Martin Hähnel

Actual Hesitation

Published:
Tags:

I will stop using AI for my personal use.[1] This has been a crazy journey to get to this decision. I'd like to think that I am reflected and that I really take things seriously if they are supposed to be taken seriously. I have written repeatedly about my AI-use, my interest in LLMs and how I view the ethics around AI (check out Hub "AI and LLMs").

In the more recent past I had stopped blogging as much, mostly because I didn't really have the time to do so (becoming a Dad), but that didn't mean I stopped thinking, reading and writing about AI in private. I will try to describe my position such as it was in as few words as I can next:

The central argument I tried to make as regards to LLMs is that the way we use the technology makes all the difference. This means two other things: 1. I judge the morality of engagement of LLMs on a personal, individual level. What are you using it for? A real use case like coding? As a mirror for reflection? Fine. As a way to scam people or making the world a worse place? Not fine. That seemed to me to be a pretty clear, pretty good criterion to judge LLM engagement. 2. This therefore doesn't really talk about anything else but use.

The background here is that I am careful to scope moral expectations to an individual. In that sense use is something we can judge individually and so is critical engagement with the news, critical commentary and other things about the LLM-industry - I didn't talk about the latter directly but use without reflection about what it is you are using, is just ignorance. For me, it always made sense to think about the things a single actor can do: I can't change society but I can start or sign a petition though. I can protest. I can write a letter, I can donate. I can use LLMs for actual good use cases with the aim to do better work. I am not allowed to use LLMs to make other people's lives worse. That kind of stuff: individual, actually doable things (or things you as an individual can avoid doing).

I have also spoken for allowing a little bit of elasticity in moral judgement. As humans, we will at times overstep and within reason, that's fine. I have also talked about the relative ease and difficulty to change your own behaviors. And that a moral judgement should not only take into account a person's situation but also their moral constitution or ability to actually stick to their values, which is easier for some than others.

The above is another way to say: If changing yourself is hard changing even one other person is even harder. And to change a whole society to do better seems hopelessly optimistic to me (hence my fatalism). Historical developments like the LLM bubble influence us more than we do them. We are in a way just the expression of the historical here and now. It may feel different, but what acts, acts (mostly) through us. That we are still acting and the fact that we do have agency is not a contradiction. It is just what we do, how we view the world and how we judge all of that and others is very much the result of our environment.

Over the last few years I have observed that the capabilities of LLMs did improve, because programming tools as a market was discovered that can actually bring in some money and so around fall of 2025 there was a noticeable difference in the quality of LLM output for coding. There have also been improvements in chat. Maybe I just got better in using these tools but even the chat without a coding harness has been helpful:

That I have found these things (coding + the chat interface) so useful also meant that I liked using LLMs genuinely. I found the technology engaging and interesting and working with it as a tool to solve other problems intellectually stimulating.

I think that it is therefore important to also say, that I didn't close my eyes before criticisms of the technology or the industry. I am myself absolutely horrified by the way this industry behaves and how little regard it has for workers rights and the environment in the creation of models, intellectual property and all the rest. I read a book about it, too.

So, in short, my position can be described as "use it for defensible reasons, don't whitewash the industry and you're okay, we all are just trying to make it through the day". Something like that.

However something has changed recently. I wanted to polish up my cv and use AI to help me maybe finding a new job. I have been living in Finland since September 2019 but I never held a job here, always worked for a German company, remotely. I don't know if now's the time to really look for something new, but by chance I saw that there was a position for a senior dev by a company I find interesting and so I thought I'd work on that using at least in part AI to get me closer to what I maybe can hand in. I didn't want to do it using my employer's ChatGPT/Gemini subscription, didn't want to accidentally use the wrong account either and so I ended up subscribing to Claude Pro for a month.

To me, using Claude by paying for it myself, was an interesting experience. I noticed immediately that I was somewhat freed from the work context and could use it more openly for theory building/research as well as for working on my job search documents and other things where I would have maybe self censored before.

That later part, the theory building took the above mentioned ethical position and tried to work on it in a back and forth style with the chat bot acting as a reflective mirror. It was fun and lead to interesting progress, I though, sharpening some of my arguments. This often took the form of a discussing a news story in light of my ethical framework. Anthropic's RUM-deal was a recent example.

Often these chats started because I had a bad feeling about what I had just read: And I tried to put my developed ethical framework through its paces and see if it'd hold up. For example in this case:

I realized after a while, that I was not really developing an ethical framework and instead was developing a justification framework. All of these chats happened to follow that template pretty often. Whatever was problematic whatever had disturbed me, I found a way to soften, widen or relativize my criterion for ethical use.

And so forth. Funnily enough this all came to a head when I revisited a footnote in one of blog posts:

"AI is a cognitive hazard" argument](https://www.baldurbjarnason.com/2025/trusting-your-own-judgement-on-ai/). I am familar with this kind of argument that a certain kind of person trained in the humanities likes to make: the self sealing kind (in German we would call this a Immunisierungsstrategie or "strategy of immunisation"). I had said before that I don't think LLMs are a tabu - which I guess is to say they probably aren't master's tools (i.e. guns, whips, etc.).

I think that "humans are just like that" is a prime example of a self-sealing argument. If humans are just like that anything and everything can be justified by that position. And all the other extensions of my ethical position made this whole thing arbitrary. I could allow myself to do (or continuing to do) almost anything in spite of very disturbing news.

What lead me down this path was a particular general pattern of my thinking, as I recognize it now:

  1. a new technology (say) catches my eye
  2. I hear of critique about this
  3. the critiques don't really convince me, mostly because they are not really nuanced
  4. I continue engaging with the technology, but I also try to develop a nuanced position that justifies my continued engagement
  5. the world continues to spin, sometimes more awful things come out about the developers of that technology
  6. the new awful things are beyond the scope of my nuanced position
  7. I work on my position (i.e. changing scope) to stay consistent, because I like using the technology

I think for almost anything else except politics and personal ethics, widening your tent is a good thing. Generalizing and incorporating new data and at the same time working on your own model to stay consistent and productive is good. But with politics and ethics something else is going on.

[I'll give myself only one paragraph here for the following. I might write about this in more detail some other time:]

AIME is an important extension of actor-network theory (ANT). In it, Latour talks, amongst other things, about different kinds of associations between actors and gives conditions under which associations of a certain kind can be successful and under which they are not. Whereas ANT was just interesting to trace associations of any kind between actors (man uses laptop -> on which runs an editor -> which can use another program -> called Git to push content -> onto the internet by an untold number of mediators...), AIME tried to sort through the different kinds of associations actors could have with each other. One of the kinds of associations would be [MOR] which is short for morality. AIME's whole deal was to redescribe modernity using types of associations between actually existing actors instead of established terms like "society" that seem to just exist without being tied to actual things in the world. Anyway, very often modern institutions, while trying to untangle the world, have established certain kinds of rules on how we can talk about things. For example we tend to distinguish facts and values which in the end is done to allow science to operate in a moral-judgement free environment. Anyway, in AIME (and before) Latour shows that, of course, maters of fact are in actuality matters of concern, i.e. nothing is ever uttered value-free or without an agenda. For the project of AIME, matters of concern exist in a crossing of two associative modes that of [MOR] and that of [REF] or reference, which is part of what builds the institution of science: This is basically the paper trail part of science: How (say) an observation is being transformed into data. It is all the practical things scientists and others have to do to be able to say something credible about something. [REF·MOR] (a so-called crossing) is what was needed to make sense of my LLM-Ethics dilemma, because this was a question that involved both ethics and my wish for more grounded technology critique. However by using the criterion for [REF] on [MOR] I was making a category mistake.

Morality in this sense is marked by hesitation and search for errors and holes in an argument. A moral actor is looking for signs that things are not right. An immoral one doesn't concern themselves with trivial details that just get in the way. The mode of association is therefore one of assembling and reassembling things in a way that makes everything fit together. If something doesn't sit quite right, something's off, that's a sign that we are associating things in a moral mode.

Doing association by reference on the other hand is work that is marked by transforming. An observation is transformed into a note. That note is in turn transformed into a document in a database. The more transforming we can do, the longer these chains of reference can be the more stable they become. What make a reference association a reference association is a constant that binds heterogenous actors to each other. A precise location, a link, a shared name, a specimen, an id, and so forth. "This is that".

Whereas reference is only possible by translating and binding things together through shared properties, morality demands awareness of actual hesitation and a reordering if such hesitation is detected. I think that I took my moral criterion about use in a direction that made me think by binding a whole lot of theory and discourse to it, I would give it weight in the moral sense. What I did in actuality was that I turned it into a big ol' bunch of references that showed what could be referenced when talking about my position, in turn making it more grounded, but none of that really helped to get rid of the increasing hesitation I felt when I traced another chain of "awful news story + justification is somehow still possible" to my ethical stance. What was missing was the reordering of things. From a pure "link things together" standpoint, I could continue forever, refining what I meant by this or that in my ethical framework, introducing new concepts if needed so that there is space for all of it. But that didn't really resolve the tension I increasingly felt.

The discomfort I felt about using AI and paying for it personally had to be interrogated and interrogate I did. The result was that while at work I couldn't really choose in my current position to not use AI, I did chose to use and pay for AI out of my own pocket basically for personal gain and my own amusement. This made me responsible and I had to weigh these personal positives against all the negatives the AI-Industry has brought or is bringing to this world and if I wanted to be associated with that. Basically: Do I use it and complain loudly, or do I leave. And because I all of a sudden had to take responsibility, I could not justify using it anymore: It would've made me complicit and more importantly: I would have continued to feel that hesitation with every moment I used it longer. There is a row of theoretical/ethical defenses against this (f.x. shared intent), but what I had to realize was that, for me, I couldn't let them count. I knew that I would have acted immorally to just ignore my uneasiness.

So that's why I decided that I can't go on just using the thing. As much fun I had, as productive as I felt, as much as I think the tech is interesting and as much as I think that for an individual their use cases are what matters most: For me I can't go on ignoring my hesitation and the bad feeling. I can't replace doing the right thing by lengthening the paper trail.

P.S.: If morality demands reexamination I guess I am keeping a backdoor open for myself. A later reexamination might lead to a situation in which I deem using LLMs ok again for me. Since we can't help but be entangled with these things the moral work is never done, indeed it would be immoral to not interrogate the feeling that something had changed, again. But at least for now, this is what it is: No personal use of AI like Claude and the like.

P.P.S.: I don't know yet what I'm going to do with "my stuff" (code, notes, this blog, etc.). Some of the features of this blog have been developed using a coding agent for example. Is it fine to just keep those or reimplement those? How about writing - like this blog post. It doesn't include even a single AI-generated word, but is of course in part based on reflective chat sessions I had with an LLM. Or how about all the little things that make my notes system run nowadays? The templater scripts and the agent sessions that moved notes around and so forth. All of it seems tainted and I have the strong desire to throw it all in the trash (or actually: Don't Delete, Abandon) and start fresh. However, I notice hesitation. So we'll see what comes of that. For now we just stop using Claude and that's it.

  1. As a programmer and fellow programmers know about this, too, it is very very hard to find any job that doesn't expect you to use AI in the form of a coding tool at the very least. I can't really change that and as I Am A Dad now, I am in no situation to simply refuse using AI at work. I need my job and the money it provides. And I do not see any chance to start a discussion at work about this. So this is what I am going to say: As long as I don't change jobs and as long as that theoretical next job doesn't offer me the chance to work without AI, I will use AI for about 8 hours a day, at work until that industry implodes. ↩︎