Welcome to alex wennerberg's website

I do my own writing

AI text is clogging our minds, and clogging the internet's series of tubes. With tools like Pangram, it is easier than ever to confirm my suspicion that a lot of the text that I encounter online is, unfortunately, AI generated.

I want to distinguish between two attitudes towards AI writing. The first is an absolute refusal to use these tools in any part of the writing process. The second, more moderate perspective, is that, with restraint, there are some aspects of the writing process that can benefit from AI tools. As I experiment with these tools more and more, I increasingly find myself in the former camp.

I am not absolutely against AI tools broadly speaking. For example, as I've written before, these tools are undeniably useful for software development. There are some who still refuse to use them on principle, but that principle, in my view, simply cannot be productivity and quality, both of which can be dramatically improved through careful use of AI assistants. I remain ambivalent and critical of these tools, and I am still exploring their strengths and weaknesses, and I'm certainly on the lower end when it comes to AI coding optimism, but they are still a major part of how I write code today.

With coding, in my view, the issues with AI tooling are real, but they can be mitigated through their judicious application. I find myself far more skeptical of their ability to be used responsibly in the domain of writing.

I have a humanist attitude towards writing. The kind of writing that is not meant to produce a measurable effect in its reader and is not oriented towards a specific and definable goal, in other words, most essay, nonfiction, fiction, and journalistic writing, broadly speaking, requires, by its very nature, a human subject. It is not something that AI can be good at because expression is not something AI is capable of. AI can, in a limited sense, "create" things, but those things have no original source, they are merely text directed towards some end (in the case of chatbots, satisfying their human prompter).

The danger that I find is that AIs are exceptionally good at fooling you into thinking that they are creative, that they do have an experience or a point of view. They will use language and produce output that gives this impression. The danger is when this impression is a facsimile of knowledge and insight. I recently encountered the phenomenon of "AI lectures" on YouTube. The shocking thing about these lectures is that in many cases they feel lucid and insightful. AI can break down Felix Guatarri's thought into something that is accessible and gives the impression of profundity, while ultimately, upon close inspection, producing gobbledygook and erasing the subtlety and context of his thought, which is only accessibly through difficulty, through struggle. Unlike AI Guattari, actual Guattari is challenging, even frustrating to read. AI Guattari produces the feeling of insight that one gets from encountering his work, without the actual struggle required to get the insight for real.

This is not an issue with AI's "capabilities." It is an issue with the nature of the endeavor itself, with applying these tools toward an end that they are simply not the appropriate tool for.

AI has the very real capability of fooling you into thinking that its ideas are your own. "Writing alongside" an agent will provide you with a ton of feedback, some of which may be genuinely useful, but none of which is yours. AIs are not a tool in a conventional sense, like a pen, keyboard or text editor that you wield, rather, they are like a robot version of a human editor of collaborator, without the actual specificity and discernment that a real editor-writer relationship has. This is something that only a human being possesses and can do.

I remember feeding Claude, as an experiment, a draft of my post Do-nothing software development. It gave, in response, a number of suggestions that were basically accurate: the first half doesn't really connect to the second half, the ending is kind of rushed, and so on. These were not my ideas though, nor the ideas of a human editor. Had I followed its advice, I would have ultimately become a "reverse centaur"—an external appendage through which AI is able to do its writing. Even if I am the one physically typing the words, as soon as I bounce an essay off an AI agent, it becomes the one writing the text, and it's very easy for this to occur without you noticing.

A similar problem arises when using AI as a "research assistant." Some people use AI to "brainstorm" in the early phases of their work. AI has no ideas though, and will very confidently give you a literature overview that lacks sufficient context. I remember one time it told me I should read a book by "the leading expert" on some subject, only to find that the author was non-academic, self-published and a total crank. This isn't just a limitation of capabilities, this is a kind of mistake that no human research assistant would make.

Because of these dangers, I am extremely wary of using AI in all but the most narrow and limited aspects of my writing process.

There are two manners in which I do use AI. I use AI as a kind of "Super Google" to help me find documents or quotes on some subject. I want to be very specific about what this means. In my previous blog post, Worse Is Better in Magic: the Gathering, I was aware of the Mark Rosewater blog post on One with Nothing, but didn't have a link, and asked Claude to find it for me. I was aware that the decks that I referenced had tournament performances, but didn't know when and where, and asked it to tell me this. I knew vaguely the Nietzsche quote that I referenced, but wasn't sure where it was in Genealogy of Morals. In all these cases, Claude basically was a highly capable search engine, and I was careful not to give it the abilities to give me ideas or to organize my research. I don't necessarily have a problem with using it in this highly narrow, directed way.

The second manner, and this one I feel much more ambivalent about, is in copy editing. At some point, I asked Claude to run through all my blog posts and apply Chicago Manual of Style formatting to them. I regret doing this, because it means, to a very small degree, these documents have AI-produced text (i.e. punctuation, capitalization, etc.). Chicago Manual of Style formatting is mostly objective, but not entirely so, and so to a small degree, I relegated copy editing to an untrustworthy external tool.

I am not sure what the best replacement for this would be. For my previous post, I asked Claude to find copy editing mistakes and then made the changes myself. For this post, I did nothing. I experimented with Vale, but it doesn't seem very good at catching the kind of issues that Claude catches. I don't object to automated tooling entirely, and I may be overly obsessing over the problem of AI copy editing. If anyone has any alternative for catching small formatting mistakes that do not involve AI tooling, I would be highly interested. Please email me at alex@alexwennerberg.com.