Part 5: Use AI for verification (Bakhtin and model collapse: How to use AI with expressive writing without generating AI slop)

by Tom Johnson on Jan 30, 2026 comments
categories: ai • ai-book-club • writing

This post is part of a series. See Part 4: Use AI as a research assistant for the previous section.

• Part 5: Use AI for verification
• Next section

Part 5: Use AI for verification

Now let’s move on to another use of AI: verification. I’m currently reading Meghan O’Gieblyn’s God, Human, Animal, Machine. It’s impressive how well-read O’Gieblyn is, not just the scope of authors she’s conversant with, but how well she simplifies the messages of philosophers and theorists such as Descartes, Kierkegaard, Bostrom, Bohr, and countless others. (I have no idea how she keeps track of all these quotes and summons them at will for the topic at hand — it’s really impressive.)

For a novice / amateur hack like me, I’m not confident in my ability to summarize major philosophers. I sometimes struggle to follow the details of the philosophical arguments, and when I do quote others, I want to be sure I’m accurately representing their ideas. It’s the same impulse I have in technical documentation: I check my content against the reference material, and I do so multiple neurotic times. Bakhtin would argue that my need to check and verify meaning is because a word is “half someone else’s” (293). More importantly, if I misrepresent a source, I’m not engaging in true dialogue, I’m just performing a monologue with a puppet. I want to be sure I’m actually engaging with a fair, legit representation of an author’s ideas.

I was recently writing a book review of If Anyone Builds It, Everyone Dies by Eliezer Yudkowsky and Nate Soares. I wanted to help remember and understand the details of their arguments better, so I found an online PDF of their content, added it into NotebookLM, then presented my draft blog post there to check that I was correctly interpreting their arguments.

This process helped in two ways. First, I felt more confident that I was accurately representing the author’s ideas. When critiquing someone, it’s easy to provide a straw man argument. AI acts as a neutral third party here, providing the author’s arguments in a thorough, fair way.

It’s common to be accused of taking ideas out of context or “cherry-picking” an argument. AI can help ensure you’re portraying the author’s actual work. This intellectual honesty is what turns a “histological specimen” of a quote into a living, “active understanding” (281), Bakhtin would say. Usually AI tends to present someone’s ideas in an almost generous light, giving them more credibility than I think they deserve, but in a constructive, helpful way.

Note that I admit that there’s benefit to wrestling with a text on your own. It can be too easy to abdicate the role of interpreter to AI and let these tools perform their magic simplification. There’s no doubt a loss of critical reading that takes place in doing so. Warner says that reading (like writing) is an emotional experience. When AI creates bullet points of another’s work, it removes our struggle and engagement with that text, which is a loss. Limiting ourselves to the summary removes our emotional engagement entirely, making the text much less interesting and meaningful to the reader. Without more of a pull towards the other, the writer’s dialogic imagination falls flat, because the writer is talking with a text he or she cares little about.

You can’t reduce a book to a list of bullet points and still get the same reading experience. As one of my colleagues pointed out, just as we experience emergence in writing (“emergence” referring to the spontaneous surfacing of ideas that have no clear origin to what produced them), emergence also happens during reading. O’Gieblyn says of her writing process, “at some point the thing I have made opens its mouth and starts issuing decrees of its own. The words seem to take on their own life, such that when I am finished, it is difficult to explain how the work became what it did.” This emergence happens at the boundary between my own mind and the text, as Bakhtin would say. If I offload the interpretation to AI, I short-circuit the likelihood of emergence during reading.

O’Gieblyn is wary of emergence, as she likes to maintain control over her own creative process, and pushes back against the black box algorithms of both AI and religion that favor the mystical. Or rather, she pushes back against obedience to systems that we can’t understand. But as even a blogger, I love the idea that when I start writing a post, I never quite know what I’ll discover or how it will all turn out.

Overall, having AI help me understand my sources makes me more inclined to engage with topics I might otherwise avoid. If AI can be a tool that helps us in that wrestling match, furthering our understanding, and thereby encourages more engagement with the text, is it not a valid tradeoff?

One more point about verification. When writing, it’s easy to fall into the mode of finding sources that support the points we already have. I find that AI pushes back against these easy interpretations, making the sources we cite more problematic. But these problems, the rough points and misalignment, are exactly what makes us engage in a more interesting, lively way. The sources that somewhat support but also kick back make these texts more of a wrestling match that we have to engage with, think through, rather than just use as puppets of pretend support. And that engagement/struggle inserts the soul into our writing.

Next section

Continue on to the next section, Part 6: Use AI for the alien context.

Comment on LinkedIn