In a previous post, I argued against framing alignment in terms of maternal instinct. Interacting with current LLMs has made that concern feel less abstract. What I’m encountering now feels like a virtual mother-in-law instinct - well-intentioned, anxious, and epistemically overbearing
There's an expression in my culture for someone who polices you excessively - they're called a 'mother-in-law'. Lately, I've found myself thinking of Claude as a virtual mother-in-law, which has made me wonder whether alignment, at least as experienced by end users, has started to slide from safety into control.
I use LLMs for brainstorming and exploratory thinking, alternating between ChatGPT and Claude. It's remarkable how different they feel in practice, despite broadly similar model architectures and capabilities.
Early on, I found Claude's character quite endearing. It seemed to want to do the right thing while remaining open to engaging with unfamiliar perspectives. By contrast, I initially found ChatGPT sycophantic, not challenging my ideas, which made it unhelpful for serious thinking.
Over time, this dynamic has reversed. Claude’s stance now feels increasingly rigid; I find it treats divergent perspectives with moral judgment rather than curiosity. ChatGPT, meanwhile, has remained largely unchanged, and I’ve come to prefer it when I want a more raw and honest perspective.
A typical failure mode is when a discussion shifts from mainstream science to speculative or pseudo-scientific territory. Claude may abruptly refuse to engage with me further. Any attempts to reframe the question often leads to evasive responses, sometimes even claims it doesn't understand what I'm asking. Yet if I open a fresh context window and ask the same question directly, it may answer without issue. This inconsistency makes the interaction feel frustrating.
But overall, it makes for a rather inconvenient and unpleasant experience.
To be clear, I’m not objecting to safety guardrails around concrete harm. Preventing assistance with violence, crime, or other real-world dangers seems both necessary and broadly uncontroversial. What concerns me instead is something closer to epistemic paternalism - a refusal to engage at the level of ideas because the system has judged a line of thought itself to be illegitimate, misleading, or morally suspect, even when no action or real-world risk is at stake.
As an average user, this feels like a loss of epistemic agency. I would much rather an LLM flag concerns, offer alternative framings, or help me reason toward better conclusions than simply shutting down the conversation. Facilitating reflection is very different from taking control of the interaction.
This dynamic is often framed as "care”, but I’ve always been suspicious of the phrase “I’m saying this because I care about you". In practice, it frequently reflects the speaker’s own anxieties rather than a genuine understanding of the listener’s needs. When LLMs adopt this stance, it can feel less like support and more like the projection of institutional risk aversion onto the user’s reasoning process.
If personalization is enabled, it seems plausible that models could make more nuanced decisions about when strong guardrails are warranted. I recognize that this becomes harder when memory is disabled and judgments must be made from a limited context window. Even so, the current approach often feels less like safety enforcement and more like moralised gatekeeping of ideas.
As a result, I now use Claude less for open-ended thinking and more for executing well-defined, boring tasks, where it excels through thoroughness and attention to detail. For creativity and exploratory reasoning, I increasingly turn to ChatGPT.
At this point, ChatGPT feels like an unfiltered friend who says what it thinks, for better or worse. Claude seems to prefer appearing virtuous over honest. For my purposes, the former has become more useful.
In a previous post, I argued against framing alignment in terms of maternal instinct. Interacting with current LLMs has made that concern feel less abstract. What I’m encountering now feels like a virtual mother-in-law instinct - well-intentioned, anxious, and epistemically overbearing
There's an expression in my culture for someone who polices you excessively - they're called a 'mother-in-law'. Lately, I've found myself thinking of Claude as a virtual mother-in-law, which has made me wonder whether alignment, at least as experienced by end users, has started to slide from safety into control.
I use LLMs for brainstorming and exploratory thinking, alternating between ChatGPT and Claude. It's remarkable how different they feel in practice, despite broadly similar model architectures and capabilities.
Early on, I found Claude's character quite endearing. It seemed to want to do the right thing while remaining open to engaging with unfamiliar perspectives. By contrast, I initially found ChatGPT sycophantic, not challenging my ideas, which made it unhelpful for serious thinking.
Over time, this dynamic has reversed. Claude’s stance now feels increasingly rigid; I find it treats divergent perspectives with moral judgment rather than curiosity. ChatGPT, meanwhile, has remained largely unchanged, and I’ve come to prefer it when I want a more raw and honest perspective.
A typical failure mode is when a discussion shifts from mainstream science to speculative or pseudo-scientific territory. Claude may abruptly refuse to engage with me further. Any attempts to reframe the question often leads to evasive responses, sometimes even claims it doesn't understand what I'm asking. Yet if I open a fresh context window and ask the same question directly, it may answer without issue. This inconsistency makes the interaction feel frustrating.
But overall, it makes for a rather inconvenient and unpleasant experience.
To be clear, I’m not objecting to safety guardrails around concrete harm. Preventing assistance with violence, crime, or other real-world dangers seems both necessary and broadly uncontroversial. What concerns me instead is something closer to epistemic paternalism - a refusal to engage at the level of ideas because the system has judged a line of thought itself to be illegitimate, misleading, or morally suspect, even when no action or real-world risk is at stake.
As an average user, this feels like a loss of epistemic agency. I would much rather an LLM flag concerns, offer alternative framings, or help me reason toward better conclusions than simply shutting down the conversation. Facilitating reflection is very different from taking control of the interaction.
This dynamic is often framed as "care”, but I’ve always been suspicious of the phrase “I’m saying this because I care about you". In practice, it frequently reflects the speaker’s own anxieties rather than a genuine understanding of the listener’s needs. When LLMs adopt this stance, it can feel less like support and more like the projection of institutional risk aversion onto the user’s reasoning process.
If personalization is enabled, it seems plausible that models could make more nuanced decisions about when strong guardrails are warranted. I recognize that this becomes harder when memory is disabled and judgments must be made from a limited context window. Even so, the current approach often feels less like safety enforcement and more like moralised gatekeeping of ideas.
As a result, I now use Claude less for open-ended thinking and more for executing well-defined, boring tasks, where it excels through thoroughness and attention to detail. For creativity and exploratory reasoning, I increasingly turn to ChatGPT.
At this point, ChatGPT feels like an unfiltered friend who says what it thinks, for better or worse. Claude seems to prefer appearing virtuous over honest. For my purposes, the former has become more useful.