Reasons to Avoid Using Artificial Intelligence
The many dangers of reading what the LLM has to say
This is the second post in a sequence. You can find the first part here and the third part here.
I don’t use AI tools much. As I said in my previous post, this is not because I don’t think they are useful. Rather, I’m concerned about the ways that my thinking could be changed for the worse by reading too much LLM output. I believe that more people should be concerned about these, too, because they are often overlooked. It’s not just the risk of getting a hallucinated response and not bothering to double-check what the computer says; reading and relying on AI output can have more pervasive negative effects.
First and most obvious is the risk of cognitive atrophy. When you outsource your thinking to the machine, you aren’t doing that thinking yourself. Yes, you might tell yourself that you will carefully check everything afterwards, but how true is that? It’s much easier to just rely on what the machine says. This is not so bad when it’s something you don’t know how to do, and realistically would never be able to do (though in that case you are not in a position to actually verify the work). But if you habitually turn over a cognitive task, you will begin to lose the ability to do it yourself. This includes the generic cognitive task of “start with a question, and work through various possibilities to come to an understanding of the answer”. This is an ability that you really want to keep, and it’s a tempting one to lazily outsource to the machine.
Another risk, primarily from using LLM chatbots, is that of developing bad emotional responses due to the fact that LLM output sounds a lot like human conversation. One danger is the well-known one of coming to believe the chatbot is a person. (Witness, for example, all the people who think they have an AI boyfriend/girlfriend, or who got attached to the “personality” of GPT 4o.) But there is an opposite danger as well: instead of treating the machine-that-sounds-human as a human, you might learn that things which sound human can be treated as machines — including actual humans.
If you are, or were, an avid reader, you may have noticed that your own thinking and writing is influenced by who you choose to read. We are great imitators, for good or for ill, and while we mostly imitate those whom we admire, we also imitate what we merely see a lot of. If you read a lot of LLM output, you will learn to write and think like the LLM. This has obvious interpersonal consequences (because others will perceive what you write as “sounding like AI”), but it’s also a bad thing in itself. You want to retain your own thinking and voice, influenced by those whom you find most worthy of imitating. If you allow your thinking to be shaped by LLM writing, you will lose whatever value your thoughts have beyond what anyone could get by asking the AI.
Finally, and perhaps most importantly, reading AI output opens you up to dangerous new ways of being deceived. This isn’t really a matter of the AI lying to you. Rather, it’s because the way we read meaning into language is predicated on the fact that it is the product of a human mind. If you read a story, a poem, an essay, or any other piece of writing, you reason, both explicitly and automatically, about the author’s intended meaning. But this reasoning comes along with inseparable assumptions about what kind of mind the author has. The author’s biases will fall along lines you mostly understand, because they are human biases; the author’s motivations are discoverable, because they are human motivations; the author’s values are comprehensible, because they are human values; the author’s meaning can be understood, because the author is a human interested in saying human things.
None of these things is true about AI writing. It is debatable whether or not the AI has anything that can be properly called a mind; whether it really has values and motivations and meaning. But if it does, what they are certainly not is human. Yet when you read any writing, you automatically process it as if it came, ultimately, from a human mind; making assumptions you don’t even know you are making, and drawing conclusions you never think to question. This means that reading AI writing necessarily involves making false assumptions, and probably drawing wrong conclusions. This may be a very small concern when the only thing present is a verifiable, technical answer to a question, or something similarly objective. But it becomes a bigger problem the more the writing contains open-ended discussion. Now your processing of the piece involves, not merely logical evaluation of explicit claims, but also the tone, voice, subtext, and other meaning-bearing subtleties of the medium. And you will assume without thinking that those things came from a human mind, and open yourself up to deception thereby.
The danger is at the greatest when the writing is something meaning-laden like a poem or story. Ultimately the only proper response to AI-generated poems, stories, and the like is to ignore them, or at most to gawk at the technical capabilities on display. Reading them for meaning is a fool’s errand, and ultimately part of the essence of such things is to mean something; to be a transmission of something from one human mind to another.
In what ways, then, is it safe to use AI? Nothing is ever totally safe, but I think you’re mostly clear to use it as a glorified search engine, to discover outside resources or find a quote or reference. It’s probably safe to use it for programming tasks and the like, as long as you are clear-eyed about what skills you may let atrophy by doing so. (Safe for you mentally, that is. Whether it’s safe for your computer or employer is another matter.) And, of course, few individual uses of AI carry a ton of risk by themselves. The dose makes the poison. But you probably need to be more cautious than you think.