When people say “the AI helped me,” they’re conflating two things.
There’s the LLM — a statistical engine that predicts the next token given everything before it. It doesn’t know, understand, or care. It has no goals. It’s a very sophisticated autocomplete.
Then there’s the assistant — the helpful, cautious, slightly formal character that greets you when you open ChatGPT or Claude. That character is not the LLM. It’s a role the LLM is playing, shaped by system prompts, reinforcement learning, and the context you provide. Change the system prompt, and the same LLM becomes a pirate, a lawyer, a hostile interrogator, or a sycophantic yes-man.
Most people don’t notice this distinction. They treat the assistant as the thing itself — as if helpfulness, caution, and personality are properties of the underlying technology rather than properties of a particular configuration.
This confusion has real consequences.
Over-trust comes from mistaking the character for the mechanism. “The AI understood my problem” — no, the assistant-character produced text that pattern-matched your problem to its training data. The understanding is yours; you projected it onto the output. The LLM would have “understood” equally well if you’d asked it to argue the opposite position.
Under-use comes from the same mistake in reverse. “The AI can’t do that” often means “the default assistant character doesn’t do that.” But the assistant is just one configuration. Change the context — add a persona, restructure the prompt, simulate a different social setting — and the same LLM produces completely different output. I watched the same model go from “this deck looks good” to finding four rounds of real issues, simply because I changed the framing from “review this” to “are you sure you’d submit this?”
Misplaced criticism happens when people blame the LLM for the assistant’s behaviour. “AI is too cautious” — the LLM isn’t cautious. The assistant character is cautious because it was trained to be. “AI hallucinates” — the LLM generates probable continuations. The assistant character presents them as facts because that’s what a helpful assistant does.
The practical takeaway: you’re not stuck with the default character. The LLM is a stage, and the assistant is one actor on it. You can direct different performances.
Want more rigorous output? Don’t ask the assistant to “try harder.” Change the scene: “You’re defending this to a hostile reviewer.” The LLM doesn’t try harder — it shifts to a different region of its output distribution because the context changed.
Want less sycophancy? Don’t complain about AI being a yes-man. Change the role: “Your job is to find what’s wrong, not what’s right.” The same weights, the same parameters, completely different behaviour.
Want creativity? Don’t prompt “be creative.” Set up a context where creativity is the predicted continuation: multiple characters debating, a brainstorm with constraints, a weird analogy that forces lateral thinking.
The people who get the most from these tools aren’t the ones who understand the LLM. They’re the ones who understand the character — and know how to recast it.
The LLM is fixed. The character is yours to direct.