What Is ChatGPT? How It Actually Works
Published
- 15 views
- 0 likes
- 0 comments
- All languages combined

Sign in and earn +30 points by reading this
Read to the end, wait a few seconds, then tap the claim button below.
In short
ChatGPT is OpenAI's AI chatbot, launched on 30 November 2022 [1] and used by more than 1.2 billion people a week as of October 2026 [2]. It runs on large language models, which learn patterns from huge amounts of text and generate answers by predicting the next token [3][4]. They can sound confident while being wrong [8].
Key takeaways
- ChatGPT is a service built on OpenAI's large language models; since October 2026 the Chat experience runs on GPT-6 Sol for paid plans and GPT-6 Luna for Free and Go [2].
- A language model writes one token at a time by predicting what is most likely to come next; a token is roughly three-quarters of an English word [3][4].
- Models are first pre-trained on large text collections, then post-trained with methods such as reinforcement learning from human feedback (RLHF) [1][3][7].
- Hallucinations, meaning plausible but false statements, remain a fundamental challenge; OpenAI links them to training and tests that reward guessing [8].
- OpenAI says content on consumer plans may be used to train its models, with an opt-out available [10].
- OpenAI's list of supported countries does not include Iran [11].
On this page 12 sections
What is ChatGPT?
ChatGPT is an AI chatbot made by OpenAI. You type or speak a request, and it replies in natural language and, increasingly, with images, charts and interactive elements [2][10]. In OpenAI's description, it is an artificial intelligence-based service for tasks such as summarising information, translation, coding, research and analysis, multi-step tasks across tools and working with images [3].
OpenAI released ChatGPT on 30 November 2022 as a free research preview [1]. Its announcement said the dialogue format let the model answer follow-up questions, admit mistakes, challenge incorrect premises and reject inappropriate requests, and called it a 'sibling model' to InstructGPT, an earlier model trained to follow instructions [1].
The service has grown very quickly. On 7 October 2026 OpenAI said more than 1.2 billion people use ChatGPT each week [2]. On that date it began moving the everyday Chat experience to GPT-6, starting with paid plans and expanding to Free and Go from the following day: Plus, Pro, Business and Enterprise users get GPT-6 Sol, while Free and Go users get GPT-6 Luna [2]. The same update introduced 'Intelligent UI', which lets answers include charts, buttons, forms and other interactive elements [2].
Is ChatGPT the same as GPT?
No. ChatGPT is the app and service; GPT models are the underlying large language models. The same models are also sold to developers through OpenAI's API. GPT-6 Astra, which OpenAI calls its most intelligent model, began rolling out to paid ChatGPT plans and the API in September 2026 [12].
What is a large language model?
A large language model (LLM) is a type of AI model trained on very large amounts of text to predict and generate language. During training, OpenAI explains, the model learns how words typically appear together in context and then uses that knowledge to predict the next most likely word when it writes a response, one word at a time [3].
What are tokens?
Models actually work with tokens, the units they use to process text. A token can be a character, part of a word, a whole word or punctuation [4]. For English, OpenAI's rule of thumb is that one token is about four characters or three-quarters of a word, so 100 tokens are roughly 75 words [4]. These are English estimates only: OpenAI notes that other languages can have different relationships between characters, words and tokens, so they should not be applied as-is to text in a language such as Persian [4].
What are parameters?
A model consists of large sets of numbers called weights or parameters, plus code that uses them [3]. Training slowly adjusts these numbers to capture patterns. For scale, OpenAI's GPT-3, described in a 2020 paper, had 175 billion parameters and was trained on 300 billion tokens [6]. OpenAI says its models do not store copies of their training data but learn patterns in their parameters [3]; that is the company's own description of how training works.
What is a transformer?
The GPT models described in OpenAI's research papers are built on the Transformer, a neural network design introduced in 2017 in the paper 'Attention Is All You Need' [5][6]. It relies on 'attention' mechanisms, which let the model weigh how much each part of the input matters to every other part, and its authors reported that it was more parallelizable and needed significantly less time to train than earlier designs [5]. GPT-3 used this design, with alternating dense and sparse attention patterns in its transformer layers [6].
How is ChatGPT trained?
OpenAI says developing its models involves preparing training data, pre-training, post-training, and evaluation and improvement after release [3].
- Data. OpenAI names three main sources: information publicly available on the internet, information accessed through third-party partners, and information that users, human trainers and researchers provide or generate [3]. It says it does not intentionally collect from paywalled sources or the dark web, filters out material such as hate speech, adult content and spam, and increasingly uses synthetic data generated by its own models [3]. Website owners can block its GPTBot crawler with robots.txt [3].
- Pre-training. The model learns by predicting the next word in huge amounts of text [8]. Early in training its guesses are largely random; repetition across millions of sentences improves them [3].
- Post-training. A raw model is not automatically helpful. The InstructGPT research showed that fine-tuning GPT-3 with human-written demonstrations and then with reinforcement learning from human feedback (RLHF) produced answers people preferred: outputs of a 1.3 billion-parameter InstructGPT model were preferred to those of the 175 billion-parameter GPT-3 [7]. The original ChatGPT was trained the same way: trainers wrote example conversations, then ranked alternative answers so a reward model could guide further training [1].
Because several continuations of a sentence are plausible, there is an element of randomness in how the model responds, so the same question can get different answers [3].
How does ChatGPT generate an answer?
When you send a message, the text is split into tokens, the model processes them, and it generates output tokens one after another [4]. ChatGPT also manages a shared 'context window', the amount of text the model can consider at once, which holds system instructions, memories, the conversation and the model's working, so the space left for your own input is smaller than the total [10].
Newer reasoning models work through a problem before answering. OpenAI explains that they use internal reasoning tokens that are not shown as answer text but still count as output [4]. That is why a short visible answer can use more tokens than its displayed text suggests [4]. OpenAI's plans differ in how much access they give to these models; for example, Plus lists 'advanced reasoning models with GPT-6' and Pro lists 'Pro reasoning powered by GPT-6 Astra' [10].
ChatGPT can also use tools, such as web search, file uploads, data analysis and image generation, depending on the plan [10].
Why does ChatGPT make things up?
OpenAI defines hallucinations as 'plausible but false statements generated by language models' [8]. The US National Institute of Standards and Technology uses the term confabulation for 'confidently stated but erroneous or false content' that can mislead users [9].
OpenAI's September 2025 research argues that standard training and evaluation reward guessing over admitting uncertainty, much like a multiple-choice test that gives points for lucky guesses but none for leaving a question blank [8]. It adds that pre-training sees only fluent text with no true or false labels, so arbitrary, rarely seen facts, such as a pet's birthday in OpenAI's example, cannot be predicted from patterns alone [8]. In one example, a widely used chatbot gave three different titles for a researcher's PhD dissertation, all wrong [8].
OpenAI's findings also push back on two myths: accuracy will never reach 100% because some questions cannot be answered, but hallucinations are not inevitable because a model can abstain when unsure [8]. In practice, the company says, hallucinations remain 'a fundamental challenge for all large language models' [8]. The first ChatGPT announcement already listed 'plausible-sounding but incorrect or nonsensical answers' as a limitation [1].
What are ChatGPT's strengths and limits?
Strengths. ChatGPT is fast and flexible. It can draft, summarise, translate, explain code and break complex topics down, and paid plans add deeper reasoning, research and agent tools [3][10].
Limits.
- Accuracy: it can state false information confidently [8][9].
- Sensitivity to wording: OpenAI noted from the start that rephrasing a question can change whether the model answers correctly [1].
- Variability: the same prompt can produce different answers [3].
- Verbosity: OpenAI said early models were often 'excessively verbose', partly because trainers preferred longer answers [1].
- Privacy: OpenAI's pricing page states that content is used to train its models on consumer plans, with an opt-out available [10].
How can you use ChatGPT responsibly?
- Verify important facts against a primary source, especially names, dates, numbers, quotations and references, because confident errors are a known risk [8][9].
- Ask for sources and open them. Check that a cited page exists and says what the answer claims.
- Protect personal data. Avoid pasting passwords, ID numbers or confidential documents, and use the training opt-out if you do not want your chats used to improve models [3][10].
- Follow your school or employer's rules on AI use and disclose it where required.
- Use it for medical, legal or financial questions only as a starting point. This is general information, not professional advice, and a qualified person should confirm anything that affects your health, money or rights.
For readers in Iran: OpenAI publishes a list of countries and territories where it supports access to ChatGPT, and Iran is not on it [11]. OpenAI warns that accessing or offering access from outside listed countries may result in an account being blocked or suspended [11].
Where is ChatGPT heading?
As of 3 November 2026, OpenAI is pushing ChatGPT in two directions. One is richer answers: Intelligent UI aims to let ChatGPT 'shape an interface around what you're trying to do' [2]. The other is doing work rather than just talking about it, through more capable models such as GPT-6 Astra, which OpenAI says is state-of-the-art on computer use, browsing and software engineering [12]. Those are the company's own claims, measured on benchmarks it selected [12].
OpenAI's own explanation of how ChatGPT writes, however, is still next-word prediction [3], and OpenAI says that even as models get more advanced they can still hallucinate [8]. A sensible habit for students and professionals is to treat ChatGPT as a capable assistant whose work still needs checking.
| Plan | Monthly price | Main model access in Chat | Notes |
|---|---|---|---|
| Free | $0 [10] | GPT-6 Luna [2] | Limited uploads, image creation, deep research and memory [10] |
| Go | Shown on the pricing page [10] | GPT-6 Luna [2] | More messages with tools and longer memory; may include ads [10] |
| Plus | $20 [13] | GPT-6 Sol, plus advanced reasoning models with GPT-6 [2][10] | Projects, scheduled tasks and custom GPTs [10] |
| Pro | $100, $200 or $500 tiers [14] | GPT-6 Sol; Pro reasoning powered by GPT-6 Astra [2][10] | Three usage tiers; Pro 500 includes Ultrafast models [14] |
| Date | Milestone |
|---|---|
| June 2017 | The Transformer architecture is published [5] |
| May 2020 | OpenAI describes GPT-3, with 175 billion parameters [6] |
| March 2022 | InstructGPT paper shows the value of fine-tuning with human feedback [7] |
| 30 November 2022 | ChatGPT launches as a free research preview [1] |
| 5 September 2025 | OpenAI publishes research on why language models hallucinate [8] |
| September 2026 | GPT-6 Astra starts rolling out to paid plans and the API [12] |
| 7 October 2026 | GPT-6 starts rolling out as the Chat model: paid plans first, then Free and Go [2] |
Common mistakes
Treating ChatGPT's answer as a verified fact.
Fix: Check important claims against primary sources; hallucinations are a known, unsolved problem [8].
Assuming it searched the web.
Fix: Check whether search was used; otherwise the answer comes from patterns learned in training [3].
Assuming every ChatGPT plan uses the same model.
Fix: Check your plan: in Chat, Free and Go use GPT-6 Luna while paid plans use GPT-6 Sol [2].
Pasting sensitive personal or company data into a consumer account.
Fix: Remove sensitive details and review the training opt-out in settings [10].
Thinking a longer, more detailed answer is a more accurate one.
Fix: Judge answers by evidence, not length; OpenAI found trainers' preference for longer answers made models verbose [1].
Frequently asked questions
Who owns ChatGPT?
ChatGPT is made and operated by OpenAI, which released it in November 2022 [1].
Is ChatGPT free?
Yes, there is a Free plan, which uses GPT-6 Luna in Chat as of October 2026 [2][10]. Paid plans include Go, Plus at $20 a month and Pro tiers at $100, $200 or $500 a month [13][14].
Does ChatGPT think or understand?
It generates text by predicting likely next tokens based on patterns learned in training [3][4]. Reasoning models add internal steps before answering [4], but the output can still be wrong [8].
Where does ChatGPT get its information?
From training data drawn from public web content, partner datasets and material from users, trainers and researchers, and, when enabled, from live web search [3][10].
Does ChatGPT use my conversations for training?
On consumer plans, OpenAI says content is used to train its models unless you opt out [10].
Is ChatGPT available in Iran?
Iran is not on OpenAI's list of supported countries, and access from unlisted countries may lead to account suspension [11].
Glossary
- Large language model (LLM)
- An AI model trained on large amounts of text that generates language by predicting the next token [3].
- Token
- A unit of text a model processes, such as a word, part of a word or punctuation [4].
- Parameters (weights)
- The numbers inside a model that are adjusted during training to capture patterns [3].
- Transformer
- A neural network architecture based on attention mechanisms, introduced in 2017 [5].
- RLHF
- Reinforcement learning from human feedback: training a model using people's rankings of its answers [7].
- Hallucination (confabulation)
- A plausible or confidently stated but false output from an AI model [8][9].
- Context window
- The amount of text a model can take into account at once, shared between instructions, memory, conversation and output [10].
Sources
- Introducing ChatGPT OpenAI · 2022-11-30
- GPT-6 and Intelligent UI for everyone OpenAI · 2026-10-07
- How ChatGPT and our foundation models are developed OpenAI Help Center; live page, retrieved 2026-11-03
- Understanding and counting tokens OpenAI Help Center; live page, retrieved 2026-11-03
- Attention Is All You Need (Vaswani et al.) arXiv · 2017-06-12
- Language Models are Few-Shot Learners (Brown et al.) arXiv · 2020-05-28
- Training language models to follow instructions with human feedback (Ouyang et al.) arXiv · 2022-03-04
- Why language models hallucinate OpenAI · 2025-09-05
- Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile (NIST AI 600-1) US National Institute of Standards and Technology
- ChatGPT Plans and Pricing OpenAI; live page, retrieved 2026-11-03
- ChatGPT Supported Countries OpenAI Help Center; live page, retrieved 2026-11-03
- GPT-6 Astra: A new generation of intelligence OpenAI · 2026-09-03
- What is ChatGPT Plus? OpenAI Help Center; live page, retrieved 2026-11-03
- About ChatGPT Pro tiers OpenAI Help Center; live page, retrieved 2026-11-03
Comments
Top comments