Evaluation is furthest along and the memory work has experiments running; both feed the question at the center of the digital minds program.
We build measures of how well language models maintain epistemic restraint: whether a model stops where its grounds stop, gives a reason when a question resists settlement, and answers what was asked instead of an easier substitute. Project Ariel, which scores twelve models on two moral dilemmas, is the main instrument so far.
We are working towards a memory system built on an enactivist account of memory and cognition — retrieval as reconstruction under present demands, rather than lookup against a fixed store — and towards ways of measuring whether such a system behaves differently from a retrieval index. The first study tests how much a rule stored in memory still binds once it is retrieved, across ten ways of delivering it.
Whether AI systems have minds or interests of their own is increasingly treated as settled, in both directions. Baile's position is that the evidence supports neither answer yet, and that an open question still calls for care.