I've no idea why it's so difficult to understand the idea that an argument against computational functionalism is not itself an empirical demonstration that a particular substrate cannot realize the relevant properties. If you want to rule out LLM consciousness, identify process or property Y, specify what would count as instantiating Y, and show empirically that substrate A cannot do so. Otherwise you have argued for a philosophical position. You have not shown that LLM consciousness is impossible.

Whether machines are #conscious, or #intelligent for that matter, depends on how we operationalize those terms. And when we do, it becomes an empirical question. You don't get to answer it a priori.

So, @naturehumbehav desk rejected my reply to @Amit_Goldenb & Gross correspondence. I guess I'm not at Harvard. Posting it here instead. The problem is that G&G set a functional bar for emotions in language models, and then judge by the substrate. Their test asks what a system does. But every time they rule LLMs out, the evidence is about what the system is made of. Goldenberg and Gross make an important contribution by trying to move the debate away from verbal self-reports and toward the functions emotions perform within a system. Their stated standard is explicitly functional: the relevant question is not whether an artificial system resembles a human brain, but whether it contains processes that play the roles emotions play in humans. The problem is that their main arguments do not consistently follow that standard. When assessing LLMs, they repeatedly appeal to properties of biological systems, neural organization, and the computational architecture of transformers rather than applying the functional tests they themselves propose. So, in the end, the bar is functional, but the judgment is ultimately based on substrate, what the system is made of. It's an important distinction. If biological implementation is doing the real work in the argument, then the conclusion that present-day LLMs lack emotions has effectively been assumed rather than demonstrated. Goldenberg and Gross say their test "does not require emotions to take any particular physical form," and they are willing to attribute emotions to a frightened crowd. Yet when they turn to LLMs, the standard changes: now brains matter, neural patterns matter, and emotions must alter the operation of the system in specifically biological ways. They cannot have it both ways. If brain-like internal organization is required, then LLMs are excluded by definition, a crowd should fail for the same reason, and "perhaps one day" just does not mean anything anymore. Now, if, on the other hand, brain-like organization is not required, then the neuroscience they invoke does little to establish whether an LLM can have emotions. Also notice that stability does not necessarily imply rigidity either. In their own example, what varies across people is what triggers fear, not whether fear must be implemented differently each time it occurs. A stable internal organization is perfectly compatible with flexible elicitation. Much the same problem appears when they write that "each forward pass through the network is computationally fixed," unlike processing in "biological systems." Having just said that their test requires no "particular physical form," they now rule LLMs out by how they are built. Yet they also say they "adopt a functional perspective, asking what functions emotions serve and whether an LLM has processes that serve those functions." On a genuinely functional account, what matters is what the system does, not what it is made of. A mechanical clock and a quartz clock can both tell time, while what they are made of is wildly different. Put it another way, if an LLM confronted with a threat narrows its attention, makes decisions more rapidly, and changes the type or likelihood of its responses, then it is performing at least some of the functions they themselves associate with emotion, whatever machinery implements them. Saying that "each forward pass" is computationally fixed describes the implementation, not the function. This is the deeper problem with their argument. Their stated method is functional, and the tests they propose (changes in attention, speed, and "response type and likelihood") are substrate-neutral. But they never actually apply those tests to an LLM. The only LLM evidence they discuss is the Anthropic preprint, which they answer with findings from brain imaging. For their second function, they cite no LLM evidence at all, only contrasts such as "unlike humans" and appeals to "biological systems." In both cases, the verdict ultimately rests not on what LLMs do, but on what they are made of. Even their apparent escape hatch, the emotional crowd, is still a collection of biological systems. They set a functional bar and then judge by the substrate. To rule LLMs out on functional grounds, they need to identify a function that emotion serves in humans and show that LLMs do not perform it. Whether they do is an empirical question.