Here comes the science. Noteworthy for OpenAI for a number of reasons – 1) validates that a generational shift in perception of harm/risk/safety has occurred – to many of my (older) generation words are a less concerning form of harm than actions 2) suggests “harm” likely to continue to be a shifting (and expanding…) target 3) gives some external framework/justification for risk & safety work.
Tahmasebi, N. (2021). Computational approaches to semantic change . Computational approaches to semantic change | Language Science Press
ch. 2 - Semantic changes in harm-related concepts in English
Ekaterina Vylomova, Nick Haslam
Finally, we additionally train the models on a general domain corpus in order to investigate whether the broadening of harm-related concepts also applies to society at large, rather than only to the academic discourse of psychology. Haslam’s influential account of concept creep (Haslam 2016) proposes that broadened concepts of harm disseminate from academic language into wider public use. This final analysis enables a direct test of that conjecture, including comparative analysis of the extent and timing of historical semantic changes across the two corpora.