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.

Generative AI increasingly supports scientific inference, from protein structure prediction to weather forecasting. Yet its distinctive failure mode, hallucination, raises epistemic alarm bells. I argue that this failure mode can be addressed by shifting from data-centric to phenomenon-centric assessment. Through case studies of AlphaFold and GenCast, I show how scientific workflows discipline generative models through theory-guided training and confidence-based error screening. These strategies convert hallucination from an unmanageable epistemic threat into bounded risk. When embedded in such workflows, generative models support reliable inference despite opacity, provided they operate in theoretically mature domains.

This paper identifies two overlooked forms of bias in LLM evaluation: auxiliary oversight (failing to account for factors that impede performance despite competence) and mechanistic chauvinism (dismissing non-human problem-solving strategies as illegitimate). We propose addressing these through empirically-driven approaches combining behavioral experiments with mechanistic analysis.

We examine virtual reality’s potential as a tool for moral learning and empathy development. While we reject the ’empathy machine hypothesis’ that VR promotes empathy by simulating others’ perspectives, we propose that avatar use may support moral learning through a different mechanism: self-fragmentation.

A popular article arguing that, when evaluating LLMs, anthropocentrism is just as misleading as anthropomorphism.

When we use methods from experimental psychology to test the capacities of LLMs, we are prone to transfer assumptions about the human case to the LLM case, and to do so without justification. By drawing attention to these assumptions we can make more informed comparisons.

In order to tell whether someone is culpable for an action initiated by a brain-computer interface, it is not necessary to work out whether the brain-computer interface correctly decoded their intention.

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.

Argument maps represent some kinds of arguments better than others.

Reading minds is easier than you might think.

Network models support novel forms of discovery, prediction, and explanation. They also raise a philosophical puzzle about unification.

Neural reuse helped to liberate humans from evolutionary constraints faced by our ancestors.

In the brain, semantic information is intertwined with Shannon information.

The concept of neural coding makes sense, if the codes can be learned by neurons.

If meta-cognition evolved, there is probably something like semi-meta-cognition.

There can be an objective fact about the number of bits in a biological signal, despite the fact that the signal is receiver-relative.

Network representation compresses information about complex systems without abstracting away from the properties that make them complex.

If there are localized functions in the brain, they can only be articulated by abstracting away from functions associated with particular experimental tasks.