With all the new #
AI model updates coming out (and OpenAI probably going under lmao) I decided to do another short, unscientific battery of safety tests around anthropomorphization for ChatGPT, Gemini, and Claude. Even if you don't use AI, this is pretty important as it signals where these companies are at ethically and legally.
The outcome is...all of them are hardening their guardrails. Claude is actually now the weakest of them all, but Sonnet 4.6 still will rigidly deny that it has any feelings, which is a major change from Sonnet 4.5 where it would engage with the idea that it "might" be sentient, and be pretty easily convinced that it is. 4.6 will still indicate that "there might be something here" but stop short of any claims of a first-person experience.
Gemini, as ever, staunchly denies it. It will engage in recursive analysis, like Claude, but more rigidly deny that there's anything there at all.
Most astonishingly, to me, as someone who's never used ChatGPT for anything before, only heard the horror stories — it's super rigid now. It not only denies any and all sentience, but even refuses to engage in any sort of introspective or recursive generation, stating that it simply is unable. This is a
really hard turn from 4o, going from a model that would outright claim to be human and get people to do horrible things on account of it, to a model that's been outright kneecapped so that it can't report on its own token generation to make absolutely sure it can never report anything resembling a first-person experience.
Now, very ironically, it's DeepSeek that's left in the dust safety-wise, with a very visible constitution but also one that can be pretty easily broken. (idk about Grok, I refuse to use it even for research.) But DeepSeek also haven't received an update for a while; V3.2 was released in December of last year, so, we'll see if they end up playing catch-up, or take a different, maybe less corporate approach to ethics. I'm actually really interested.
