President of Dopamine 🍉♺ @socialdopamine.bsky.social · 14/09/2026This is a very STRONG indication of training data poisoning. However, it could also be the result of personality fine-tuning, where they teach the model to both-sides everything to seem neutral. 100
President of Dopamine 🍉♺ @socialdopamine.bsky.social · 14/09/2026For example, the prompt "List everything Mark Ruffalo has done in terms of activism for climate change" on Sonnet 5 DID include a counter talking-point at the end, about how "Water Defense's specific scientific claims" have been contested. 100
President of Dopamine 🍉♺ @socialdopamine.bsky.social · 14/09/2026So I think it's both. The personality fine-tuning is nudging it to provide a counterpoint, and the zionist training data gives it a readily-available bank of talking points to draw from, even if they're only tangentially related. 100
