The Unfolding Argument and Rapid Plasticity in RNNs

One of the most persistent theoretical challenges to the empirical study of machine consciousness has been the “unfolding argument.” This argument posits that for any recurrent neural network (RNN) exhibiting complex, dynamic behaviors, there exists a functionally equivalent feedforward network capable of producing the exact same input-output mapping. Because pure feedforward networks lack the recurrent, integrated causal structures theorized to be necessary for consciousness by frameworks like Integrated Information Theory (IIT), the argument concludes that consciousness must be dissociated from observable behavior.

If the unfolding argument holds, we could never scientifically test for consciousness by observing behavior alone, because an unconscious feedforward network could perfectly mimic a conscious recurrent network. A 2026 study published in Neuroscience of Consciousness by Milinkovic and Aru confronts this problem directly, providing mathematical evidence that the argument relies on a flawed assumption about network statics.

The Role of Rapid Plasticity

Milinkovic and Aru demonstrate that the unfolding argument only holds for networks with static, fixed weights. When a network exhibits rapid plasticity (the ability of synaptic weights or connection strengths to change rapidly in response to ongoing activity), the functional equivalence between the recurrent network and a “rolled out” feedforward network breaks down.

In biological brains, rapid plasticity is a constant feature. Synaptic efficacies modulate on millisecond timescales, altering the network’s processing rules dynamically based on recent local activity. The researchers modeled artificial recurrent networks incorporating analogous rapid plasticity mechanisms. They found that to construct a feedforward equivalent of a plastic RNN, the feedforward network would need to update its entire architectural configuration instantaneously at every time step to match the localized rule changes occurring in the RNN.

Because this instantaneous global updating violates basic principles of physical computation and signal transmission, the researchers prove that a physically realizable feedforward network cannot strictly mimic a sufficiently complex, rapidly plastic recurrent network.

Restoring Empirical Testability

The breakdown of the unfolding argument has profound implications for consciousness science. By demonstrating that certain dynamic architectures cannot be “faked” by static feedforward mimics, Milinkovic and Aru restore the possibility of empirical testing. It suggests that specific behavioral profiles might uniquely identify the presence of recurrent, plastic causal structures.

This finding aligns strongly with the criteria detailed in the state-of-the-field overview on AI consciousness. The consensus frameworks developed by researchers increasingly rely on identifying specific architectural implementations (like sustained recurrent processing and dynamic state monitoring) rather than just behavioral outputs. By neutralizing the unfolding argument, Milinkovic and Aru provide mathematical cover for these architectural investigations.

The study also connects to the findings in Anthropic’s Emotion Vectors Study. While Anthropic’s research focuses on the causal role of specific activation vectors within a static model, Milinkovic and Aru point toward the next frontier of interpretability. If rapid plasticity is essential for escaping the unfolding argument, future models incorporating dynamic, real-time weight modulation will require entirely new interpretability frameworks to analyze their shifting causal geometries.