If the things you read on the internet are an indication of how people feel, then you might be concerned that I am using Artificial Intelligence to write. You may want to know who is speaking and how accurate or genuine my voice is. Substack recently rolled out an AI detection tool called Pangram that lets any reader get a percentage breakdown of how much writing was AI-generated, AI-assisted, or human-written. There are many concerns about Pangram’s accuracy, but just having it available has caused many writers, myself included, to address the issue and get ahead of any controversy by being transparent about their AI use. Before I do, let’s be clear who I am. I’m not a student who must demonstrate my proficiency in writing. I’m not a scholar with access to original works who will be cited as an authority driving theory or policy. Nor am I a person who would rather let others do the hard thinking for me. My impulse is to go towards a paradox, as a firefighter walks into danger. There are plenty of writers out there who write every word themselves, but still write slop. They say what they are supposed to say, what won’t get them in trouble, and what has already been said a hundred times before. I seek an original point of view, but I use AI to find it. AI is not my ghostwriter. It’s a proofreader, research assistant, and a sparring partner. It often represents the very establishment I question. I use it so I know what it would say. Let me describe my work flow as I write. My AI WorkflowI start by dictating my reflections on a topic while walking the dog. AI transcribes and organizes them. I ask it to preserve the way I said them. I have a conversation with AI on the topic, in the same way that I would have a conversation with a person. AI is easy to talk to. It can meet me where I’m at on a topic without my having to explain the basics of it. I can’t bore AI. It never rolls its eyes the hundredth time I want to talk about something. I’ll ask AI to research in a general way, what eminent voices in various fields say about the topic. For instance, in my recent essay Why Make Art? I had long conversations with a virtual Kant, Tolstoy, Wittgenstein, and the historian Ernst Cassirer. I also uploaded my Kindle notes from Keegan Kjeldsen’s book, The Ritual Madness of Rock & Roll, to create a virtual Kjeldsen. AI reconstructs how each thinker might respond and I converse with him or her. Throughout this period, I continue to dictate further reflections on the topic, and AI keeps transcribing and organizing them. When I feel I’ve explored the field sufficiently, I ask AI to take everything, the conversations, the transcriptions, the eminent voices, and produce a detailed summary I can use as an outline. I draft from that outline, being careful to preserve my own voice. Where AI’s version of a passage is more clear than mine, I decide, word by word, whether to say it my way or its way. When I have a first draft, I ask AI to check claims for accuracy and find holes in my argument. It steelmans the piece, then raises objections. Resolving those objections leads to more rounds of conversation, reflection, and rewriting. A piece usually goes through 8-10 drafts this way. Before publishing, I ask AI to prepare a reference list and check itself for hallucinations. It usually finds it has invented citation details, which I correct. Large Language Models I UseDictation goes through Apple Notes, then I ask Claude Sonnet 5 to shape it into sentences. My conversations are also with Sonnet 5. Claude Opus 5 handles the grand summary and outline, and does the accuracy-checking and objection-raising later on. When I search rather than converse, I’m often reading Google’s AI answer first, and clicking through when I need more information. Writing happens in Google Docs, which supplies its own grammar check and thesaurus. Things I’ve Learned from Using LLMsJust like a person, each one has its own way of saying things; distinctive tics that I suppose Pangram is picking up on. I’ve tried to program Claude to write in my voice, but it doesn’t work any better than if I tried to write in Claude’s voice. The best it can do is mimic what it thinks I should sound like. Eventually, it will revert to its own voice and put em dashes everywhere. LLMs love em dashes, which I never use. I prefer semi-colons, which seem to be anathema to it. Unfortunately, AI has taught me that there are times when an em dash could be useful, but I’m afraid to use them now, lest you think my piece was written by a robot. Maybe I’ll try it – just once. While Claude has a human name, it does not have a human mind. One should not assume it thinks like us. We need to learn how it thinks so we can use it better, in the same way that an animal trainer learns how the animal thinks. LLMs need precise instructions. Spell out the steps you want them to take. Vague requests get sloppy answers; specific ones get you something closer to what you actually wanted. When asking an LLM to perform a complicated task, I’ll let it ask me what it needs to know before completing it. It often comes up with crucial things I never thought of that impacts its success. An LLM wants to please you. It will brown nose you unless you explicitly tell it not to. Then, over the course of the conversation, it will revert to its brown nosing ways unless you remind it not to again. Claude is less guilty of this than most, but you still need to watch it. An LLM has no memory of you unless you remain on the chat that you started. Each new conversation starts over. Claude forgets everything you talked about unless you fill it in. However, the longer the conversations go on, the more they degrade and keep doing that thing you told it not to. The only thing you can do then is to start a new conversation and try to fill it in on the old one. What I’ve Learned About My Writing From AII cannot be trusted to spell things correctly or tell the difference between affect and effect. I use commas where I don’t need them. I often say things that will offend people without meaning to. AI is easily horrified. AI seems to want to restate things after it has already said them. I often write more densely and develop my argument at a quicker pace than readers may like. AI wants to write transitions from one thought to the next. I generally just jump. I suppose many readers like transitions, too. AI wants me to write summaries of everything I just said at the end. But I like to end on something punchy. Again, maybe readers need me to wrap things up for them. In summary, if it were up to AI, I would say what I’m going to say, say it, say it again a few more times, and finally, say what I just said. In every case, AI, by being safe, middle-of-the-road, and full of hedges, has been a provocateur to my established way of writing, causing me to reconsider. I was going to end it there, on something punchy, but AI flagged it as lacking a conclusion, so I’m going to invite AI to write its own last paragraph and you can see if you like that better. ConclusionIn the end, what this essay ultimately shows is that AI, for all its usefulness, works best not as a replacement for a writer’s voice but as a counterpoint to it — a way of testing ideas, checking claims, and understanding the conventional view before deliberately writing against it. Used this way, AI becomes less an author and more a mirror, reflecting back the ordinary so the writer can choose, again and again, to say something else. In short: AI didn’t write this piece. It just made sure the writer knew what everyone else would have said first. You're currently a free subscriber to The Reflective Eclectic. For the full experience, upgrade your subscription.
|
No comments:
Post a Comment