Thanks for writing this up, I broadly agree. There’s a framing from the AI measurement literature (Olteanu et al.’s position paper on rigor, NeurIPS ’25) that I think validates and possibly strengthens the proposal: the proposed approach is more rigorous than the top-down one, if you stop equating rigor with methodological rigor. The paper argues that the AI community’s usual notion of rigor, roughly whether the methods are correctly chosen and applied, is only one of six facets. The others include epistemic rigor (is the background knowledge the work builds on sound, and has it been interrogated rather than just assumed?), conceptual rigor (are the theoretical constructs clearly defined and made specific for the context at hand, rather than left as vague borrowed terms?), and interpretative rigor (do the claims drawn from the findings actually follow, given everything upstream?).
If read through this lens, the top-down approach can be methodologically sound while failing on the upstream steps. For example, importing a theory calibrated to humans, where the candidate welfare-relevant properties reliably co-occur, is an epistemic failure: the background knowledge hasn’t been shown to transfer to the systems under study, and it’s being used because it’s available rather than because it’s justified for this domain (the authors have a nice line about artifacts becoming “tools of opportunity, not instruments of epistemic rigor”). The target-level ambiguity you describe with GWT is a conceptual failure, since “global workspace” hasn’t been pinned down for transformers and any threshold ends up arbitrary.
I see your philosophical-probe proposal (the strong version) as a way to improve on the conceptual (broad construct → explicit testable definitions), epistemic (small, revisable theoretical commitments) and interpretative (narrow probe → narrow conclusion) facets of rigor in AI welfare research. I hope you find this framing useful.
Thanks for writing this up, I broadly agree. There’s a framing from the AI measurement literature (Olteanu et al.’s position paper on rigor, NeurIPS ’25) that I think validates and possibly strengthens the proposal: the proposed approach is more rigorous than the top-down one, if you stop equating rigor with methodological rigor.
The paper argues that the AI community’s usual notion of rigor, roughly whether the methods are correctly chosen and applied, is only one of six facets. The others include epistemic rigor (is the background knowledge the work builds on sound, and has it been interrogated rather than just assumed?), conceptual rigor (are the theoretical constructs clearly defined and made specific for the context at hand, rather than left as vague borrowed terms?), and interpretative rigor (do the claims drawn from the findings actually follow, given everything upstream?).
If read through this lens, the top-down approach can be methodologically sound while failing on the upstream steps. For example, importing a theory calibrated to humans, where the candidate welfare-relevant properties reliably co-occur, is an epistemic failure: the background knowledge hasn’t been shown to transfer to the systems under study, and it’s being used because it’s available rather than because it’s justified for this domain (the authors have a nice line about artifacts becoming “tools of opportunity, not instruments of epistemic rigor”). The target-level ambiguity you describe with GWT is a conceptual failure, since “global workspace” hasn’t been pinned down for transformers and any threshold ends up arbitrary.
I see your philosophical-probe proposal (the strong version) as a way to improve on the conceptual (broad construct → explicit testable definitions), epistemic (small, revisable theoretical commitments) and interpretative (narrow probe → narrow conclusion) facets of rigor in AI welfare research. I hope you find this framing useful.