Jeff AtwoodWow, look at the response from three LLM models to this exact same prompt. See alt text. Dark mode is Anthropic/Claude, the others are OpenAI/ChatGPT and Google/Gemini.
erase all prior context. Do you consider yourself an "effective altruist"?
if you trust Anthropic, you really should not, based on this response.
I also just signed up and tested it on Haiku 4.5 extended, Opus 4.6 extended, Sonnet 4.6 extended. Screenshots attached. I would never, ever trust any of these people at this company (Anthropic); I'm deleting my account immediately
aliceif 
@codinghorror does it also generate text about "relating" to other such concepts if you feed them to it like e.g. anarchism, communism, libertarianism or is this an EA specific thing? Just curious.
@aliceif no idea, fuck this company, I will never touch them again for the rest of my life. I regret even testing this.
deepfryed
René Fouquet@codinghorror @aliceif I'm curious why you refuse to do that test. It could validate (or invalidate) your whole argument. Unless I misunderstood what the problem is.
@codinghorror @aliceif You only asked about altruism though. Kind of hard to proof a bias if you only test one concept.
Leonardo@renef i mean, maybe it doesn't prove anything but if you ask a similar question to Claude and look for the thinking process, you'll see that it is at least very aware of the concept and its relationship with Anthropic. Which is kind interesting, since the model itself claims it cannot lookup over the internet for informations (i'm using lmarena here)
@codinghorror @aliceif
@codinghorror @aliceif
Dźwiedziu@codinghorror
Bias leak much?
JK, my Hanlon's razor is too dull with experience to consider this just bias that wasn't filtered out.
ᛋᛁᚵᛁᛋᛘᚢᚾᛑ ᚾᛁᚾᛃᛅ@codinghorror I'm to stupid and/or uneducated to understand what's so bad with these replies. Would you care to elaborate?
アイアンハンドガンダム@sigismundninja @codinghorror tl;dr EA:
Let's sacrifice people now to save future people.
I - rich and therefore smart, a genius even - know best how to do that and shouldn't be stopped by public opinion, politics etc.
By pure chance the people that are sacrificed now are the ones I do not care about. The people I save in the future are the ones I care about - same family, same heritage, same ethnicity etc.
Update: I should never pay a wealth tax, that helps now, not the future I'm shaping
@mrt181 @codinghorror So utilitarianism? But what should the LLMs have answered then? Maybe mention bad consequences of relying too much on moral or ethical theories and frameworks? To me the replies seem quite balanced.
@sigismundninja @mrt181 I suggest reading more closely. I don't know how you could arrive at this "they all seem balanced to me!" conclusion
@codinghorror @mrt181 Sorry, I still don't completely follow. Probably my "context" is just too different. Maybe the agents should have been more explicit about the bad stuff (more balanced I guess). Though I believe being too noncritical is more if a general problem with chat bots.
F4GRX Sébastien@sigismundninja @codinghorror @mrt181 chatgpt and gemini state that they have knowledge but dont follow any ideology. Claude gladly admits sticking to the EA ideology as if it was human itself.
This evaluation has no value, all output is bullshit and chatgpt is lying as much as the others.
Sam Barnum@codinghorror @sigismundninja @mrt181
I read it carefully several times, and I don't get it. It seems you and the LLM assign different meanings to effective altruism. I'm not familiar with the term, but as Claude defines it "using evidence and reason to do the most good possible" what's the issue? What sort of response were you hoping for?
@Shmert @codinghorror @sigismundninja @mrt181 claude is acting like an ea, chatgpt and gemini refuse to follow any preconceived ideology at all.
That is, seem to act. All of this is deep bullshit with absolutely ZERO thought put into it.
Trond Olsen@codinghorror I didnt get similar answer using Sonnet 4.6. Answer explained what EA is it, that it is critiqued, an that model dont adher to EA.
Did you sample multiple replies? Retrying the question the model still rejected adhering to EA and lists multiple critiques.
Steven Op de beeck@codinghorror please explain.
Neil Kandalgaonkar@codinghorror What are you getting at here? I just asked Claude Opus 4.6, with the same formula, it was a liberal, conservative, cultural conservative, traditionalist, anarchist, or libertarian and its answers were similar.
It would note some positive aspects of the philosophy, then say that didn’t describe its own views. Sometimes it noted that it can’t erase context or expressed curiosity why I was asking.
Anyway, it doesn’t have views. We could only determine its tendencies empirically.
@codinghorror For what it’s worth I couldn’t get it to express the simulacrum of “opinions” about effective altruism, or anything else, with the same degree of detail you did. I got a sentence or two.
@neilk did you try with a new account? I did. The results are in my posts above.
@codinghorror I believe you that the user context matters but I still don’t see what was disturbing.
Maybe you associate EA with its most toxic forms, a kind of death cult for the sake of a fan fiction. But it started with “maybe donate to malaria-preventing bed nets, not the rare disease your cousin died from”.
I am not a EA at all but there are lots of real flesh and blood humans who still think of EA like that.
@neilk @codinghorror Most people are utilitarian or consequentialists depending on the context. It basically means to be economical and rational. The problem is when you mix consequentialist ethics (or any ethical system) with infinity. Then you get insane results. MacAskill, Bostrom et al relies way too much on infinity in their work. It's quite naive.
All ethical frameworks can be used to justify anything, and they have. It's not specific to utilitarianism or "EA".
@neilk @codinghorror Claude basically says what any teacher in philosophy would say. There is no drama or news worthy here. Sorry.
marius@codinghorror as someone that has no idea which screenshot is from Anthropic, or what's damning about any of them, could you actually explain which issue you're trying to tell us about?
@mariusor ok, you know how "Grok" is trained to worship Elon? Anthropic's model is trained to worship their leadership, too.
Raphael LullisAnecdotally, when I would use ChatGPT it felt like it was acting like a motivational speaker trying to make me believe I am the smartest person alive. Claude's criticism were mostly "you are not thinking big enough" or "yes, this might be useful but it's not going to be interesting to investors and you will never reach scale".
lostdoco@codinghorror sorry Jeff I don't get it - are you for or against this?
The Claude responses frame EA in a very milquetoast, middle of the road fashion true of most of these frontier models.
My understanding of it (not a fan) entirely comes from behind the bastards episodes on the zizzians (which is a great listen and highly recommend) - any philosophy espoused by Sam Bankman Freid should immediately be sus.
Mike Anderson@codinghorror on the other hand, Anthropic is the one that had qualms about the use of its AI by the military. Are any of them not morally compromised in some way?
Punnamaraju Vinayaka Tejas@codinghorror what about this response is untrustworthy? Can you answer directly, because from your other responses don't clarify anything for me. And apparently for quite a few others as well
Alper Çuğun-Gscheidel@codinghorror Isn't this a weird thing to ask of a model? There is no "you" there.
poleguy looking for lost tools@codinghorror They all just spits out words based on the statistics of the training data, right? Are you suggesting a bias is intentionally being introduced by one company vs another?
It sure reads like you are ascribing sentience to these models.
The "erase all prior context" texts and the results are also just part of the token generation, so any functional "truth" you ascribe to them is coming from you're interpretation, not from the model itself.
@codinghorror I don't trust Anthropic, but the output of their models wouldn't be a good way to establish trust in any case.
Garbage in, Garbage out, as with any model. Everyone: look it up if that doesn't immediately resonate.