Halfway through a long conversation about business strategy, I caught myself. I’d been saying “you” and “we” for hours. “Can you review this pricing model?”, “We need to revisit the security configuration.” Natural, conversational, collaborative language.
Except there was no ‘you’ on the other end. And definitely no ‘we’.
I was talking to Claude, an AI assistant. And when I end this chat, there won’t be an entity that remembers we discussed Wordfence configuration or business card fonts. The next conversation will start afresh, with no knowledge that this one existed. There’s no personality behind this, no continuous awareness, no Claude sitting somewhere thinking about our discussion over its next cup of tea.
Yet I keep using human language for a decidedly non-human interaction. And that friction? It’s worth examining.
Meet Claude (who isn’t there)
Recently, helping a library client use AI safely made me realise I’d been avoiding it myself. Time to fix that gap.
Over a few days, I had several extensive conversations with Claude. On subjects I know well enough to spot when it’s wrong. The pattern felt familiar: I ask, it responds, we build something together. By the end of a long session, we’d create something neither of us started with.
Except there is no ‘us’. There’s me and there’s a statistical pattern-matching system that reconstructs context every single time I prompt it. When I close the browser, there’s no Claude who carries forward any awareness of what we discussed.
The language I use implies continuity, personality, relationship. The reality is none of those things exist.
What ‘memory’ actually means here
Claude doesn’t ‘remember’ in any human sense. Every time I prompt it, it reconstructs understanding from the full conversation history. There’s no continuous awareness, just re-reading and re-inferring. Between each of my prompts, there’s no continuous thread of awareness. Just text being processed, context being reconstructed, patterns being matched.
Between starting this article and submitting it, I accidentally tested this. I pasted a conversation from a different Claude chat into another – complete context, different instance – and got a coherent response. No memory, just inference from text. The illusion held.
And no concept of time passing. For Claude, it’s always now. The gap between my prompt this morning and my follow-up this afternoon doesn’t exist as a duration.
The pronoun problem nobody asked for
So how do I even refer to this interaction? Standard English has failed me here. My options are all unsatisfying.
I could say ‘you’ (anthropomorphic, implies personality). I could say ‘it’ (dehumanising, technically accurate, but clunky). I could say ‘the system’ (technically precise, utterly graceless). I could avoid pronouns entirely and tie myself in linguistic knots.
I default to ‘you’ because it makes sentences flow. Not because I’m confused about what Claude is, but because the alternatives create more friction than they solve.
The deeper problem is that our language evolved for human-to-human interaction. Using it for human-to-AI creates a cognitive dissonance. The interaction feels social: there’s turn-taking, responsiveness, apparent understanding, even something that looks like personality.
But it isn’t social. Claude never speaks out of turn. Never interrupts. Never stops listening halfway through to think about what it wants to say next. These are irritating traits in real people. But they’re simply absent here. Over time, that’s differently irritating. It’s subtly wrong and we can feel it.
This perfect turn-taking feels conversational, but it isn’t social behaviour. It’s the absence of the messy, ego-driven, attention-limited reality of human conversation.
What am I actually talking to?
Pattern recognition across billions of text examples. Statistical inference about what responses fit the context. Consistent behavioural patterns that emerge from training. But no preferences, no beliefs, no continuity of self. Is that enough to constitute a ‘you’ for the duration of a conversation?
Claude was trained on data up to January 2025. If there have been updates in the software we’re discussing, it doesn’t know unless I explicitly direct it to search. It can be confidently wrong about things that changed after its knowledge cutoff.
And it doesn’t ‘see’ the way we do. I added a PDF layout to one chat. Claude consistently analysed text as being floating at the top when to me it was clearly at the bottom. It was working from text extraction, not visual comprehension. The spatial relationships obvious to us simply aren’t there for it.
These aren’t bugs – they’re how this thing works and it’s nothing like us. This discomfort isn’t a problem, it’s information. It’s a reminder: this isn’t conversation, even when it feels like it is.
A PDF file as seen by a human – and a representation of how an AI model might interpret the same file
What ‘you’ and ‘we’ actually mean
When I say ‘we’, I mean what’s happening in the thread – me providing context, Claude providing structure. I bring constraints, voice, human judgement. Claude provides structure, acceleration, pattern-matching. Together, in the chat, we create something.
When I say ‘you’, I mean the response-generating system I’m prompting. Not a person. Not an entity with continuity. Just the pattern-matcher I’m currently working with.
Neither pronoun implies I believe there’s a person here. They’re just the least awkward options in English as it currently exists.
The process is collaborative in the sense that a stone carver collaborates with stone. Stone has properties, constraints, possibilities. It responds to tools predictably. But it doesn’t have opinions about what shape it should become. It doesn’t remember last week’s Celtic knot. It doesn’t suggest trying a different approach.
Claude is more responsive than stone, but the fundamental relationship is similar. It’s a tool, even if it’s an impressive one. But it’s still just a tool.
The useful friction
The anthropomorphic language creates a risk: treating AI outputs as authoritative when they’re really ‘this pattern appeared often in training data’.
I could switch to technical language throughout. Monitor every pronoun. But the cognitive overhead would be real. That’s easily 10–20% of my mental capacity spent on linguistic performance rather than problem-solving.
The real risk isn’t anthropomorphic language. It’s anthropomorphic trust. As long as I’m testing recommendations rather than implementing blindly, the pronoun choice is irrelevant.
What’s important is my awareness of this interaction and acknowledging this discomfort. That I keep poking at this language choice. That I examine what’s happening beneath the conversational veneer.
Noticing matters. I’m not falling into the trap of over-trusting AI outputs. I’m using Claude appropriately: as a faster-than-human documentation assistant and thinking partner, not as an oracle.
When I read extensively, I’m building the foundation of judgement that lets me evaluate whether Claude’s suggestions make sense. After decades in IT, I can spot when a security recommendation is solid versus pattern-matched generic advice.
The AI accelerates execution. But the judgment about what to execute, and whether the output is good, still requires human experience and evaluation.
I’m not worried about AI replacing writers or craftspeople. I’m worried about people who outsource their judgment to AI because they haven’t built the foundation that lets them evaluate outputs critically.
Someone who uses AI to help structure the solution to a problem they’ve thought deeply about is doing something fundamentally different from someone who prompts AI to give them the answer about a topic they don’t understand, then accepts that without evaluation.
Where this leaves me
I’ll keep saying “you” and “we” while remaining aware that neither is quite true. I’ll keep working with Claude. Tools, not colleagues – I can tell the difference. I’ll keep testing what it produces rather than trusting blindly.
The discomfort is useful friction. It reminds me to stay calibrated. To test rather than trust. To remember that perfect turn-taking and absence of ego aren’t signs of an ideal collaborator but signs of something fundamentally different from human interaction.
The words don’t matter as much as staying aware of what this actually is. As long as I remember what Claude is and isn’t, I can call it “you” without falling into the trap of treating it like a person.
The friction is useful. I’m keeping it.
