Editors and AI, Part III: How Generative AI Really Works

In part 1 and part 2 of this series, I explored what the term “artificial intelligence” actually means and looked at which editorial tools use AI features. Today, let’s peek under the hood to understand how AI really works. This isn’t just theoretical—understanding AI’s mechanics helps editorial professionals (like you!) make informed decisions about which tools to trust and when to rely on human expertise instead.

How Generative AI Works (It Has Nothing to Do with Robots!)

As much as I like to joke about “robot overlords” taking over the world, those robot and cyborg images you see alongside AI news stories are misleading at best, fearmongering at worst.

The tech industry’s habit of anthropomorphizing AI tools by giving them human names, attributing emotions to them, and marketing them as “personalities” is partially to blame for the anxiety surrounding these technologies. I get it: It’s fun to interact with AI tools that have names and “personalities.” I even program some of my own custom AI tools to tell me jokes or give me music trivia.

But it’s crucial to understand that generative AI is more like Excel than anything from science fiction. It analyzes vast amounts of text to understand how language fits together, similar to how we might learn language patterns by reading thousands of books. It doesn't have actual thoughts, desires, or a personality; it’s just a sophisticated pattern-matching tool.

The Eager Intern Analogy

Think of generative AI like a brand-new, excited, eager intern. It has an encyclopedic memory but no practical experience, and it really wants to make you happy. This “intern”:

• Has memorized previous versions of style guides like the Chicago Manual of Style, but doesn’t have access to the current, live versions (so, it’s often using outdated guidelines)
• Knows the basics of writing, but doesn’t understand when to break the “rules”
• Can recite grammar rules but struggles with context and tone
• Wants so badly to be helpful that it will make things up (“hallucinate”) rather than admit that it doesn’t know something
• Has read extensively but has zero real-world experience
• Remembers everything it’s read but doesn’t truly understand it
• Can’t give the exact same answer twice—every time you ask a question, the answer will be different

This last point is particularly important for editors to understand. Unlike traditional tools that give consistent results (think Find and Replace or spell check), generative AI produces different output each time—even when you ask the exact same question. This inconsistency makes it challenging to rely on for tasks that require precision and reproducibility, which is a lot of what we do as editors.

How AI Learns

Generative AI systems like ChatGPT and Claude learn through a process called training, where they analyze vast amounts of existing text to identify patterns. Understanding this process matters because:

1. Generative AI can only suggest what it’s seen before—it can’t create truly original ideas
2. Its suggestions reflect patterns in its training data
3. It doesn’t actually understand language or context; it’s matching patterns

When ChatGPT writes you a poem or generates content, it’s using probabilistic reasoning based on the patterns it’s seen—like that eager intern drawing on everything they’ve read but without true understanding. This is why generative AI can write something that sounds perfectly plausible but is entirely made up, or create content that reads well but is inaccurate.

Why Generative AI Makes Mistakes (and Why This Matters for Editors)

Understanding how generative AI works helps explain why it fails in ways that directly impact our editorial work. As editors, we need to recognize these patterns of failure to protect our clients and maintain our professional standards.

Here are the most common mistakes AI makes—and why they matter for your editing practice.

Hallucinations

Generative AI might confidently cite a nonexistent source, completely make up a reference number for a CMOS rule, or make up statistics because it wants to please you and keep you engaged. It could add fictional details to a technical manual or invent steps in a process because it’s matching patterns of how instructions are typically written, without understanding that every detail must be accurate.

For editors, these hallucinations are particularly dangerous when we’re fact-checking or verifying style guide rules. Imagine confidently telling a client that CMOS recommends something based on an AI’s made-up rule. Yikes!

Context Confusion

AI might suggest changing contractions like “don’t” to “do not” in a casual blog post because it’s matching patterns of formal academic writing, not recognizing that your informal tone is intentional.

Or, on the opposite spectrum, AI might flag technical terms in a medical manuscript as “unnecessarily complex language” because it’s matching patterns from general writing guidelines about clarity and plain language. It doesn’t understand that in this context, precise technical terminology is exactly what’s needed—“myocardial infarction” shouldn’t be changed to “heart attack” in a journal article for cardiologists.

For a third example, generative AI will often revise dialogue in fiction to be grammatically perfect because it’s matching patterns of formal writing, not recognizing that characters often speak in fragments or use regional dialects.

As editors, we make these judgment calls constantly. We know when technical terminology serves the reader and when plain language would work better. We understand when informal contractions build rapport and when they undermine authority. These nuanced decisions require real-world experience that AI simply doesn’t have.

Overcorrection

Ask an AI tool to “lightly copyedit” a few paragraphs of text, and don’t be surprised if it does a way heavier edit than you wanted—it may even rewrite the entire passage. While it is possible to get generative AI tools to show restraint and only edit for a few items at a time, this takes a lot of training, patience, and coaxing. This is a big reason why generative AI doesn’t do well with detail-oriented levels of editing, like copyediting.

This matters for editors because we need to maintain the author’s voice while fixing actual errors. When AI swoops in with heavy-handed rewrites, it can strip away the very qualities that make a piece compelling or effective.

False Confidence

Because AI tools give suggestions in an authoritative tone, it’s easy to forget they’re just making probabilistic guesses based on patterns. They sound confident even when they’re wrong—just like that eager intern who doesn’t want to admit they’re unsure.

Understanding these limitations helps explain why human editors remain essential. We don’t just match patterns—we understand context, audience, intent, and nuance. We know when to break rules for effect and when consistency matters most. Most importantly, we understand that editing is about more than just following patterns—it’s about making meaningful choices that serve the text and its readers.

Our clients trust us to make informed decisions. When we rely on AI without understanding its limitations, we risk compromising our professional judgment and, ultimately, our clients’ trust.

This brings us to a critical question: How do we move forward? Understanding generative AI’s limitations isn’t just about knowing when not to use it—it’s about making informed decisions about if and how these tools might fit into our work. In my next post, I'll explore why many editors take a "just say no" stance on AI—and what we might gain from a more nuanced approach.

Recommended Reading

• Jones, Nicola. “AI hallucinations can’t be stopped — but these techniques can limit their damage.” Nature. January 21 2025. Accessed Jan 23 2025.

Previous Posts in This Series

• Editors and AI, Part I: What Is AI? A Primer for Editorial Professionals
• Editors and AI, Part II: AI in Editorial Software—Which Editing Tools Use AI and Which Don’t

This post was published on February 5, 2025.