The short answer
An AI chatbot is a program you type a question into and get sentences back. Underneath it is a large language model: a system that has read an enormous amount of text and learned, in extraordinary statistical detail, which words tend to follow which other words in context.
That is genuinely all it does at the mechanical level. It predicts the next fragment of text, over and over, until the answer is finished. What makes it feel like understanding is that predicting the next word well, across billions of examples, turns out to require something that behaves a great deal like knowledge of grammar, facts, reasoning patterns and tone.
It is not a search engine, and this trips people up more than anything else. A search engine finds a page somebody wrote. A chatbot composes an answer that has never existed before. That is why it can help with your specific situation, and also why it can state something false with complete confidence.
How an answer actually gets made
Four things happen between your question and the reply, and knowing them explains most of the behaviour people find mysterious.
- Your text is broken into tokens
- Roughly word-fragments. This is why a model can miscount the letters in a word: it never saw the letters, it saw the chunks.
- The conversation is read as one block
- Not just your last message, but as much of the thread as fits in its context window. This is why a long conversation stays coherent, and why a very long one starts forgetting the beginning.
- It may go and look things up
- A modern chat can search the web mid-answer. When it does, the answer should arrive with sources next to it, and an answer with sources can be checked in a way that one without cannot.
- It writes one fragment at a time
- That is why text appears to type itself, and why asking it to think step by step often produces a better answer — more intermediate text genuinely gives it more room to work.
Why it sometimes invents things
The failure has a name — hallucination — and an unglamorous cause. A model trained to produce plausible text will produce plausible text even when it has no basis for it. Fluency and accuracy are separate properties, and only one of them is what the system was optimised for.
It shows up most in the places you would expect: exact figures, citations, case law, quotations, prices, dates, anything obscure, and anything that happened after the model was trained. It shows up least in explanation, restructuring, summarising a document you provided, and drafting.
The practical defence is not to stop using it. It is to ask for sources on anything factual, to prefer answers that link out, and to check the two or three claims in a reply that would actually cost you something if they were wrong. Everything else can be read the way you would read a well-informed colleague talking quickly.
Free, paid, and what the money buys
Nearly every assistant has a free tier and a paid tier around twenty dollars a month. The money does not usually buy a different kind of intelligence; it buys access to the newest and largest models, higher limits, longer documents, and features like deep research or file handling.
- Which models you reach
- Free tiers usually run smaller or older models. The gap is real on hard reasoning and mostly invisible on everyday questions.
- How much you can ask
- Limits are the honest differentiator. Here it is 600 messages a month free and 10,500 on the paid plan, published as numbers rather than described as “generous”.
- Whether one lab or several
- Most subscriptions give you one company's models. A router-based product gives several, which matters because no single lab is best at everything.
- What happens to your text
- Read this part before paying. Some services train on your conversations by default; some let you opt out; some encrypt history so it cannot be read at all.
For reference, BlueAI charges $5 a month rather than twenty, gives access to models from five labs in one thread, and encrypts conversations in the browser before they are stored. The free plan needs no card and no email address.
What it is genuinely good at
A useful rule: it is strong wherever the raw material is language and the answer can be checked, and weak wherever it needs to be exactly right about something obscure and has nothing in front of it.
- Explaining
- A contract clause, a diagnosis letter, an error message, a concept a teacher went past too quickly. Ask for a simpler version and it will give you one.
- Changing the shape of text
- Notes into an outline, an outline into a draft, a draft into a summary, a summary into an email. This is the most reliable thing it does.
- Comparing and researching
- Several sources at once, with the sources attached. Much faster than reading twelve tabs, provided you spot-check the conclusions.
- Working with your own files
- A PDF, a spreadsheet, a screenshot. Questions about a document you supplied are the safest category there is, because the answer is grounded in something real.
Frequently asked questions
A program that takes your question as text and writes an answer, using a model that has learned patterns from a very large amount of writing. It composes the reply rather than retrieving it from a page.
No. A search engine points you at pages other people wrote. A chatbot writes something new for your situation, which is more useful and less verifiable — which is why sources next to an answer matter.
Because it is built to produce plausible text, and plausibility and truth are different targets. It is most likely to invent exact figures, citations and recent events, and least likely to when summarising something you gave it.
Yes, for most everyday work. Free tiers typically use smaller models with tighter limits. BlueAI's free plan is 600 messages a month with no card required.
That depends entirely on the service. Check whether conversations are used for training and whether history is encrypted. On BlueAI, conversations are encrypted in the browser before they are stored and are not used for training.
No. Writing the question the way you would say it out loud works. A few habits help, and they are in the prompts guide, but nothing here requires special phrasing.
Keep reading: how to check an answer, how to write prompts, getting started, the learning hub.