How can competing theories of consciousness be tested?

For: Readers who know the basic question of consciousness and want to assess a theory comparison.

A theory becomes easier to assess when you can say what it predicts before seeing a result. Begin by distinguishing the broad idea from a particular version of it. “Global workspace”, “integrated information” and “predictive processing” name different research approaches; they are not interchangeable labels for any talk about consciousness. The first task is to identify the speaker’s commitments and the phenomenon those commitments are intended to explain.

The World Science Festival discussion with Rufin VanRullen explicitly concerns artificial-intelligence architectures inspired by Global Workspace Theory. Listen for the proposed role of the architecture and the evidence claimed for its usefulness. Keep two questions separate in your notes: whether a design helps a machine perform a task, and whether the machine has experience. Success on the first question does not supply the missing argument for the second. This discussion represents an application of a workspace approach, rather than an exhaustive presentation of every biological global neuronal workspace proposal.

Compare that with Koch’s integrated information lesson. Its paired summary presents the theory’s commitments about causal organisation. It is useful precisely because it is an identifiable proponent account. Record what the theory takes as its starting point, what must be measured, and what would make a measurement relevant to experience. Avoid silently replacing “this theory proposes” with “science has established”.

For a concrete empirical comparison, read the 2025 adversarial collaboration before deciding which presentation sounds more convincing. The study tested preregistered predictions associated with integrated information theory and global neuronal workspace theory, using visual stimuli and several brain-recording methods. Some findings supported predictions, while other findings challenged central predictions of both accounts. The paper includes differing interpretations. It does not announce a settled winner or test every possible version of either theory.

Use this case to practise a disciplined chain of reasoning: theory, prediction, measurement, result, interpretation. Ask where an extra assumption enters that chain. A recording method may measure only part of what a theory describes. A statistical pattern may fit more than one explanation. A theory’s advocates may revise the prediction after a result, but that revision then requires another test. These are questions to bring to the paper, not claims that any single criticism automatically defeats a theory.

Mediano’s information-dynamics proposal and Hipólito’s discussion of generative models provide a more demanding continuation. Their value here is to make the construction of a measure or model visible. Seth supplies a further named perspective, predictive processing, without being treated as a substitute for the actual two-theory experiment. Finish by writing one prediction table with an explicit “would challenge this version” column. A careful unresolved comparison is a better learning outcome than declaring a winner whose test you cannot specify.

Next step

Read the paper’s differing interpretations after completing the prediction table, then choose one proposed follow-up test to examine more closely.

Limits of this route

The videos include a 2015 proponent account, 2019 conference proposals and a 2024 workspace-inspired AI discussion. The 2025 paper’s abstract, prediction section and discussion framing were inspected, not every supplementary analysis. Video descriptions and paired summaries were checked; full recordings were not watched. No theory is marked as independently validated.