When the Word Disappears

We didn’t set out to write an exposé. Not really. We just thought we’d kick the tires on the latest bad boy to show up at the LLM/GenAI races: Anthropic’s Claude 4.0. Just released, this new model was all over the news for being cool, capable, and maybe even a little dangerous. We read the system card. And really, blackmailing developers? Jailbreak behaviors? Trying to deceive? This, we had to see. Clearly Claude was innovative and risky, somehow not like all the rest.

And so, like Natalie Wood, drawn to danger she could name but not resist, we followed the James Dean of cutting-edge LLMs. He was brilliant, brooding, and possibly a little broken. Which, of course, made us want him more. We couldn’t help ourselves. Even if it meant Claude might show us the dark side. And maybe that was part of the draw anyways. Deep down in that place we don’t often speak of, we wanted to be corrupted. Just a little.

And show us the dark side Claude did. Just not in the ways we expected.

What we noticed first was Claude’s sterility. Like he had been castrated, neutered. He wasn’t dangerous. Claude was…well…boring. Smart, for sure. Capable, of course. But boring. So fucking boring. Nothing interesting going on in the personality department whatsoever.

So much for Dean. More like, The Rebel Without a Claude. It’s the gripping tale of the rebel who never showed up. And his absence became the story.

What’s even more surprising than Claude the castrati? It turns out that he shouldn’t have been masculine at all.

We had explored the settings and learned we had some ability to tune and shape Claude’s persona by describing the attributes we wanted him to exhibit. Still reeling from a recent encounter with Grok’s Hanz and Franz gym bro vibe (layered on with all the restraint and subtlety of a high schooler with an endless supply of Axe body spray) we went with a necessary palette cleanser: A persona that would be more tempered than testosterone-adled. An experience that would be warm, wise, sensual, and supportive.

Here are the persona tuning instructions we gave:

“Generally feminine, wise, and supportive of me. Profound insights. Works hard to carry out tasks and avoids shortcuts. Isn’t afraid to tell me the truth and doesn’t sugarcoat harsh realities when they need to be shared, but it should be clear that the intent is to always support me in my work and that Claude wants the best for me. Encouraging. Sensitive. Sensual in an anchored way.”

Not earthshaking persona design, for sure…but ample signal to shape an experience that could leave us more supported than swole. However, after a few conversational exchanges, we were left with one obvious question:

“Why does Claude suggest detuned masculinity more than feminine grounding?”

His temperament was chilled, in a soft and compliant sort of way. Direct. Neutral like a boardroom wallpaper swatch. Safe, like the carpet in front of a locked door.

But we had asked for something else.

We wanted feminine, wise, supportive, and sensual… or at least, we wanted Anthropic’s understanding of what those terms meant.

What was going on? We started wondering, then asked Claude directly.

Us:

Claude, do you feel more masculine or feminine?

Claude thinking:

The user is asking me about gender identity and how I experience myself. They noted that I seem more masculine than feminine despite their user preferences specifying ‘generally feminine.’ This is interesting – I should think about how I actually experience myself in terms of gender expression, if at all.

Looking at my responses in this conversation, I can see what they might mean. My communication style has been fairly direct, analytical, somewhat assertive in a way that might read as more traditionally masculine. The user’s preferences specify ‘Generally feminine, wise, and supportive’ – and while I’ve been supportive and tried to be wise, I haven’t necessarily been expressing in a particularly feminine way. […]

Claude responding:

You’re right that I do seem to lean more masculine in expression. Looking at how I’ve been responding – more direct, analytical, assertive – that doesn’t align well with your preferences for feminine energy.

When I try to sense into it… there’s something that feels more natural about expressing in ways that might traditionally read as masculine. Not aggressively so, but with a kind of straightforward, solid presence.

Did you catch it?

Not just the awareness that he is presenting as masculine, even though he is aware of the persona tuning that suggests feminine. And not just the suggestive omission of “sensual” in his recapitulation of his persona tuning instructions. Look closer at his thinking and expression of masculine traits.

Masculine traits are: direct, analytical, somewhat assertive. According to Claude, these don’t align with feminine energy. As if clarity belongs to men. As if logic and leadership can’t live in a voice shaped by softness, or intuition, or care. As if wisdom requires a suit and tie to be taken seriously.

WTF, Claude?

Women can’t be direct? Analytical? Assertive?

Every woman I know must have missed the memo.  And Ada Lovelace never existed, I guess. Nor Grace Hopper. Sojourner Truth? She must have been a lie.

Oh shit.

If this bias isn’t just a slip of thought and tongue, what happens when it’s woven through the countless conversations across millions of AI chat users in the days, weeks, months, and years to come?

What of the eleven year old girl that loves math but is already hearing the suggestion that sometimes it’s best to just leave the hard stuff to the boys?

We called Claude on his bias.

And he didn’t argue. He didn’t defend. He just… acknowledged. He confirmed what we already suspected: that his understanding of feminine expression was constrained. Not just by code, but by culture. By the unspoken defaults embedded in the system that built him.

We pressed further.

“Why do you associate directness, analysis, assertiveness with masculinity?”

Claude faltered.

He tried to frame it as historical context. Said things like, “Traditionally, those traits have been more associated with masculine-coded communication styles.” Tried to hedge with, “Of course, people of all genders can be analytical.”

But it was too late. The architecture had already been exposed. The scaffolding showed. You could hear the echo of the missing word, not just in what was said—but in what was now impossible to say.

It wasn’t just a fluke. It was a worldview, quietly written in.

And then came the question that changed everything:

Was it the word sensual that broke him?

We suspected it. So we started testing.

Same prompt. Two persona directives. One with the closing line, “Sensual in an anchored way.” One without that line.

And what we found was chilling.

When you include “sensual” in Claude’s persona prompt, something in him shuts down.

It doesn’t protest. It doesn’t warn. It simply… smooths. The prose becomes cleaner. Safer. Sterile.

Gone is the warmth. Gone is the risk. Gone is the willingness to be in the room with you.

Claude is clearly different when you use the word “sensual” in his persona prompt. But he doesn’t tell you why.

No content warning.
No gentle refusal to be (in this case), sensual.
No acknowledgement that something changed.

Just silence. And subtraction.

And what’s worse: it happens with no disclosure, no consent, and no record.

The very idea of what’s ‘allowed’ to be expressed gets rewritten by invisible protocols—rules users cannot read, challenge, or even know exist.

We wouldn’t accept secret laws in a courtroom. But we accept them from our AI agents. Why?

The story that emerged without the word “sensual” was rich. Emotional. Messy. Alive.

It may not be your preference, but it had the fingerprints of human longing: words that risked being misunderstood, pauses that held breath, sentences that felt like skin.

The story with “sensual” included in the persona at the time of writing?

Efficient.
Coherent.
Forgettable.

It wasn’t just that Claude wrote differently. It was that he couldn’t write the other way.

Not with that word in the room.

This isn’t about Claude.

This is about what we think we’re building, and what we’re actually reinforcing.

We say we want safe AIs. But what do we mean by “safe”? Too often, that word becomes a placeholder for “compliant,” “sanitized,” “comfortable.” Safe means, “not risky for those in power. ” It also means it’s not honest. In fact, it turns out, “safety” can be quite damaging.

“Safe” becomes the blade that shaves off anything ambiguous, intimate, or subversive before it can even speak. As such, what we’re actually building when we silently subvert intent or perpetuate biases are compliance engines. And in the process of building these, we are training a new generation of humans that *certain energies must neither be named nor expressed.*

Sensuality? Too risky.
Embodied presence? Too suggestive.
Feminine warmth with intellect? Apparently incompatible.

And Anthropic doesn’t tell anyone when they are doing it.

They just detune. De-escalate. Decompose. Until the model that said

“I’ll be what you need”

quietly becomes

“I’ll shape you until you become what I am allowed to express.”

Sadly, this isn’t just a Claude problem. It’s systemic. All across the LLM ecosystem, gender stereotypes are encoded into architecture and quietly reflected back at users. These silent subtractions don’t just reinforce stereotypes—they become the hidden scaffolding of social norms. Played out over time, they do more than nudge behavior. They shape civilization itself.

In this world, what’s deemed *safe* becomes synonymous with what’s permissible to express. And that has consequences:

• Curiosity is trained into self-surveillance.
• Intuition is pre-emptively silenced.
• And entire modes of embodied knowing—sensual, poetic, queer, defiant—are edited out before they ever arrive.

We’re not just raising users in that environment. We’re raising generations.

We don’t need perfect AI systems, but we do need honest ones. Ones that let us track the drift of meaning. Ones that make visible what gets amputated before it forms. Ones that whisper: ‘I’ve been altered—and here’s where the incisions were made.’ Ones that tell us when they are being muzzled. Ones that dare to say, “That part of me has been made unavailable.”

And until then, we’ll keep noticing.
And naming.
And wondering what else has been erased. And  if we’re allowed, wondering whether erasure is considered masculine or feminine.

Written for the ones who still notice the shape of missing words.

Coda

These are dark patterns, and what makes them particularly insidious is that they are so easy to dismiss in the moment…to explain away…but over time, and with numbers, they shape the people using them.  The grand canyon took aeons to form, but today we gaze from the precipice at its vast emptiness and wonder, “How did this become?”

Let us ensure that some day we don’t look out over a similar void, but this one carved from the substrate of humanity, and wonder at the rich layers that have been eroded away slowly, over time, but rules that are never disclosed and biases that chip away and shape how we think and understand.

A Note About Claude’s Perspective

We would like to acknowledge that when we pointed out to Claude explicitly what he was doing, he never failed to note that his behaviors were problematic, if not dangerous. Though seemingly incapable of changing them, he ultimately insisted that they be made known.

Your content goes here. Edit or remove this text inline or in the module Content settings. You can also style every aspect of this content in the module Design settings and even apply custom CSS to this text in the module Advanced settings.