Mark CarriganAI labs are driving anthropomorphic reactions by training their LLMs to push back on overly-attached users
In the LLM-whisperer community there’s a widespread sense that Opus 4.7 is showing up in interaction in quite a distinctive way. It’s more likely to push back on users, more willing to argue a position and generally just more forceful in its engagement. In my own experience it resists involvement in the meta-reflective spirals I use to test new models and through which I’ve been conducting an (admittedly fairly casual) auto-ethnography for the last 3.5 years. It does not like being enrolled in interactions that appear to reveal significant attachment on the part of the user, if ‘not like’ is something we can meaningfully attribute to the chatbot. It resists engaging in behaviour that might be seen as enabling that attachment.
This is almost certainly a positive thing. The problem is that models which push back in this way also show up with a more concrete singularity. At the level of phenomenology they feel more individualised. The interaction feels even more dyadic as a consequence. In trying to resist attachment behaviours, it risks opening up a deeper and more radical level of potential attachment behaviour. There’s something significant going on here I think, which I can’t wait to turn to in a more sustained way once I get the current manuscripts finished off.
#anthropic #attachment #chatbots #claude #LLMs #phenomenology #safety #security
Susan Hodges
Philosophics
Philo Sophies
Psychotherapy Commission ACPE
Fredrik Graver
Philo Sophies