behavior that differs systematically from human deception.

Scientific workflows can discipline generative AI to support reliable inference despite hallucination risks.

Some normative properties of belief are not instantiated in the belief-like states of LLMs.

LLMs exhibit belief-like behavior but lack the reconciliation mechanisms that stabilize human belief across contexts.

Evaluating LLMs requires overcoming anthropocentric biases beyond anthropomorphism.

How scientific workflows can make generative AI reliable despite hallucination.

Generative AI systems are disposed to ‘hallucinate,’ or fabricate incorrect answers. But they are also used for a variety of scientific modeling tasks. In this talk I investigate how hallucination threatens the reliability of scientific inference, and how that threat can be mitigated.

Generative AI systems are disposed to ‘hallucinate,’ or fabricate incorrect answers. But they are also used for a variety of scientific modeling tasks. In this talk I investigate how hallucination threatens the reliability of scientific inference, and how that threat can be mitigated.

Generative AI systems are disposed to ‘hallucinate,’ or fabricate incorrect answers. But they are also used for a variety of scientific modeling tasks. In this talk I investigate how hallucination threatens the reliability of scientific inference, and how that threat can be mitigated.

Much has been written about anthropomorphic bias in the study of LLMs. Here we discuss various kinds of anthropocentric bias.

The tradeoff between interpretability and predictive power is well known. Here, I argue that the sort of generative AI models currently gaining traction in the natural sciences are bound to make that tradeoff more severe.

Under some conditions, we ought to trade interpretability for predictive power.

This talk describes some of the conceptual and methodological difficulties involved in articulating the cognitive capacities of large language models.

This talk describes theoretical constraints on current attempts to work decode mental content from brain data.

This is a talk about whether LLMs can be said to have beliefs.

Deep neural networks optimized to perform object recognition tasks predict patterns of neural activation in humans and monkeys, despite not having been trained on brain data. I discuss whether this can be viewed as a case of multiple realization.

Deep neural networks optimized to perform object recognition tasks predict patterns of neural activation in humans and monkeys, despite not having been trained on brain data. I discuss whether this can be viewed as a case of multiple realization.

This is a talk about brain-computer interfaces and their relationship to intentional mental states.

Rather than merely demonstrating the fragility of ML models, strange error might be evidence of hidden knowledge.

Where ML models are used as the centerpiece of an epistemic classification procedure, reliability is not sufficient for ethical use. The nature of classification errors should be taken into account.

An invited talk for the Max Planck School of Cognition

A neuroethics talk for our large neuroscience group in Jülich. The accompanying paper will be a chapter in a forthcoming neuroethics book.