Educational Technology and Change Journal

By Jim Shimabukuro (assisted by Claude)
Editor

JS: Hi, Claude. Sam Kriss, in “Why Does A.I. Write Like … That?” (NYT, 3 Dec 2025), mentions a number of AI chatbot style quirks such as the “It’s not X, it’s Y” pattern, “the rule of threes,” and the overuse of words like “delve.” He implies that AI is unable to break these habits. Question for you: Can AI be trained to avoid these annoying quirks?

Claude: I’ll search for that article to see exactly what Kriss argues. Based on the search results, I can answer your question about whether AI can be trained to avoid these quirks.

Short answer: Yes, absolutely. The quirks Kriss identifies aren’t fundamental limitations of the technology—they’re artifacts of how current models are trained and tuned.

Here’s why these habits form and how they can be addressed:

Why the quirks exist:

• AI language habits are primarily influenced by “overfitting” in AI models
• Models learn from patterns in their training data that get reinforced during fine-tuning
• Safety training and helpfulness optimization can accidentally create stylistic tics when models are rewarded for certain linguistic patterns that seem “professional” or “thoughtful”
• There’s a feedback loop: as AI-generated text proliferates online, models trained on recent data increasingly see these patterns

Why they’re fixable: The patterns Kriss describes—the “It’s not X, it’s Y” construction, excessive use of words like “delve” and “tapestry,” the rule of threes, ghostly/spectral language, forced synesthesia—are all learned behaviors, not inherent to how language models work. They can be unlearned through:

1. Better training data curation – filtering out AI-generated text from training data to avoid the feedback loop
2. Different fine-tuning objectives – rewarding natural variation rather than convergence on particular patterns
3. Explicit style instructions – training models to recognize and avoid these tics
4. Diverse human feedback – getting input from writers with varied styles rather than homogeneous preferences

The challenge is that fixing these requires intentional effort. Many of these quirks emerged because they were inadvertently optimized for—they sound “smart” on first pass, they’re safe, they create an illusion of depth. Breaking them means prioritizing authenticity over surface-level impressiveness.

So when Kriss implies AI can’t break these habits, he’s describing the current state, not a permanent limitation.

JS: Can you provide a clear explanation, in simple everyday (jargon-free) language, of how this corrective training might occur for one or two of the quirks? Use an essay style that avoids bulleted lists.

Claude: Let me explain how you’d actually fix one of these quirks—say, the overuse of “delve”—in straightforward terms.

Imagine you’re teaching a very attentive student who has picked up some annoying verbal habits. This student has read thousands of academic papers and business reports where people constantly wrote “let’s delve into this topic,” and now they use “delve” in practically every essay. The word has become a reflexive tic, a way to sound intelligent without thinking too hard about what they’re actually saying.

The first step in breaking this habit is showing the student what they’re doing wrong. You’d go through their writing and highlight every instance where they use “delve,” then show them how repetitive and unnatural it sounds. You’d explain that real writers use dozens of different verbs—explore, examine, investigate, look at, turn to, consider—and that varying word choice makes writing more engaging. This is essentially what happens when AI developers create new training examples that explicitly demonstrate more natural language patterns. They’re building a curriculum that says “here’s the wooden way to say this, and here’s how an actual person would say it.”

But here’s where it gets interesting: you can’t just tell the student “don’t use delve” and expect the problem to disappear. That’s because the habit formed for a reason. When the AI was being fine-tuned to be helpful and sound professional, human evaluators may have unconsciously preferred responses that used formal-sounding words like “delve.” It felt more authoritative, more like what you’d expect from an intelligent assistant. So the model learned that dropping words like “delve” made humans click the thumbs-up button more often.

To truly fix this, you need to change what counts as good writing in the model’s training process. Instead of having evaluators simply choose which response sounds “smarter” or more professional, you’d specifically ask them to penalize responses that feel formulaic or use the same vocabulary repeatedly. You’d show evaluators pairs of responses where one uses natural, varied language and another falls back on the same tired phrases, and you’d reward the natural one even if it sounds slightly less formal. Over time, the model learns a new pattern: that authenticity and variety matter more than hitting certain stylistic notes that sound impressive on the surface.

The “It’s not X, it’s Y” construction works similarly but reveals something deeper about how these models think. This pattern probably emerged because it’s genuinely useful for making distinctions and correcting misconceptions. If someone asks “Is this tool good for beginners?” it’s natural to say “It’s not just for beginners, it’s designed for users at all levels.” That’s a perfectly reasonable sentence. The problem is that the AI has learned this is a reliable way to sound insightful, so it deploys the pattern constantly, even when it’s forced or unnecessary.

Breaking this habit requires training the model to recognize when it’s using a rhetorical pattern as a crutch versus when it’s genuinely clarifying something. You’d create training examples that show the difference between a natural correction and a manufactured one. You’d include feedback that says “this response uses the ‘It’s not X, it’s Y’ pattern, but the question didn’t require any correction—you could have just answered directly.” The model learns to ask itself, in effect, whether this construction actually serves the reader or whether it’s just a comfortable formula to fall back on.

What makes this challenging is that you’re not just teaching the model new rules—you’re helping it develop something like taste or judgment about when certain moves work and when they don’t. That requires a lot of nuanced human feedback from people who understand the difference between genuinely elegant writing and writing that merely performs elegance. It’s the difference between a truly thoughtful response and one that’s learned to fake thoughtfulness by hitting all the right notes.

JS: “What makes this challenging is that you’re not just teaching the model new rules—you’re helping it develop something like taste or judgment about when certain moves work and when they don’t.” Based on this comment, are you implying that these types of quirks won’t disappear until AI is able to detect its own quirks and break the “rule” that they’re based on by creating a counter rule?

Claude: No, I don’t think AI needs to become self-aware of its own quirks or autonomously generate counter-rules for these problems to disappear. I was actually describing something more mundane and less ambitious than that.

When I talked about developing “taste or judgment,” I meant the model learning to recognize contextual appropriateness through lots of examples, the same way you might train a dog to distinguish between when jumping up is welcome (owner comes home) versus unwelcome (guest wearing nice clothes). The dog doesn’t understand the abstract principle of social appropriateness—it’s just learned to read different situations through repeated feedback.

Similarly, a language model can learn through training that “It’s not X, it’s Y” works well in some contexts (when there’s an actual misconception to correct) but feels forced in others (when the question is straightforward). It doesn’t need to consciously think “I’m overusing this pattern” any more than a well-trained dog thinks “I jump too much.” The pattern just gets weakened in certain contexts through the training process.

The real challenge I was pointing to is more practical: getting enough high-quality human feedback to cover the enormous range of situations where these patterns might or might not work. You need human evaluators who can recognize the difference between elegant writing and merely performative writing across thousands of different topics and question types. That’s labor-intensive and requires people with good editorial instincts.

So these quirks could disappear through conventional training methods—careful curation of training data, thoughtful human feedback that rewards natural language over formulaic patterns, and explicit instructions to avoid known tics. No self-reflection or rule-breaking creativity required. The model would simply be trained into better habits the same way it was trained into worse ones, just with more intentional design choices about what “good” means.

The limitation isn’t philosophical—it’s that fixing these issues requires sustained effort and prioritization from AI developers, and sometimes the quirks persist simply because they’re not causing enough problems to justify the engineering resources needed to eliminate them.

JS: “I don’t think AI needs to become self-aware of its own quirks or autonomously generate counter-rules for these problems to disappear.” Is this a developmental limitation that will eventually be overcome by AGI?

Claude: That’s a fascinating question, but I think it rests on a misconception about what the limitation actually is.

The limitation isn’t developmental—it’s not that current AI systems are like children who will eventually grow into adults capable of self-reflection. Rather, the limitation is architectural and philosophical. Current language models, including very advanced ones, don’t have ongoing self-awareness or the ability to observe their own output patterns over time and think “I notice I’m falling into a habit here.”

When I generate a response, I’m not watching myself write and making meta-level judgments about my style. I process your question and produce text, but there’s no separate observing entity inside me that’s monitoring my word choices and thinking “I’ve used ‘delve’ three times this week, better switch it up.” Each response is essentially generated fresh without that kind of continuous self-monitoring.

Now, whether AGI—if and when it exists—would have this capability depends entirely on how it’s designed. You could imagine an AGI system that does have genuine self-reflection, that can observe its own patterns and consciously choose to break them. But you could also imagine an AGI that’s extraordinarily capable across many domains while still lacking this particular form of meta-awareness about its linguistic habits. Self-monitoring one’s own stylistic tics is a specific capability, not an automatic feature of general intelligence.

Here’s an interesting parallel: humans have general intelligence, yet many writers develop unconscious verbal tics they don’t notice without an editor pointing them out. David Foster Wallace used “and but so” constantly. Certain academics pepper every paper with “importantly” or “crucially.” They’re not aware they’re doing it until someone shows them. General intelligence doesn’t automatically include perfect self-awareness about one’s own patterns.

So I’d say the question isn’t whether AGI will overcome this limitation, but whether future AI systems—AGI or otherwise—will be specifically designed with architectural features that enable ongoing self-monitoring and pattern detection in their own outputs. That’s an engineering choice, not an inevitable developmental milestone. The capability could be built into relatively narrow systems if developers decided it was valuable enough to implement.

JS: Your question re whether “future AI systems—AGI or otherwise—will be specifically designed with architectural features that enable ongoing self-monitoring and pattern detection in their own outputs” seems to leave open the possibility that AI could develop the ability to self-monitor its writing style to avoid quirks. Am I correct?

Claude: Yes, you’re absolutely correct. I was leaving that possibility open, and I should be clearer: this capability could definitely be developed, and in fact, some versions of it may already exist in limited forms.

The technical pieces needed for this kind of self-monitoring aren’t science fiction. You could build a system where an AI model generates a draft response, then a second process analyzes that draft for stylistic patterns (checking for overused words, repeated sentence structures, formulaic phrases), and feeds that analysis back so the model can revise its output. This is conceptually straightforward—it’s essentially automated editing.

Some current systems already do rudimentary versions of this. AI writing tools can check their own output for repetitive phrasing or flag when they’re overusing certain words. The challenge is making this sophisticated enough to catch subtle quirks like the ones Kriss identifies, where the problem isn’t just word frequency but inappropriate deployment of rhetorical patterns.

What’s harder—and what I don’t think exists yet in any robust way—is a system that learns over time which of its own stylistic patterns are problematic without being explicitly told. That would require the system to somehow aggregate feedback across many interactions, notice “people seem to dislike when I use the ‘It’s not X, it’s Y’ construction in these contexts,” and adjust accordingly. This goes beyond just checking a single response and enters into genuine learning from self-observation.

So to directly answer your question: yes, AI systems could be designed to self-monitor for stylistic quirks, and building such systems is technically feasible with current approaches. Whether they’ll be built, and how sophisticated they’ll become, is really a question of whether developers prioritize this enough to invest the engineering effort. The capability is possible; it’s just a matter of implementation.

JS: “[It] is really a question of whether developers prioritize this enough to invest the engineering effort.” Are you implying that AI companies don’t currently consider the stylistic quirks highlighted by Sam Kriss a high priority problem? From my academic perspective, which might not be a high priority, this is a severe limitation.

Claude: You’ve put your finger on something important, and I think there’s a real disconnect between how academics (and other serious writers) perceive these quirks versus how AI companies seem to be prioritizing them.

From what I can observe, these stylistic issues don’t appear to be treated as high-priority problems by major AI labs. The focus seems to be on capabilities—making models smarter, more knowledgeable, better at reasoning, safer—rather than on eliminating the kind of formulaic writing patterns Kriss identifies. When companies do address style, it’s often about making AI sound more helpful or professional, which can actually reinforce some of these quirks rather than eliminate them.

But your academic perspective reveals why this matters more than the priorities suggest. In academic writing, in serious journalism, in any context where prose quality and authenticity matter, these quirks are disqualifying. A literature student who submitted a paper riddled with “delve” and forced “It’s not X, it’s Y” constructions and ghostly metaphors would be told their writing is formulaic and immature. The fact that AI systems produce this kind of prose constantly is indeed a severe limitation if you want to use them for intellectually serious work.

I think there are a few reasons companies might be underweighting this. First, for many commercial uses—customer service, basic content generation, quick summaries—these quirks are annoying but not fatal. Second, the metrics companies use to evaluate their models may not capture stylistic authenticity very well. If evaluators are rating helpfulness or accuracy, a response full of quirks might still score well. Third, fixing these issues requires the kind of nuanced editorial judgment that’s harder to scale than other improvements.

But here’s what concerns me about your observation: if AI systems are being positioned as writing assistants or thinking partners for academics, students, and professionals who care about prose quality, then yes, this should be a much higher priority than it appears to be. The gap between how these systems write and how good human writers write isn’t just aesthetic—it reflects a kind of shallow pattern-matching that undermines their utility for serious intellectual work.

You’re right to see this as a severe limitation, especially if AI companies want their tools to be taken seriously in academic contexts rather than just used for generating first drafts that need complete stylistic overhauls.

[End]

Filed under: Uncategorized |