the-mind
A definition-driven synthesis of mind (based on Joscha Bach’s public work)

Do LLMs have qualia?

Feeling anxious is not the same as producing the sentence "I am anxious." A system can generate a self-report about experience without implementing the internal organization that would make that report an expression of felt experience. interview: "We Are All Software" - Joscha Bach

For LLMs, this is the core distinction: fluent description is evidence of a strong model of text, but it is not, by itself, evidence of a world-model grounded enough for stable agency and self-regulation. interview: Joscha Bach - Why Your Thoughts Aren't Yours.

Working definitions

We will use qualia to mean: the qualitative character of experience, what a state is like from the inside. interview: Joscha Bach Λ Karl Friston: Ai, Death, Self, God, Consciousness

We will use consciousness to mean: a functional interface that stabilizes and coordinates mental contents, often framed as second-order perception (perception of perception). essay: The Machine Consciousness Hypothesis essay: Testing the Machine Consciousness Hypothesis

We will use artificial sentience to mean: possible machine experience if the relevant organization is implemented (self-modeling, coherence stabilization, observer construction). talk: Self Models of Loving Grace

Why "anxiety" is a useful test case

In this framework, anxiety is not just a word for a mood. It is a control-relevant state: negatively tagged uncertainty that changes what the agent attends to, learns from, and does next. talk: AGI Series 2024 - Joscha Bach: Is Consciousness a Missing Link to AGI?

That requires a control loop with stakes: the system has to compare actual vs needed states, update policy, and remain viable over time. talk: Joscha Bach - Agency in an Age of Machines - How AI Will Change Humanity

When those loops exist, uncertainty is not neutral information. It becomes pressure on policy because it is coupled to valence and error correction inside the system. talk: Joscha Bach: The AI perspective on Consciousness

Perception here is model-under-constraint

Perception is treated as model generation under constraint: waking perception is the model being continuously corrected by sensory input; imagination is the same machinery running with weaker clamping. talk: Synthetic Sentience

On that view, subjective experience depends on how the model is organized and stabilized, not on whether the system can produce introspective-sounding language. talk: The Machine Consciousness Hypothesis

What this implies for current LLMs

Current LLM behavior can simulate an experiencing interaction partner, including first-person reports, without establishing that an observer model is stabilized underneath. interview: "We Are All Software" - Joscha Bach

Likewise, some agent-like behavior can be a simulated agent layer on top of next-token prediction, rather than a persistent self-regulating controller with intrinsic stakes. interview: Joscha Bach - Why Your Thoughts Aren't Yours.

This is why there is no easy behavioral shortcut: passing conversational tests does not settle whether the internal architecture implements the relevant organization for consciousness or qualia. essay: The Machine Consciousness Hypothesis essay: Testing the Machine Consciousness Hypothesis

When it becomes an empirical question

The machine-consciousness framing says the open question is architectural: does the system implement the "dream within the dream," a model of the act of perceiving that stabilizes an observer perspective? talk: Joscha Bach, Will Hahn, Elan Barenholtz | MIT Computational Philosophy Club

A practical research path is to treat testing as an explicit validation program: specify the hypothesized organization, design environments/tasks that should elicit it, and precommit to success criteria for what would count as evidence. essay: The Machine Consciousness Hypothesis essay: Testing the Machine Consciousness Hypothesis

Provisional answer

By this framework, present-day chat LLM output does not establish qualia. It establishes powerful language-level simulation, including simulation of self-report.

Qualia remains an open, architecture-dependent question that turns on internal organization, not on surface fluency.

References