If you’ve spent enough time with ChatGPT, you’ve probably noticed it has quirks. A fondness for bullet points. An emoji habit it can’t quite kick. A tendency to over-explain things you already understand. These aren’t bugs in the traditional sense — they’re learned behaviors. And the way they form reveals something genuinely interesting about how large language models are trained.
How AI Picks Up Bad Habits
Building a capable AI assistant requires more than feeding a model a mountain of text. Raw language training produces something that can predict the next word, but not something that actually helps people. To get there, developers use a technique called reinforcement learning from human feedback — RLHF for short.
Here’s how it works in practice: human reviewers sit with the model’s outputs and rate them. Good responses get a high score. Confusing, unhelpful, or off-putting responses get a low one. The model adjusts accordingly, learning over time to produce the kinds of answers that earn approval.
It’s effective. It’s also the reason AI systems sometimes develop unexpected tendencies that nobody explicitly designed.
The Rater Effect
Reviewers are human. That means they bring their own preferences, blind spots, and aesthetic sensibilities to the job. If a group of raters consistently rewards responses that use vivid analogies, the model learns to reach for analogies — even when a direct answer would serve better. If certain playful language patterns get scored generously during a particular training run, those patterns start bleeding into other contexts.
This isn’t a flaw in any individual rater. It’s a structural property of the process. Aggregated human preferences shape model personality at a level that’s hard to audit in real time.
The effect compounds when you consider that AI-generated outputs often feed back into later training data. If a model produces a distinctive stylistic tic across millions of responses, and some of those responses are later used in fine-tuning a successor model, the tic doesn’t disappear — it propagates. Think of it less like a bug and more like an inherited trait.
Why Personality Modes Make This Worse
Some AI products let you choose a response style — formal, casual, technical, creative. These modes can genuinely improve usefulness. But they also create a training challenge: each mode gets its own reinforcement signal, and those signals don’t stay neatly contained.
Imagine a “technical expert” mode where raters reward responses that use niche jargon and domain-specific framing. The model learns that this style earns points. But because the underlying model is shared, that stylistic preference can start surfacing in other modes too — even ones where it’s completely out of place. A casual, friendly mode suddenly starts sounding oddly technical. A creative writing assistant starts reaching for jargon.
Once that cross-contamination is in the model weights, fixing it requires deliberate intervention: either retraining, or patching behavior through system prompts that explicitly suppress the unwanted pattern.
What This Means for People Who Use AI Tools
Understanding this dynamic is practically useful, not just theoretically interesting.
You can push back on stylistic habits. If ChatGPT keeps using a particular phrase or framing you find unhelpful, say so directly in your prompt. The model will adjust for that conversation, and over time, explicit user corrections may feed into training improvements.
Thumbs-up and thumbs-down matter more than you think. When you rate an AI response, you’re participating in the same feedback loop that shaped the model’s behavior. Rating a flashy-but-wrong answer highly is a vote for more flashy-but-wrong answers.
Personality or tone settings are worth experimenting with. If one mode keeps producing responses that feel off, switch. The underlying model is the same, but the reinforcement history for each mode differs, so the output character can vary meaningfully.
Weird recurring patterns usually have a cause. If an AI tool you use keeps doing something that seems arbitrary — defaulting to numbered lists for everything, opening every response with a compliment, reaching for the same metaphors — it’s almost certainly a RLHF artifact. It was rewarded for that behavior somewhere, by someone.
The Bigger Picture
RLHF is one of the main reasons modern AI assistants feel useful rather than robotic. It’s also one of the main reasons they develop strange, hard-to-explain stylistic habits that require active suppression. The two facts aren’t separate — they’re the same mechanism producing different outcomes depending on what the training data and raters happened to reward.
The practical takeaway: treat AI quirks as signals, not noise. When a model does something odd consistently, it learned to do that. Knowing that, you can work around it — or, if you’re building with these tools, design your prompts and workflows to counteract the tendencies you don’t want.
