Essays by Glen Cantrell on AI, writing, judgment, and human responsibility.
Diderot spent the better part of twenty-five years building an encyclopedia meant to gather what was known and put it where more people could reach it, and in the process he kept running into the questions that come with that ambition: who gets access to knowledge, how it gets sorted, what the categories conceal, and who gets to claim authority over the answers. I’ve been thinking about those questions a great deal lately, because the machines we’ve built now do much of that work at a speed and scale he never imagined, gathering, connecting, reorganizing, interrogating and redistributing what we know, until, in crude terms, the encyclopedia answers back.
These essays follow what that has meant in one ordinary life, a writer’s and a small publisher’s and a small-town Chamber president’s, where AI made curiosity and correction cheap enough to try, and where the harder questions about context, evidence, accusation and identity turned out to be ones no machine could answer for me. I use these tools openly and argue with them constantly, and neither habit has proved sufficient on its own. Cheap cognition makes gathering and interpreting easier, but the responsibility for what we do with the answers is proving much harder to automate, and that is where these essays keep ending up.
I use AI openly to examine my writing and handle publishing work I could not otherwise afford. That assistance makes a small press possible, but it also raises questions about authorship, disclosure, and what copyright protects. “Did you use AI?” tells us too little; the useful argument begins with where we draw the line.
AI made a book for a tiny local audience affordable enough to publish and circulate. Readers elsewhere recognized parts of the experience while exposing limits and mistakes. A working experiment in automating the AI book club then had to be dismantled. Cheaper experimentation helped only when I accepted what checking revealed, including answers I didn’t want.
I gave four AI models six scenes built around “bless your heart,” holding the words nearly still while changing relationships and circumstances. Their disagreements exposed different assumptions about affection, criticism, and belonging. Then the transcripts revealed that two models had brought prior context into the test: I was partly measuring what the machines already knew about me.
Jerry Falade’s disputed account of an authorship accusation sent me to detector documentation and tests with my own writing. The results showed why sample length, surrounding text, and the presentation of certainty matter. A detector can supply useful information; people still have to decide what evidence justifies an accusation, and how much harm they are prepared to authorize.
A book-marketing pitch knew my reviews, career, elf costume, and forgotten Substack. Following its headers, tracking pixels, and duplicated websites exposed the machinery surrounding that apparent intimacy, without establishing exactly who or what wrote it. Making my public life legible to search and AI had also made it easier to assemble a convincing sales prospect.
A reader liked my book’s story and setting but stopped halfway through because the prose sounded “prefabricated.” The suspected AI tells were rhetorical habits I learned in school. I can accept the criticism and remain responsible for the writing, including AI-assisted editorial choices, without accepting that familiar sentence patterns establish who wrote it.