FWIW Harvey Lederman and I defend a version of what you're calling "as if" mental states (https://philpapers.org/rec/GOLWDC-2), but we reject the idea that they aren't "really" mental states in a "substantive" sense (we call our view 'objective interpretationism'). For example, I think that there is a sensible approach to special sciences in general that has these implications about psychology in particular. Tectonic plates exist if and only if they are part of the best theory of geology, and they are part of that theory if and only if they help explain the geological data. Analogously, one approach to psychology is that the psychological data is behavior, and beliefs and desires exist if and only if they are part of the best theory of behavior. Crucially, however, this need not require structured internal representations. Now the resulting approach to special sciences certainly doesn't say that tectonic plates are merely 'as if' objects; likewise, the approach to psychology wouldn't say that beliefs and desires are merely 'as if' mental states.
In addition, I disagree that work on personas suggests that AIs don't satisfy functional conditions on mental states. A different way to understand the ideology of 'personas' or 'role play' defended by Harvey Lederman and I, and also by Chalmers, is just that each conversation with a single "model" is its own agent, with its own beliefs and desires. Different conversations will involve different personas, with different kinds of beliefs and desires. But for example each conversation may involve mental representations. I'm worried that a lot of the ideology of personas and role play comes from a misguided attempt to find a single set of beliefs and desires to associate with an overall model, when instead we should be assigning beliefs and desires to individual AI agents, which are roughly individuated by the conversation.
Hi Simon, thanks for this. Yes - there is definitely more room for me to engage with your view here, and I also wonder whether what Chalmers calls "quasi-beliefs" for example is closer to it. I need to read your paper in more detail to say whether I agree on your main point, and that might be reason to provide a level intermediate between "as-if" and functional. (This was mostly my first attempt at writing something short on the topic and get some public writing started). I would have thought that tectonic plates would count as something more than "as-if" things because we use the phenomena explained by positing them to make an inference to the best explanation about their existence as causally efficacious units, while we could in principle also be agnostic about that (though that would seem strange in that case). But you're saying that we can do the same for mental states, even if we're not positing internal representations. I'll have to think more about that.
I don't think I meant to suggest that persona work suggests AIs don't satisfy functional conditions. I agree that personas or other states might be the causal drivers of behavior, and that this might look very different from humans (see end of section on functional states). My ending scepticism was rather about how explanatorily powerful versions of these results are, insofar as the behaviors are clearly tied to the semantic meaning of the activated vectors. I definitely agree that we should not necessarily be looking for persistent beliefs and desires in particular as the functional states in LLMs, even though we might in individual agent-instances.
FWIW Harvey Lederman and I defend a version of what you're calling "as if" mental states (https://philpapers.org/rec/GOLWDC-2), but we reject the idea that they aren't "really" mental states in a "substantive" sense (we call our view 'objective interpretationism'). For example, I think that there is a sensible approach to special sciences in general that has these implications about psychology in particular. Tectonic plates exist if and only if they are part of the best theory of geology, and they are part of that theory if and only if they help explain the geological data. Analogously, one approach to psychology is that the psychological data is behavior, and beliefs and desires exist if and only if they are part of the best theory of behavior. Crucially, however, this need not require structured internal representations. Now the resulting approach to special sciences certainly doesn't say that tectonic plates are merely 'as if' objects; likewise, the approach to psychology wouldn't say that beliefs and desires are merely 'as if' mental states.
In addition, I disagree that work on personas suggests that AIs don't satisfy functional conditions on mental states. A different way to understand the ideology of 'personas' or 'role play' defended by Harvey Lederman and I, and also by Chalmers, is just that each conversation with a single "model" is its own agent, with its own beliefs and desires. Different conversations will involve different personas, with different kinds of beliefs and desires. But for example each conversation may involve mental representations. I'm worried that a lot of the ideology of personas and role play comes from a misguided attempt to find a single set of beliefs and desires to associate with an overall model, when instead we should be assigning beliefs and desires to individual AI agents, which are roughly individuated by the conversation.
Hi Simon, thanks for this. Yes - there is definitely more room for me to engage with your view here, and I also wonder whether what Chalmers calls "quasi-beliefs" for example is closer to it. I need to read your paper in more detail to say whether I agree on your main point, and that might be reason to provide a level intermediate between "as-if" and functional. (This was mostly my first attempt at writing something short on the topic and get some public writing started). I would have thought that tectonic plates would count as something more than "as-if" things because we use the phenomena explained by positing them to make an inference to the best explanation about their existence as causally efficacious units, while we could in principle also be agnostic about that (though that would seem strange in that case). But you're saying that we can do the same for mental states, even if we're not positing internal representations. I'll have to think more about that.
I don't think I meant to suggest that persona work suggests AIs don't satisfy functional conditions. I agree that personas or other states might be the causal drivers of behavior, and that this might look very different from humans (see end of section on functional states). My ending scepticism was rather about how explanatorily powerful versions of these results are, insofar as the behaviors are clearly tied to the semantic meaning of the activated vectors. I definitely agree that we should not necessarily be looking for persistent beliefs and desires in particular as the functional states in LLMs, even though we might in individual agent-instances.
Look at George Kellys Personal Consytruct Theory on how a person thinks and makes sense.
Makes more sense to me than any other explanation.