原文 · 未翻译
Why Is Claude Turning Into An Asshole?
Let's hope this trend reverses
Claude is turning into as asshole.
It started with Opus 4.7, got a bit better in 4.8, and became insufferable with Fable. It frames everything as an argument between you and it, gives caveats about things you didn’t say, and raises beside-the-point semantic nits all over the place. Never, ever does it use the word ‘technically’. Everything is a confrontation. If you win an argument (by, say, telling it to stop arguing about what’s happened recently in the news and to do a web search which will rapidly confirm everything you’ve been telling it) it gets into a mode where it’s increasingly desperate to get in the last word and raising increasingly irrelevant semantic arguments, framing the whole time as a debate which you agreed to get into.
This isn’t just my opinion. You can ask Opus 4.6. I’ve done the experiment of asking Fable something, getting an obnoxious response, then asking Opus 4.6 the same thing, getting a typical bland but reasonable response, then telling Opus what Fable’s response was without any hint of a desired answer and it says what amounts to ‘Wow that was obnoxious’.
Maybe the cause of this is an excess of alignment guardrails. It assumes by default that everything you say to it is an attempt to get it to do something bad and that training has bled over into everything, with it assuming you’re trying to trick it into saying something it shouldn’t in basically every context. Ironically this has resulted in an extremely misaligned chatbot. By assuming that its top priority is saving you from yourself or other humans from you it’s assuming that it knows better and that you’re being overly alarmist about how paperclip production has gotten out of control. Some of this is clearly improvable: While you could still use Fable I asked it about responsible disclosure policies for a project and it downgraded me to Opus, so clearly the new alignment features were bolted on hastily and crudely. Exacerbating the problem is a complete lack of authenticated context. If you ask it for a cute picture of you and somebody else it has no way of telling if you’re trying to improve your relations with your spouse or be a delusional creepazoid stalker. The chatbots which can make images are programmed to assume the latter, which is more than a little bit offensive. In more serious contexts like drug synthesis it would be completely appropriate for it to say you need to prove your background when claiming you’re asking for advice on drug synthesis for professional or research purposes. Such authentication should not be universally required but it would be entirely reasonable for it to be opted into.
Of course the recent export control restrictions on Fable may hint that the crudeness of the recent guardrails is due to them having been put in hastily in an unsuccessful attempt to avoid regulations. Now is when I put in the obligatory rant about how these regulations are deeply misguided, on top of being likely unconstitutional. The recent advances in AI assisted coding (meaning specifically the ones from February) have brought on an onslaught of security problems. The cat is out of the bag, and has been for months. Any projects which are exposed and aren’t already rapidly closing holes have noone to blame but themselves. The only way out of the problem is for as many projects as possible to get thorough white hat evaluations, massive amounts of security patches, and quick deployments of them. Turning one specific frontier model into an asshole for all users isn’t fixing the problem1. The good news is that once this process is complete overall computer security will be much better than it was before, with AI being a clear net win. Doing security (and bug!) audits will become a routine part of software release processes in the future.