What is ChatGPT, and how is it different from a search engine?

Yesterday OpenAI put a chatbot called ChatGPT online at chat.openai.com, and it is free to use while the company calls it a research preview. You type a message, it types back, and you can keep the conversation going: ask a follow-up, ask it to try again, or tell it that it got something wrong. OpenAI says the point of the public release is to collect feedback on what the system does well and where it fails.

The first thing most people do is treat it like Google. That works some of the time, and it is also the fastest way to be disappointed. ChatGPT is not looking anything up. It is a language model: a program trained on a very large amount of text to predict which words should come next. That difference shapes everything about how to use it.

What ChatGPT actually is

ChatGPT comes from a model in OpenAI's GPT-3.5 series, which finished training early this year. OpenAI describes it as a sibling of InstructGPT, an earlier model tuned to follow written instructions. To make it good at conversation, OpenAI used a method called reinforcement learning from human feedback, or RLHF. In plain terms: human trainers wrote sample conversations playing both the user and the assistant, then ranked several answers from the model from best to worst. Those rankings were used to nudge the model toward the kind of replies people preferred.

The result is a model that keeps track of the conversation. If you ask about a theorem and then say "how is it used in cryptography?", it understands that "it" means the theorem. OpenAI's launch post also shows it asking for more detail before fixing a piece of code, and rewriting a friendly note in a more formal tone when asked.

How it differs from a search engine

A search engine and ChatGPT answer very different questions behind the scenes.

	Search engine	ChatGPT
What it does	Finds existing pages that match your words	Writes a new reply, one word at a time
Where answers come from	The live web, ranked by relevance	Patterns learned from its training text
Recent news	Yes, within minutes	No; its knowledge thins out after 2021
Shows sources	Always, as links	No links; you cannot see where an idea came from
Best at	Facts you will verify by clicking	Explaining, drafting, rewriting, brainstorming

The Verge's first look at the release notes that the bot's own interface warns it may produce incorrect or misleading content, and that its knowledge of events after 2021 is limited. The same report shows the bot inventing details in a biography a professor asked it to write about himself. OpenAI's limitations section admits this directly:

"ChatGPT sometimes writes plausible-sounding but incorrect or nonsensical answers." (OpenAI, November 30, 2022)

So the honest comparison is this. A search engine gives you somewhere to look. ChatGPT gives you a fluent first draft of an answer, and you decide whether it holds up.

Why the wording of your question matters

OpenAI notes two other quirks worth knowing on day one. First, the model is sensitive to phrasing: it may say it does not know something, then answer correctly after a small rewording. Second, when your request is ambiguous it usually guesses what you meant instead of asking. Both point the same way. The more you tell it about who you are, what you need and what shape the answer should take, the better the reply.

Here is the difference in practice:

Prompt

Explain photosynthesis.

Prompt

I teach a class of 10-year-olds. Explain photosynthesis in five short sentences they can follow, then give one simple experiment we could do with a houseplant and a sunny window.

The first gets a textbook paragraph. The second gets something you can use on Monday morning.

Try it: Ask ChatGPT the same question twice, once in a single short line and once with two sentences about who the answer is for and what format you want. Compare the two replies side by side.

What should a beginner try first?

Start with tasks where you already know enough to judge the answer. Ask it to explain a topic from your own work as if to a newcomer, and see whether the explanation is right. Paste in a paragraph you wrote and ask for a clearer version, or for a friendlier one. Ask for ten ideas for a birthday party, a blog post or a team meeting, and keep the two that are good. These are the things it does well: language, structure and variations.

Then practise the conversation itself. After the first answer, push back: "make it shorter", "use a table", "you assumed I live in the US; I'm in Canada." Watching it adjust is the quickest way to understand that you are steering a writing tool, not querying a database.

Hold off on using it for facts you cannot check, such as recent events, statistics, quotations, medical or legal details, and anything about a specific person. If an answer matters, confirm it with a real source. Also keep private information out of the chat; it is a public research preview, and OpenAI is gathering what people do with it to improve the system.

Treated this way, as a fast and articulate assistant whose work you review, ChatGPT is already useful for drafting, learning and thinking out loud. Many expect tools like it to change how people write and study over the coming year.