What Is ChatGPT? How It Works, What It Is Good At and Where It Fails

Key takeaways

• ChatGPT is an AI chat assistant built on large language models from OpenAI.
• It generates likely text; it does not look up truth by default, so it can produce confident mistakes.
• Good prompts state the goal, context, format and constraints.
• Do not paste confidential, personal or client data unless your organisation's policy and settings allow it.
• Features, models and plans change often. Check OpenAI's help pages for the current options.

What is ChatGPT?

ChatGPT is a chat-based AI assistant made by OpenAI. You type a question or instruction, called a prompt, and it replies in natural language. It was released to the public on 30 November 2022 and quickly became one of the most used AI tools. Over time it has gained abilities such as handling images, voice, files and web search, depending on the plan and model. Because OpenAI changes models, features and limits regularly, this article avoids naming a "current" model. See OpenAI's help centre for what is available now.

How does ChatGPT work?

ChatGPT is built on a large language model (LLM), a neural network trained on very large amounts of text to predict what comes next. The core ideas are:

1. Tokens: text is split into small pieces called tokens.
2. Transformer architecture: the model uses attention to weigh which earlier tokens matter for the next one. The idea comes from the 2017 paper "Attention Is All You Need".
3. Pre-training: the model learns patterns of language, facts and reasoning styles from a large body of text.
4. Fine-tuning and feedback: it is then trained to follow instructions and give helpful answers, including with human feedback.
5. Generation: for each prompt, it produces a reply one token at a time, choosing likely continuations.

Because it works by prediction, fluent text is not the same as verified truth. That single fact explains most of its limitations.

What is ChatGPT good at?

• Drafting and editing emails, reports and study notes.
• Explaining concepts in different ways, such as "explain TCP like I am new to networking".
• Summarising long text.
• Helping with code: writing small functions, explaining errors, suggesting tests.
• Brainstorming, outlines and practice questions.
• Translating and changing tone.

Where does ChatGPT fail?

• Hallucinations: it can invent facts, quotations, sources and function names that sound real.
• Out-of-date knowledge: without search tools, it only knows what it saw in training.
• Arithmetic and exact logic: it can slip on long calculations unless it uses a calculator or code tool.
• Bias: it can repeat biases from its training data.
• Overconfidence: the tone stays certain even when the content is wrong.

How do you write better prompts?

Then iterate. Ask it to correct itself, give an example of what you want or paste your own text and ask for edits.

Is ChatGPT free, and is it safe and private?

OpenAI has offered free and paid tiers, and the details change, so check its pricing page rather than relying on a blog. On privacy, treat any chatbot like an external service. Do not enter passwords, client data, source code you are not allowed to share or personal identity numbers. Settings may let you control whether chats are used for model improvement, and business plans have different terms. Follow your employer's AI policy. Also remember that attackers can try to trick AI tools with hidden instructions, known as prompt injection.

How can students and professionals use it responsibly?

1. Use it to learn rather than simply copy. Try the exercise first, then ask for feedback.
2. Verify technical claims in official documentation.
3. Test any code in a safe environment before using it.
4. Do not submit generated work as your own where rules forbid it.

Common mistakes

• Trusting numbers, dates and citations without checking.
• Giving vague prompts and blaming the tool.
• Pasting sensitive information into it.
• Believing it understands your company or system without being told.

Next steps

To understand the models behind such tools, read what is new in OpenAI GPT-5 and useful hidden ChatGPT features. If you want to build with LLMs, see our LLM engineering course.

Related reading

• ChatGPT image generation: what it does, how to use it and its limits
• ChatGPT Explained: Your AI Companion for Every Task