Full bibliography

Artificial neural networks are becoming more advanced and human-like in detail and behavior. The notion that machines mimicking human brain computations might be conscious has recently caused growing unease. Here, we explored a common computational functionalist view, which holds that consciousness emerges when the right computations occur—whether in a machine or a biological brain. To test this view, we simulated a simple computation in an artificial subject’s “brain” and recorded each neuron’s activity when the subject was presented with a visual stimulus. We then replayed these recorded signals back into the same neurons, degrading the computation by effectively eliminating all alternative activity patterns that otherwise might have occurred (i.e., the counterfactuals). We identified a special case in which the replay did nothing to the subject’s ongoing brain activity—allowing it to evolve naturally in response to a stimulus—but still degraded the computation by erasing the counterfactuals. This paradoxical outcome points to a disconnect between ongoing neural activity and the underlying computational structure, which challenges the notion that consciousness arises from computation in artificial or biological brains.