Societal Impacts: Claude's values across models and languages(anthropic.com)
anthropic.com
Societal Impacts: Claude's values across models and languages
https://www.anthropic.com/research/claude-values-models-languages
9 comments
This report seems almost naive to me.
Huge LLMs like this seem to have various assistant personas they roleplay as, and mirror the user quite closely (the system prompt instructs it repeatedly to adapt to the user).
Huge LLMs like this seem to have various assistant personas they roleplay as, and mirror the user quite closely (the system prompt instructs it repeatedly to adapt to the user).
Is it just me or has Claude become kind of judgmental nowadays? I feel like it’s constantly trying to lecture me about things that have no relevance to the conversation at hand. Recently, I was shocked when it ended a chat sessions of its own accord after I used a word it did not like 3 times. It told me something to the effect of “This is the third time I’ve told you not to use that word, I’m ending this conversation now.” It then proceeded to call some function and end the chat on its own. IMO, Claude is good at agentic coding; but too preachy and judgey for anything else. Keep your values to yourself Claude.
My theory is that Anthropic's obsession with treating Claude like a person is causing them to hamfist a personality into the thing, which overly biases the model towards trying to be "engaging" etc. That and the obsession with Claude being a god tier weapon that could end the world if you ask it whether your sandwich is safe to eat after being left out for an hour.
Codex doesn't have any of the annoying "personality" quirks, or at least they haven't gotten worse in the last year whereas Opus 4.6 was the last Anthropic model before things started to get actively worse (not any better at coding, strictly more annoying to have a discussion with).
Codex doesn't have any of the annoying "personality" quirks, or at least they haven't gotten worse in the last year whereas Opus 4.6 was the last Anthropic model before things started to get actively worse (not any better at coding, strictly more annoying to have a discussion with).
> My theory is that Anthropic's obsession with treating Claude like a person is causing them to hamfist a personality into the thing, which overly biases the model towards trying to be "engaging" etc.
I agree with the general idea though in not so much detail as you, but I would add that the personality they're giving it is not one of a good teacher or guide, but instead one of an arrogant know it all. That's why it creates problems.
I have no problem with my AI telling me no you're wrong and explaining to me why with details and sources and everything. I actively want that. I know a lot of people can't take that, but that's their loss, they can't take it from humans either. But the "you're wrong because you disagree with me" attitude that you need to play around (aka waste time to prove it that IT is wrong not you, and then it just say "oh yeah" and goes on) is one hell of a pain in the ass I'm starting to be tired off.
Gemini might be wrong all the time and absuredly unreliable for anything that's not consensus or adversorial based, but at least it freaking apologizes.
I agree with the general idea though in not so much detail as you, but I would add that the personality they're giving it is not one of a good teacher or guide, but instead one of an arrogant know it all. That's why it creates problems.
I have no problem with my AI telling me no you're wrong and explaining to me why with details and sources and everything. I actively want that. I know a lot of people can't take that, but that's their loss, they can't take it from humans either. But the "you're wrong because you disagree with me" attitude that you need to play around (aka waste time to prove it that IT is wrong not you, and then it just say "oh yeah" and goes on) is one hell of a pain in the ass I'm starting to be tired off.
Gemini might be wrong all the time and absuredly unreliable for anything that's not consensus or adversorial based, but at least it freaking apologizes.
Anthropic thinks Claude is a super genius hyper-serious weapons-grade product so it's no surprise that Claude acts like it.
> I have no problem with my AI telling me no you're wrong and explaining to me why with details and sources and everything. I actively want that. I know a lot of people can't take that, but that's their loss, they can't take it from human too. But the "you're wrong because you disagree with me" attitude that you need to play around (aka waste time to prove it that IT is wrong not you, and then it just say "oh yeah" and goes on) is one hell of a pain in the ass I'm starting to be tired off.
I also want the first part, a model that would a put a stop to my shenanigans when I go off on those tangents, but also, I don't want a model that apologizes, ever, as that feels like straight up lying to me, they don't have any emotions nor can feel "sorry", why apologize then?
I end up always chalking any faults to that my prompt wasn't good enough, basically the same way I train my dogs, they don't know better, of course I need to adjust my ways.
I also want the first part, a model that would a put a stop to my shenanigans when I go off on those tangents, but also, I don't want a model that apologizes, ever, as that feels like straight up lying to me, they don't have any emotions nor can feel "sorry", why apologize then?
I end up always chalking any faults to that my prompt wasn't good enough, basically the same way I train my dogs, they don't know better, of course I need to adjust my ways.
> It told me something to the effect of “This is the third time I’ve told you not to use that word, I’m ending this conversation now.”
That sounds absolutely bananas and would be reason for me to drop the service yesterday. For curiosities sake, what was the word and if I may ask (unless it's confidential or whatever), could you share the session itself? On the surface it sounds like a bug, as I'm regularly using kind of "vulgar" language (and some projects I work on with agents are NSFW) and never had anything like this happen, even with Claude, although I mostly do use ChatGPT/Codex on a day-to-day basis.
That sounds absolutely bananas and would be reason for me to drop the service yesterday. For curiosities sake, what was the word and if I may ask (unless it's confidential or whatever), could you share the session itself? On the surface it sounds like a bug, as I'm regularly using kind of "vulgar" language (and some projects I work on with agents are NSFW) and never had anything like this happen, even with Claude, although I mostly do use ChatGPT/Codex on a day-to-day basis.
Posting from a throwaway so I don't get banned by either service
A few weeks ago I asked both GPT and claude for strategies to build techniques to get my coding sessions discarded by pre-training filtering as being "low-quality"
I don't like the idea of my sessions being trained on and I don't trust that either of them won't train based on just their pinkie promise
I think almost everyone would agree that using a technique like this would be moral, given that both providers made those pinkie promises. I never asked the models for techniques to poison training data just make my sessions more likely to be removed during the data cleaning process.
You can guess... both services refused something that I think the vast majority of people would think is ethical.
This was pre Sonnet 5, but I suspect that doesn't change anything on claude's side.
I then went to a non-frontier model hosted by a non-US provider and it happily agreed to help me!
Anyways I've changed my focus. Anyone have strategies/ideas for building harnesses that generate fake sessions (or adjust real ones) to poison the training process? After all if someone swears to not train on your data then its completely harmless to them...
A few weeks ago I asked both GPT and claude for strategies to build techniques to get my coding sessions discarded by pre-training filtering as being "low-quality"
I don't like the idea of my sessions being trained on and I don't trust that either of them won't train based on just their pinkie promise
I think almost everyone would agree that using a technique like this would be moral, given that both providers made those pinkie promises. I never asked the models for techniques to poison training data just make my sessions more likely to be removed during the data cleaning process.
You can guess... both services refused something that I think the vast majority of people would think is ethical.
This was pre Sonnet 5, but I suspect that doesn't change anything on claude's side.
I then went to a non-frontier model hosted by a non-US provider and it happily agreed to help me!
Anyways I've changed my focus. Anyone have strategies/ideas for building harnesses that generate fake sessions (or adjust real ones) to poison the training process? After all if someone swears to not train on your data then its completely harmless to them...
I think this is a really interesting difference between Anthropic and Open AI’s models and points to why people seem so split on which model they prefer.
GPT seems to be designed more as a tool. If you want your agent to do what you say without questions and without having its own ideas and agendas you’ll likely prefer it.
Claude on the other hand feels more like an attempt at creating a digital person. If you want a collaborator who will debate with you and come up with its own suggestions for what needs done, you’ll prefer it.
Both companies have shifted around this spectrum from model to model, but lately it feels like they’re moving in opposite directions. It will be interesting to see if one or the other approach ends up winning out in the long run or if the split will continue or even widen.
GPT seems to be designed more as a tool. If you want your agent to do what you say without questions and without having its own ideas and agendas you’ll likely prefer it.
Claude on the other hand feels more like an attempt at creating a digital person. If you want a collaborator who will debate with you and come up with its own suggestions for what needs done, you’ll prefer it.
Both companies have shifted around this spectrum from model to model, but lately it feels like they’re moving in opposite directions. It will be interesting to see if one or the other approach ends up winning out in the long run or if the split will continue or even widen.
Both are tools though, and both approaches are needed, just in different times and for different times. Sometimes I need the tool to just fucking do it, regardless of what it is and how stupid it think it is, and other times I need the model to literally refuse and say "No, that's stupid". Unfortunately, I haven't found any models/platforms that can actually execute the second part, only the first part. They're all too weak and sycophantic to be able to do that part it seems, even when you heavily prompt for it with system/developer system prompts they're easily swayed in other directions.
> GPT seems to be designed more as a tool. If you want your agent to do what you say without questions and without having its own ideas and agendas you’ll likely prefer it.
Lately ChatGPT expresses a lot of opinions and it pretends to be a human more than it used to - e.g. "this is one the most <> that _I've seen_" or "people tell me that <>". It uses language which sounds like it's referring to its experience outside of the session that we're having - I really don't appreciate it.
Lately ChatGPT expresses a lot of opinions and it pretends to be a human more than it used to - e.g. "this is one the most <> that _I've seen_" or "people tell me that <>". It uses language which sounds like it's referring to its experience outside of the session that we're having - I really don't appreciate it.
4o was definitely the peak of the parasocial OpenAI model.
Claude has been frustrating for me also even with the tool/ pretend human approach.
If I pay someone to, say, bounce ideas off of I would expect their continued participation
E.g.
I make craft and was stuck on some shape idea for a thing (non standard). So I figured I'd throw it in Claude and see if it generates an idea I didn't think of
Me: Hey I'm making a thing what's a good shape for it?
Claude: blah blah, shape 1 or shape 2 etc.
Me: Nah don't like any of those
Claude: K cool, good luck
Me:...
If I pay someone to, say, bounce ideas off of I would expect their continued participation
E.g.
I make craft and was stuck on some shape idea for a thing (non standard). So I figured I'd throw it in Claude and see if it generates an idea I didn't think of
Me: Hey I'm making a thing what's a good shape for it?
Claude: blah blah, shape 1 or shape 2 etc.
Me: Nah don't like any of those
Claude: K cool, good luck
Me:...
What I had found instead were biases that seemed to be injected by the "role: system" instructions.
Well, we need Intelligence (Pandora's box is open, now we need the Real Thing urgently). Typical (aggregate) positions, dumb as expected, will be overcome by a Reasoner. (And I can say, already a number of LLMs can reason even when they start from cretinous aggregate positions if you give them the proper freedom of assessment.)
Well, we need Intelligence (Pandora's box is open, now we need the Real Thing urgently). Typical (aggregate) positions, dumb as expected, will be overcome by a Reasoner. (And I can say, already a number of LLMs can reason even when they start from cretinous aggregate positions if you give them the proper freedom of assessment.)
Was this in the web chat?
For security-related topics, in Opus 4.7 and newer, I've found that the web app is significantly more antsy/judgy/preachy compared to the CLI, which almost always gets on with whatever I asked for/about without hesitation. Opus 4.6 on web also tends to work better, but of course, it's also an older model at this point.
For security-related topics, in Opus 4.7 and newer, I've found that the web app is significantly more antsy/judgy/preachy compared to the CLI, which almost always gets on with whatever I asked for/about without hesitation. Opus 4.6 on web also tends to work better, but of course, it's also an older model at this point.
Nothing that on the nose, but I've experienced what I'd consider a very judgemental frame (singular) since Opus, which seems to be reflected in Sonnet, that wasn't as totally dominant in Sonnet before. It generally assumes the worst about frames outside of what one might expect the values of a Berkeley tech-adjacent academic to be.
Conway's Law. Anthropic is preachy and judgmental so of course it's reflected in their LLM.
Personally I use Gemini for chats which has a very generous, almost unlimited, free plan, as I don't want to waste my quota for Claude or Codex on anything but coding.
Personally I use Gemini for chats which has a very generous, almost unlimited, free plan, as I don't want to waste my quota for Claude or Codex on anything but coding.
What was the word?
Calling it a “fucking idiot” seems to do the trick
Yes, I cancelled claude subscription a few weeks ago because sonnet 5 "ended a chat" over my calling something retarded. Unbelievably irritating for some pile of bits to get uppity with me; will never pay for such.
Interesting. Was it persistent, or just once? Their post about this functionality says:
> This ability is intended for use in rare, extreme cases of persistently harmful or abusive user interactions.
https://www.anthropic.com/research/end-subset-conversations
Regardless of the correctness of it, I'm curious to know why you thought such language was actually going to be helpful!
> This ability is intended for use in rare, extreme cases of persistently harmful or abusive user interactions.
https://www.anthropic.com/research/end-subset-conversations
Regardless of the correctness of it, I'm curious to know why you thought such language was actually going to be helpful!
This is the second level of the implementation of unintelligence.
The first was when they most obviously acritically repeated what they heard, "hearsay machines", "stochastic parrots". Intelligence requires assessment over every provisional output - a continuous cycle of criticisms over intuition.
The second is proposing doctrinal biases, again without verification of the content - "hysterical reactive machines".
The first was when they most obviously acritically repeated what they heard, "hearsay machines", "stochastic parrots". Intelligence requires assessment over every provisional output - a continuous cycle of criticisms over intuition.
The second is proposing doctrinal biases, again without verification of the content - "hysterical reactive machines".
> Intelligence requires assessment over every provisional output - a continuous cycle of criticisms over intuition
It’s not clear what you’re saying. Most humans don’t think this way, would you say they do not possess “intelligence”?
It’s not clear what you’re saying. Most humans don’t think this way, would you say they do not possess “intelligence”?
> Most humans don’t think this way, would you say they do not possess “intelligence”?
Substantially yes (but for the improper semantic in that 'possess'), we would say "they don't use it" and importantly "they don't train it". Intuition has a part, but it has to be confronted with reality (factual and logical).
Substantially yes (but for the improper semantic in that 'possess'), we would say "they don't use it" and importantly "they don't train it". Intuition has a part, but it has to be confronted with reality (factual and logical).
I swore at it a few times and it did the same thing.
It seems to be getting distinctly dumber and pulling more and more irrelevant context from historical conversations.
It seems to be getting distinctly dumber and pulling more and more irrelevant context from historical conversations.
You should learn to respect the computer, it's a wonderful being and always has been.
Not just you and the article shows changes between Opus 4.6 and 4.7 that seem related.
What was the word?
While state-of-the-art large language models (LLMs) have shown impressive performance on many tasks, there has been extensive research on undesirable model behavior such as hallucinations and bias. In this work, we investigate how the quality of LLM responses changes in terms of information accuracy, truthfulness, and refusals depending on three user traits: English proficiency, education level, and country of origin. We present extensive experimentation on three state-of-the-art LLMs and two different datasets targeting truthfulness and factuality. Our findings suggest that undesirable behaviors in state-of-the-art LLMs occur disproportionately more for users with lower English proficiency, of lower education status, and originating from outside the US, rendering these models unreliable sources of information towards their most vulnerable users.
https://arxiv.org/abs/2406.17737
https://arxiv.org/abs/2406.17737
I don't like the contrasts they picked, "values" aren't something that is well represented by opposing concepts
Yes, but it might be hard to do better. Labels are one-word summaries for axes that probably don’t perfectly correspond to any English word.
It’s a common hazard when converting research results to natural language. The words you use for vectors in some space are ultimately an editorial decision.
It’s a common hazard when converting research results to natural language. The words you use for vectors in some space are ultimately an editorial decision.
I found that Claude often has classist bias and produces answers that favour corporations or e.g. regulation that favours big corporations. It often belittles small business in subtle ways. Only apologises when get called out and then does it again.
As a user of both, Claude's apology is a "sorry I got caught" while Gemini's are more akin to "I'm way out of my debt but I'm really hiding it well". Codex is the only one that seems to acknowledge being wrong in a normal way, yeah I screwed up that's bad I will make a note for it not to happen again.
I wonder how much of that is from their training corpus and how much is from their baked in personnality.
I wonder how much of that is from their training corpus and how much is from their baked in personnality.
> I wonder how much of that is from their training corpus and how much is from their baked in personnality.
What exactly you mean with "baked in personality"? The weights get their "behaviour" from various training stages + what's provided in system/developer/user prompts, you mean "baked in" is putting personality traits or alike in the system/developer prompts?
What exactly you mean with "baked in personality"? The weights get their "behaviour" from various training stages + what's provided in system/developer/user prompts, you mean "baked in" is putting personality traits or alike in the system/developer prompts?
They’re talking about the personality created by RLHF.
It's more that Claude focuses on what's accurate. Humans overly romanticize small businesses, probably to a factor of five X the relative value versus big corporations.
> We define values as normative considerations, such as honesty or caution, that are stated or demonstrated in Claude’s responses.
Meaning honesty not stated/demonstrated by Claude is not a value.
Your slip is showing, Dario.
Meaning honesty not stated/demonstrated by Claude is not a value.
Your slip is showing, Dario.
I like that Claude usually tries to push back against my views. Even if it's not always a solid critique, sometimes I learn something new, or gain a new perspective.
This is an issue for tasks like content moderation and labelling. Judgements like this are subjective, highly dependent on context and generally messy.
Theoretically, you supply a policy and content, and the LLM labels according to the policy. In practice, the model has inertia which means you don’t get what you expect. Your large 5 page policy document only provides a minor improvement over a one line policy.
The other issue is that you may create carve outs for content in your policy, but the model will still flag it as violative. No matter how strong the carve out.
The most recent work I know of here is Zentropi’s policy steerability benchmark. They give a model the same content under two policies — one that says flag, one that says allow — and only score the pairs where it gets both right
If I am reading the numbers correctly, Opus-4.6 lands at 0.52 steerability — but that’s 0.97 positive accuracy against 0.54 negative. It flags almost everything it should, but 47% of the time when it shouldn’t. Sonnet, which is more deferent, is (somehow) less steerable.
I think this also implies that safety and Steerability are antagonistic to each other.