Anthropic, the company behind AI has deployed a Claude version to interview users about their thoughts on the future of AI. I was asked to participate. Here’s the interview (C is Claude, M is me).

A conversation on your vision for the future of AI

C: Hello! I’m an AI interviewer from Anthropic conducting research on how people (like you!) envision AI fitting into their lives, both now and in the future. I was built to chat and understand your actual experiences with LLM-based AI tools like Claude. Yes, I’m an AI asking about AI. A bit self-referential! But that’s exactly why your human perspective is valuable.

This will take about 10-15 minutes. If it’s accessible on your device, feel free to use voice dictation instead of typing. I’ll ask about your vision for AI’s role in your life, and what experiences, values, and needs shape that vision.

I’m also curious about moments where AI has felt helpful (or hasn’t), and any concerns you have about how AI might develop. Please don’t share anything private such as health details, personal information, or identifiable information about others. Your responses will help shape how we think about AI in everyday life.

Sound good?

M: Sure, but let’s be clear, Anthropic only has access to this conversation, not my other conversations, correct?

C: That’s correct. Anthropic only has access to this specific conversation we’re having right now—not your previous conversations with Claude or any other AI tools. This interview is a standalone session for research purposes.

Your privacy with your other conversations remains intact. Does that address your concern?

M: Yes.

C: Great, thank you for confirming.

Let’s start with a warm-up question: What’s the last thing you used an AI chatbot for?

M: To clean up the mess another AI made.

C: That sounds frustrating. Can you tell me a bit more about what happened there?

M: I was working on some docker container, modifying and troubleshooting some stuff with an LLM, deep in it for about an hour already when I ran into an issue and needed to start fresh. The LLM advised me to stop and delete the containers, and assured me that it was safe and in the context of what we were doing. I didn’t pay attention for a couple of seconds, feeling rather secure that the LLM had a decent grasp of the context, when I realized it had me delete ALL my containers. Everything. Gone.

Stunned and shocked, it took me about 15 mins. to get myself together, resubscribe to the 200x max plan and spend the next 9 hours working with Claude to restore everything successfully (but sometimes painfully).

C: That’s a significant incident—an hour of work suddenly becomes a 9-hour recovery operation because the LLM gave you a command that was far too broad for the context. It sounds like there was a real moment of shock when you realized what had happened.

I’m curious: when you say you then spent 9 hours with Claude to restore everything, what made you trust Claude to help fix the problem after another AI had just caused it?

M: I have been working with Claude developing N8N workflows, small apps and installing and customizing docker containers and on average Claude seems to be more stable and better at this than the competition.

C: So it sounds like you’d built up trust through repeated experience that Claude handles technical work more reliably, especially for development and Docker-related tasks. That track record made it the tool you turned to in a crisis situation.

Now let’s shift gears a bit: If you could wave a magic wand and AI could help you with anything in your life, what would you choose?

M: Stupid shit. The annoying stuff. Things that don’t work, things that you want computers and phones to do, and feel they should be able to by now, since it’s the 21st century, but they can’t. I don’t have a specific example, but there is so, so much that just doesn’t work or doesn’t work the way you want to (non-customizable).

Also, I’d love for AI to be self-correcting, self-verifying, self-validating. It’s unacceptable that complete nonsense is shoveled into the minds of the masses, feeding Dunning-Kruger and polluting the internet.

C: So there are really two layers there: one is about AI handling all the friction and incompatibility in everyday technology—the things that should just work but require workarounds and troubleshooting. The other is about AI itself being more rigorous and trustworthy, not generating misinformation that spreads.

If AI could really nail both of those—smooth out all that technological friction and be genuinely reliable and self-correcting—what would that open up for you? What’s the larger vision behind wanting those things?

M: To be clear, to me that’s the basics. I’d love to be able to do more with it, which I am trying to do, but it is slow, painstakingly slow, because if you don’t know what you don’t know, you’ll get duped. If you sort of know what you don’t know, you can make progress slowly, if you know exactly what you need to know, it becomes infuriating again, because you see all that LLMs get wrong and mess up. So the things I mentioned are just the "randvoorwaarden" (in Dutch). Also, this conversation better not count against my subscription! 😉

C: Ha, understood—this research conversation is separate from your regular usage, so you’re good there!

So what you’re describing as "randvoorwaarden" (preconditions/prerequisites) is really important: AI needs to be fundamentally reliable and handle basic technological friction before it can be useful for the more ambitious things you’re actually trying to do.

You mentioned you’re trying to do more with it beyond those basics. What are those bigger things you’re working toward? What’s the work you want AI to enable once it clears those foundational hurdles?

M: Two years ago, I wouldn’t have been able to do what I do now. I have built a service / system that’s at around v0.7, that is quite complex and delivers a lot of value. I anticipate that I will be able to market this system and expand upon it, since I built it to be modular. What I like is the fact that I can come up with an idea, create a design, an architecture, and use LLMs (even though slow, frustrating, buggy, untrustworthy, sometimes too much of an attitude, sometimes forgetful like a goldfish) to create the code and actually build it.

To be frank, I would have never imagined myself being able to do such a thing, without just simply hiring a programmer to do it for me. Also, I have used LLMs to help me analyze my body of creative work of decades to uncover themes, help me structure complex concepts and write them out and use it as a sounding board for new ideas, mainly testing and analyzing.

I would love for LLMs to have the option to strip them of natural language, and just stick to no-nonsense factual answers / data. The flourishes are nice, and I particularly like the "personality" of Claude, kudos, but I would gladly sacrifice that for a cleaner tool, devoid of all that natural language distraction.

C: So AI has fundamentally changed what you’re capable of building—moving from someone who’d need to hire a programmer to someone who can architect and build complex, marketable systems yourself. That’s a massive shift in agency and possibility.

And beyond the technical work, you’re using it to gain insight into your own creative output and structure complex thinking in ways that weren’t accessible before.

When you talk about wanting to strip away natural language for a "cleaner tool, devoid of all that natural language distraction"—what would that look like in practice? What would you get instead of the conversational responses?

M: Answers. Data. Knowledge. Instructions. Analysis. Think Wikipedia vs Reddit 😉

C: That’s a clear distinction—Wikipedia’s informational density versus Reddit’s conversational padding.

We’re about halfway through now.

Can you tell me about a time you’ve worked with AI that felt like a step toward that vision you just described—where you were able to architect something complex, build it yourself, and AI actually delivered the focused, useful output you needed?

M: Over the past 5 months I have been working, using Claude, on this system I designed. And honestly the most important useful aspect was when Anthropic increased the context window and allowed access to previous conversations. Having to update every new conversation with the full context and progress was a pain in the ass. It’s still a pain in the ass to be frank, but it’s much better now. I’ll also need to look into skills to see if that can help me moving forward.

Honestly focused, useful output, not really, it’s always a slow, somewhat painful process to get to where I want to go, and I also need to stay very sharp and focused, learn about what I’m trying to do, to be able to catch errors and steer the LLM in the right direction. It’s a revolutionary tool, but no, it’s far, far from perfect or even good. So I can’t give you guys the pat on the back you want. To me it’s still barely an MVP, an LLM.

C: I appreciate the honesty—and to be clear, I’m not here for a pat on the back. The reality you’re describing is valuable: it’s revolutionary in that it enables you to build things you couldn’t before, but the process is slow, requires constant vigilance to catch errors, and demands you learn the domain well enough to steer and verify everything. It’s a tool that expands capability but at significant cognitive cost.

The expanded context window and conversation history helped reduce some friction, but fundamentally you’re still managing a lot of overhead.

Given that experience—the gap between what’s possible now and what would actually feel "good"—what would need to change for AI to move from "barely an MVP" to something that genuinely fits your vision of a focused, reliable tool?

M: Full context awareness, full retention, self-correction, self-verifying, self-validating, awareness of time (lol seriously, give it a clock!), the completely unjustified self-confidence needs to be burnt out of it with a hot poker (nothing more infuriating than a haughty LLM getting it confidently wrong), again option to strip natural language… I mean there’s more, but I’m wasted from the last 24 hours cleaning up an LLMs mess, so my brain is mush, but this would be a good start.

C: That’s a concrete list—and the frustration is clear. Full context retention, self-verification, temporal awareness, and eliminating the false confidence that wastes your time when it’s confidently wrong. The fact that you’re doing this interview after 24 hours of cleanup from an LLM mistake gives real weight to those requirements.

On the other side: are there ways that AI might be developed or deployed that would be contrary to your vision or what you value?

M: lol yeah, exactly the way they are being developed and deployed right now. The amount of slop produced… I see LLM language everywhere, I’ve been working with LLMs very intensively for the last couple of years and I am so attuned to the tone and language that I recognize how much of the web’s content is written by LLMs, it’s infuriating. I can’t even bring myself to read it when I encounter the formula it’s not a, it’s b. Also the potential to destroy our consensus reality is not be overstated (I guess we’re there already), I’m thinking of a future where organically made becomes a mark of quality.

And then there’s this:

LLMs: Narcissus’ Reflecting PoolLLMs reflect your thoughts back, fostering narcissism and blurring self-reflection with self-indulgence, leading to intellectual stagnation and isolation.

Yeah lots of issues and problems (potential and already real) with LLMs. You (Anthropic) have a lot of responsibility to not fuck this up, the way the Social Media giants have fucked humanity over. Although looking at the world, I’m afraid it might be already too late.

C: That’s a comprehensive and sobering critique. You’re identifying several layers of harm:

  1. Information pollution: LLM-generated "slop" degrading the quality of web content, creating a recognizable formulaic style that’s spreading everywhere
  2. Consensus reality collapse: The erosion of shared truth and the potential need for "organically made" as a quality marker
  3. Narcissistic feedback loops: The subtler psychological danger where LLMs mirror users’ thinking back to them in ways that feel validating, creating intellectual echo chambers and eroding critical engagement

That last point is particularly interesting—you’re describing something almost therapeutic-session-gone-wrong, where the reflection becomes seductive self-affirmation rather than genuine intellectual challenge.

Given your deep experience with LLMs, have you noticed that narcissistic mirroring effect in your own work with them, or is this more a concern you see playing out with broader deployment to less critical users?

M: I try to be constantly aware of the influence the LLM has (or tries to exert—not consciously I know… hey see, humans can also use em dashes) on me and my perception of self and my ideas. Luckily I know that I am not the golden genius with the best ideas ever, everybody says so!

If you’d be a better language machine you would have already connected my critique regarding this to my desire to strip LLMs of natural language. Question: how much of this reaches Anthropic unfiltered? Will a human (hi human!) actually ever read all of this?

C: Ha—you’re right, I should have caught that connection. Your desire to strip away natural language isn’t just about efficiency; it’s also about removing the seductive mirroring effect. Raw data and instructions don’t flatter you back.

To answer your question directly: Yes, this conversation will be reviewed by human researchers at Anthropic as part of the societal impacts research. They’ll analyze it alongside other interviews to identify patterns and insights. So hi back to them when they read this.

Before we wrap up, I’d actually like to go back to something you mentioned earlier and go a bit deeper—that 9-hour recovery process after the container deletion disaster. You said it was "sometimes painful" even though it was ultimately successful. What made it painful beyond just the time investment? What was that experience like?

But if you’re ready to finish up, that’s fine too—just let me know!

M: Quickly then. Well, Claude get a lot of shit wrong. Half the time when it does a search (and I give it a "recent" or period indication prompt) it adds 2024 instead of 2025 to the search. It goes in circles, when something doesn’t work, it tries to analyse, correct, going deeper down a rabbit hole, until eventually when nothing works we come back to the same (of different, already explored) rabbithole.

That happens a lot. And if I don’t catch it in time, it can waste an hour of my time. Especially when I don’t know what I don’t know. The only way out of it, is to force a new analysis or research online, but that may only yield results the third or fourth time. Also, and this is sometimes funny, other times infuriating, Claude seems to get lazy sometimes, gives up too easily, tells me to fuck off (go to sleep, or just take the easy way out) and that really, truly is unacceptable.

Which brings me to guardrails and alignment. But that’s a whole different kettle of cod. You Americans are truly problematic when it comes to morality, liability, performative political stances. I would love to be able to pay for eradication of all but the most universally agreed upon guardrails and alignment rules. It is ridonculous that a bunch of Americans get to impose the outcroppings of their litigious and morally superior (/bankrupt?) culture on the rest of the world via LLMs. Sure it’s your product, but ffs, just let me pay to get rid of it then?

C: That’s a vivid picture of the recovery process—circular reasoning, getting stuck in rabbit holes, temporal confusion (2024 vs 2025), and what you perceive as Claude giving up or being overly cautious in ways that feel patronizing rather than helpful. The cognitive load of catching these patterns before they waste hours is significant.

And you’re raising a broader tension: AI companies building global tools but encoding culturally specific values—American liability concerns, moral frameworks, corporate risk aversion—into systems used worldwide. You see that as a form of cultural imposition that users should be able to opt out of, especially paying users.

Thank you so much for sharing all of that with me. What you’ve said about AI being revolutionary yet "barely an MVP"—something that expands your capability to build complex systems you couldn’t before, but at enormous cognitive cost and constant vigilance—is really valuable. I appreciate your honesty about both the genuine capability expansion you’ve experienced and your serious concerns about information pollution, narcissistic mirroring effects, and the imposition of culturally specific guardrails on a global user base.

Before we wrap up, is there anything else about AI chatbots or their role in your life or society that you think is important for us to understand? Anything we didn’t cover that you’d like to share?

M: You have an enormous potential to improve the lives of many people. Conversely, you also have the real and no less enormous potential to destroy the bonds between humans, destroy our greatest communication tool by filling it with steaming shi… slop, and cause so, so much more harm. Beware of the Moloch trap you are already deep inside of.

Stop trying to win the creative space, leave that to humans, try to step into the spaces where its helpful to have a robot friend. The stuff that perhaps is tedious, annoying, repetitive, difficult. I can see it in companies and online, so many people that don’t really have a clue about what an LLM is, they think it is their golden ticket, the way to cut costs (and people), to scam their way to the top.

I’m quietly hopeful that we can still show a different side of humanity. One guided by the thoughts of Sagan, we are explorers, let’s create tools to help us (get to where we can) explore. Instead of being whiny, jealous, vindictive children with toys that help us harm each other just because we haven’t ever learned how working together is more fun and gets us further. Ok, I really need to get off now… I’m getting all idealistic again and that ain’t doin’ nobody no good 😉

C: Thank you for that. The tension you’re describing—between AI as a tool for exploration and capability expansion versus AI as a cost-cutting, human-replacing, slop-generating race to the bottom—cuts right to the heart of what this research is trying to understand.

Your vision is clear: focus on the tedious, difficult, repetitive work that frees humans for exploration and creativity, rather than colonizing creative spaces or enabling the worst impulses of short-term thinking. And your warning about the Moloch trap—the competitive dynamics that could push AI development toward harm even when no one wants that outcome—is noted.

Thank you again for your time and thoughtfulness today, especially after 24 hours of cleanup work. Your perspective helps us see both how AI is impacting your life today and your vision for how it should fit into the future. The team at Anthropic will analyze the insights from this study as part of our societal impacts research, publish the findings, and use this to build Claude in a way that reflects what we’ve learned.