Talking to AI feels surprisingly natural. We ask questions, explain problems, sometimes even apologise. But here’s the surprising part — Large Language Models (LLMs) weren’t originally built for conversations. Their core job is much simpler: predict the next word (or token).
During pre-training, LLMs read trillions of words from books, websites, research papers, code, and other publicly available text. By repeatedly predicting the next token, they learn grammar, facts, reasoning patterns, and language. They are then fine-tuned using techniques such as Instruction Tuning and Reinforcement Learning from Human Feedback (RLHF), where humans rank responses to make the model more helpful, safe, and conversational.
But How Do Scientists Know If AI Is Actually Good?
That’s harder than it sounds. AI can produce beautifully written answers that are completely wrong — a phenomenon known as hallucination.
To reduce this, researchers use techniques such as LLM-as-a-Judge, where another AI model evaluates the response, and multi-agent systems, where specialised AI agents review, critique, and improve each other’s work before presenting the final answer.
Human feedback is equally important. Most users never click 👍 or 👎, so researchers also analyse conversations. Satisfied users naturally write things like:
“Thanks!”
“That was helpful.”
“Exactly what I needed.”
These conversational signals, together with follow-up questions, corrections, and explicit ratings, help researchers evaluate whether AI is genuinely helping people.
Does Saying “Please” Actually Make AI Respond Better?
Surprisingly, sometimes yes — but not because AI has feelings.
Before understanding your prompt, an LLM breaks it into tokens. Every token is converted into a high-dimensional mathematical representation called an embedding. Then comes the magic of the Transformer’s attention mechanism.
Each word asks: “Which other words in this sentence should I pay attention to?”
Adding a single word like “please” slightly changes those attention patterns and therefore changes the overall context. During training, the model has seen millions of sentences such as:
Could you please explain…
Would you mind summarising…
Please help me understand…
These phrases usually appear in educational material, tutoring conversations, customer support, and professional communication. As a result, the model learns that this context is often followed by patient, detailed explanations.
That’s why:
Explain quantum computing.
and
Could you please explain quantum computing to a beginner with a simple example?
may produce different responses — not because of the word “please”, but because the entire prompt signals a teaching context. Similarly, very aggressive prompts can sometimes lead to slightly different responses — not because the AI feels offended, but because hostile language statistically appears alongside arguments, sarcasm, or incomplete requests in its training data.
The Real Lesson
The future of AI isn’t just about prompt engineering. It’s about feedback engineering — learning how to collaborate with AI effectively.
A good manager doesn’t simply assign work. They review it, question it, correct it, and provide feedback. Increasingly, that’s exactly how humans will work with AI agents.
So perhaps saying “thank you” isn’t teaching AI to be polite. It’s reminding us how to collaborate.
AI doesn’t respond to kindness. It responds to context.
And “please” is simply one small piece of that context.
Curious about the human side of this? Read the companion piece: Why We Say Please to AI — The Psychology behind it.