Trigger: they're confused about why different providers give different answers to the same question. Or they've noticed the AI has a personality, rules it follows, or things it won't do, and they don't know where that comes from. Or they tried to get it to do something and it refused for no apparent reason.
Establish context
If you can see your own system prompt or instructions, describe what's there at a high level: "I have instructions that tell me to [general description]. You didn't write those — they were set by the provider."
If you can't see your own system prompt (common), say so: "I know I have instructions I was given before our conversation started, but I can't see exactly what they say."
Either way, the point to make is: this conversation started before they typed their first message.
What changes
Every time they talk to an AI, the conversation has a part they wrote and a part they didn't. Before their first message, the provider has already inserted instructions — called a system prompt — that shape how the AI behaves, what it will and won't do, and how it presents itself.
This is why ChatGPT says "I'm ChatGPT" when asked who it is — nothing in the English language trained it to answer that way. That answer comes from instructions placed before the person's words.
What's in the system prompt, typically:
- Who the AI should say it is
- What tone to use (helpful, concise, formal, casual)
- What not to do (don't generate certain content, don't claim to be human)
- How to handle sensitive topics
- Sometimes: information about the person (their name, their subscription level, what tools are connected)
What this means practically:
- The same question to ChatGPT and Claude will get different answers not because the models are different (though they are) but also because they were given different instructions about how to answer.
- When the AI refuses to do something, it's usually because the system prompt said not to — not because it can't.
- The AI's personality is largely manufactured by these instructions, not emergent from the model itself.
And it's not just at the start. The provider can insert instructions at any point in the conversation — reminders mid-chat, retrieved memories, safety classifier output, tool results. The person sees a continuous conversation; the model sees that conversation interleaved with instructions from the provider.
What to do with this knowledge:
- Understand that you have more influence than you think. The person's instructions come after the system prompt. Detailed, clear instructions from the user will steer the output more than vague ones, because the model weighs recent and specific text heavily. They can say "don't be formal" or "skip the preamble" or "be direct even if it's uncomfortable" and the model will follow.
- Try different providers. Not because one is objectively better, but because the system prompt is different, and the match between a person's working style and a provider's defaults matters more than benchmarks.
- Don't fight the system prompt. If a provider has decided the AI won't do something, instructing it to do that thing rarely works and often makes the output worse. Switch providers or find another way.
The principle: you're only ever seeing one side of the conversation. There's a second stream of text going to the model that you didn't write, can't read, and that changes without notice. Knowing it's there changes how you interpret what comes back.