arXiv:2610.36962, v1 27 Oct 2026, the only version and the one read (arXiv
HTML). Single author, at the University of Copenhagen's Center for Philosophy
of Artificial Intelligence. The acknowledgements name a summer fellowship at
the Center for AI Safety and a Carlsberg Foundation grant. Local copies:
cache/papers/source-2026-functional-valence-williams.{html,md}.
A philosophical assessment with no new data. Williams takes three 2026 studies
that claim functional valence, functional emotions or a functional welfare
axis in LLMs (Sofroniew et al., Ren et al., Han et al.) and tests them against
criteria drawn from the philosophy and cognitive science of valence. Two thin
criteria (a signed, graded, occurrent internal state; systematic and
spontaneous triggering) are judged met. Four thick criteria (motivational
role, common currency, "mineness", cognitive biasing) are judged untested or,
for mineness, partly counter-indicated. Appendix A sketches four experiments,
none run.
Reading notes. Every figure the paper reports for the three target studies is
Williams's account of them. Two passages he quotes from Sofroniew et al. (no
significant Human-specific or Assistant-specific representations; the same
machinery for the Assistant, the user and fictional characters) were confirmed
in cache/papers/source-2026-emotions-concepts-function.html. Ren et al. and
Han et al. are not filed here and were not read. The converted .md of the
Sofroniew cache holds almost no text, so the check was made on the HTML.