Heterophenomenology

In the AI Story

The method was introduced in Consciousness Explained (1991) as Dennett's answer to the bootstrapping problem in consciousness research: you cannot use first-person reports as ground truth because the reports are themselves part of what needs to be explained. His solution was to treat reports as data about how the subject represents her experience, without assuming the representation is accurate.

The method has always been controversial. Phenomenologists and consciousness researchers accused Dennett of denying the reality of experience. His response — consistent for thirty years — was that he was denying only the assumption that subjects have privileged epistemic access to what their brains are actually doing. The experience itself was real; the claim that introspection was a reliable scientist of the experience was what he rejected.

Applied to AI, the framework becomes newly urgent. The Orange Pill, and the viral confessions and Substack essays it documents, generate an enormous corpus of first-person reports about what it is like to work with AI tools. These reports contain signal — they are not fabrications — but they are not simple descriptions of what the tools are doing. They are interpretations, shaped by the frameworks the reporters bring, the feelings the reports attempt to manage, and the social positions from which they are issued. Heterophenomenology takes the reports seriously as data without taking them as oracular.

The same method applies, provocatively, to the AI itself. When Claude reports on its own processing — when a model generates reflective text about what it is doing — the text is data. It is not a direct window into whatever computation is actually occurring. It is a report, produced by a system whose reports are themselves part of what we are trying to understand.

Origin

Dennett developed the method across the 1980s and gave it its systematic statement in Consciousness Explained (1991). He refined the position in debates with David Chalmers, Thomas Nagel, and others through the 1990s and 2000s.

Its application to AI emerged in the 2010s as the need to interpret both user reports and model outputs grew urgent. Dennett's position, maintained to the end, was that his method was made for exactly this kind of interpretive problem — taking reports seriously without taking them as authoritative was the central discipline of any science of mind, biological or artificial.

Key Ideas

Reports are data, not testimony. What a subject says about her experience is evidence about how she represents her experience, not a direct read-out of what the underlying process is doing.

The bracketing move. The scientist of consciousness must suspend the question of what is really happening experientially and focus on the publicly available textual and behavioral data.

Symmetry across substrate. The method applies identically to human and machine subjects — take their reports seriously as data, do not treat the reports as infallible.

Interpretation is the work. The reports require framework, context, and disciplined reading; they do not decode themselves, and the person who produced the report is not always the best interpreter of it.