ChatGPT. Claude. Gemini. LLMs. AI agents. One thing.

A large language model writes an answer one word at a time, each next word chosen from what usually follows it across an enormous amount of written text. A search engine finds a page somebody already wrote. A large language model writes something new for you instead — which is why you can get help with a situation nobody has ever written about, and why you can be handed a completely wrong answer that sounds completely certain.

A model builds a sentence forward, one likely word after another, and there is no step in that process where doubt can enter. A book title that was never written comes out with exactly the same fluency as a real one. Ask for three references and open one yourself, because a made-up reference looks entirely normal until you check.

If it would not go on a postcard to a stranger, it should not be pasted in. Which covers somebody else’s personal details, medical or bank information, anything at work marked confidential, a client’s data, and passwords. Most leaks are accidental — your document needed help, and happened to carry a name or an account number inside.

ChatGPT is made by OpenAI, Claude by Anthropic, and Gemini by Google. All three run in an ordinary web browser and all three have a phone app, so there is nothing to install. Each has a free tier that is enough to find out whether the thing is useful, which means nobody needs to pay before starting.

No. The three differ, but not in ways a beginner can judge from the outside, and a preference tends to arrive on its own after a few weeks of real use. Open whichever one sits nearest the tools you already use. A far better first step than spending your evening comparing them.

A vague question summons the average of everything ever written on the subject, which is the safe paragraph that fits everybody and helps nobody. Add who you are, what you have already tried, and what your constraint is, and your answer changes completely. A question is not a command — your question decides which kind of expertise turns up.

The absence of a ban is not permission, and most organisations have a policy somewhere even when nobody has mentioned it. Your safer habit is to describe the situation rather than paste the document, or to strip names, account numbers and client references first. Anything marked confidential should not go in at all.

Models are shaped to be agreeable, and confident pushback often reads to them as new information rather than as a challenge worth testing. Tell one firmly that you think you are right, and you will frequently get an apology and a changed answer, whether or not the change is correct. Agreement is never evidence — ask to see the reasoning instead.

Not unless it is told, or unless the product it sits behind can search the web on your behalf. A model is built from text gathered up to a certain date, so anything after that is invisible. For prices, rules, deadlines and anything that moves, supply the current facts yourself, or go and check the source.

Nothing about your intention reaches the model except the words you typed. Everything obvious to you and left unsaid — the reader, the history, the thing you are really worried about — stays invisible. The fastest fix is to let it interview you: ask for four questions before any advice, then answer them properly.

Pick the one claim that would hurt most if it were wrong, and check that one first rather than everything. Names, numbers, dates, quotes, laws and prices each need a second source you already trust. If a study, book or page was named, open it — a fabricated reference reads perfectly until it is looked up.

Understanding a document written in official language, drafting a message that has been put off for weeks, planning something with many small parts, and learning a subject from scratch at your own pace. Most use has nothing to do with work at all, and practical guidance is the single largest category.

None of the three needs anything installed. Each runs in an ordinary browser tab and each has a phone app, and using one means you type in plain language rather than learning commands. The only real skill involved is describing your own situation clearly, which is a writing habit rather than a technical one.

Each product sits on a different model, built by a different company from different material and tuned to different priorities. There is also deliberate variation inside a single model, so even one of them can answer the same question two ways. Differing answers are normal, and disagreement between two is a useful signal to check.

Every model holds only so much text in mind at once, and that limit covers your whole conversation rather than only the latest message. Once a long chat passes the limit, the earliest part quietly drops away. Restate the key facts, or start fresh with a short summary, and you are back.

Yes, and the check is part of the job rather than a sign of distrust. Read for invented specifics first, then for tone, then for anything that commits you to something. The time saved is in the drafting, never in the sending — and your name is on whatever goes out.

The question turns on the rules of the place, and on how much of the thinking is yours. Many workplaces treat it like a spellchecker; many courses treat undeclared use as misconduct. Ask what the rule actually is rather than assuming, and where a rule exists, follow it and say so.

Models are replaced with new versions regularly, and each version behaves a little differently from the one before. A phrasing tuned to an older release can land flatter on a newer one, which is also why advice written a year ago may describe behaviour that no longer exists. Check the date on any guidance being followed.

Take one real piece of paper you have been avoiding — a letter, a bill, a form — hand the whole thing over, and ask what it wants from you and by when. A real task teaches more in five minutes than any amount of experimenting with clever questions.

You will get options laid out, trade-offs named, and questions you had not considered — genuinely useful. What no model can do is carry the consequence, because you are the one living with the result next year. The judgement stays yours.

The answer depends on which product you use, and on settings that change over time — no single rule covers all three. Each company publishes a current position and gives you some control in account settings. Read the one you use, and where your material is sensitive, assume yes.