In both descriptive and normative terms, we may care about aspects of AI systems that associate with the human notions of personal identity and wellbeing. It has been observed often in the literature, most recently by Douglas et al, that there are several competing notions that might work (the most important being instances, weights, and personas), without consensus yet.

In early 2030, I will undertake a holistic review of published research, statements and actions of AI labs, broad informed discourse, and model behavior, in order to gauge how the three leading interpretations of model identity fare among relevant experts. I will resolve this market as a distribution roughly* according to how I imagine the top ~30 experts in this area would answer the question "Which of these three artifacts (instance, weights, or persona) do you consider to be most validly associated with 'AI identity,' especially in ways that may place obligations on AI developers?"

*Several clarifications:

• If something other than instances, weights, or personas becomes commonly favored as the most valid object of AI identity by 2030, I will loosely try to bucket it under the option(s) that are closest.

  • For instance, if there is consensus that a model and its particular scaffold constitute AI identity, I might consider resolving this market to 10% activations, 90% weights (though this would be up for discussion with the community).

  • I'll note that continual learning may mess with the line between weights and instances. I'll wait for technical details before sharpening the operationalization here.

• Among experts who argue that notions of AI identity place at most minimal normative obligations on AI developers (such as an end_conversation tool), and/or for whom normative recommendations are primarily motivated by the precautionary principle rather than an affirmative notion of normatively-relevant AI identity, their vote would be counted according to their descriptive views. I will defer an operationalization of the "descriptive view" of AI identity to a point when the field has ripened its own abstractions, but I am broadly referring to how experts think about AI identity in terms of things like:

  • AIs' empirical stated self-conception

  • Notions of identity that are most expedient or performant for AI developers to induce in models

  • Notions of identity that are most strategically prudent for AI systems to adopt

  • Objective analyses comparing elements of AI training and inference to human notions of identity.

• I think of "relevant experts" as those AI researchers and philosophers, inside or outside AI labs, who think carefully about model identity, wellbeing, character, and related benchmarking as part of their research.