What is ChatGPT?

One of the most well-known implementations of conversational AI, ChatGPT runs on OpenAI’s generative pretrained transformer (GPT) models. As of this writing, ChatGPT is based on the multimodal GPT-4o AI model, enabling it to generate images and hold audio conversations in addition to text generation. 

ChatGPT has garnered significant media attention since its initial release in 2022. Its popularity has contributed much to the current AI boom, including the development of competing products such as Google Gemini (formerly known as Bard) and Anthropic’s Claude. 

However, ChatGPT has also inspired concern around issues related to large language model (LLM) technology, including plagiarism and the use of generative AI to create misinformation and replace human workers.

ChatGPT was developed by OpenAI, a machine learning and AI startup founded by CEO Sam Altman, President Greg Brockman and others including Elon Musk and Ilya Sutskever. In addition to ChatGPT and the GPT model family, OpenAI also created the DALL-E image generation model and text-to-video model Sora. 

OpenAI was initially created as a nonprofit entity but now exists as a hybrid venture: the 501(c)(3) nonprofit OpenAI, Inc. controls a holding company that is the majority owner of the for-profit subsidiary OpenAI Global LLC. Since its founding, OpenAI has received considerable investment from Microsoft, totaling at least USD 13 billion since 2019. 

Much of Microsoft’s investment in OpenAI came in the form of Azure cloud compute credits.¹ In return, Microsoft provides access to the OpenAI API—which other developers can use to create their own generative AI tools—through its Azure cloud computing platform and shares in OpenAI’s profits. The Microsoft Copilot and Bing AI services are also powered by GPT models.

ChatGPT works by using generative pretrained transformers (GPTs), a type of large language model (LLM) developed by OpenAI. Transformers are a type of deep learning neural network architecture that specializes in maintaining context over large inputs through self-attention mechanisms that enable the models to identify the most relevant portions of an input sequence. 

Broadly speaking, ChatGPT creates responses through a two-stage process: 

1. Language understanding: The model converts user input into a numerical representation through the embedding process. Embedding enables ChatGPT to capture and process user intent and semantic and syntactic meaning. 
2. Response generation: GPT models combine the input sequence with the greater context of the user interaction to come up with a response. The response is informed by the knowledge that the model gained through its training process. The model predicts its responses based on similar content in its training data. 
   

AI-powered chatbots such as ChatGPT work by using complex machine learning algorithms that compare user input to the contents of their datasets. During training, GPTs identify patterns and relationships in the training data and then use those findings to predict real-world outcomes—such as the next word in a sentence. ChatGPT’s outputs represent its predictions based on the patterns of its training data. 

GPT-3.5, the model that powered the first version of ChatGPT, was part of the third generation of the GPT model family, with the first introduced in 2018. Now with access to GPT-4o, users can use ChatGPT to generate both audio and image content. Users can access ChatGPT through their internet browser or through apps on macOS, Windows, iOS and Android. 

ChatGPT is trained on huge datasets of content taken from the internet. This includes forum posts, news articles, images and websites. For example, GPT-3 was trained on over 45 terabytes of text data, including the entirety of Wikipedia at the time. This enables it to understand the patterns of human-generated language and understand the relationships between different topics. 

GPT models are trained with a mixture of techniques including supervised learning, unsupervised learning and reinforcement learning from human feedback (RLHF), in which human trainers rank the model’s responses and create reward mechanisms for optimal performance. The initial pretraining stages are unsupervised, with RLHF employed later for fine-tuning. 

ChatGPT is further fine-tuned from the base GPT models with additional conversational datasets, such as one consisting of movie dialogues. The unsupervised training process forces the model to digest reams of unstructured data and come to its own conclusions about the patterns and meaning in that content. 

ChatGPT responds to human feedback. Users can instruct ChatGPT to behave in specific ways, such as a therapist or career guide. User feedback also extends to thumbs-up and thumbs-down buttons that can further tailor ChatGPT’s responses. 

ChatGPT is a versatile example of AI technology that can handle a wide range of tasks. While the free version of ChatGPT offers considerable utility, users who subscribe to one of ChatGPT’s paid tiers with an OpenAI account gain access to more powerful GPT models. 

ChatGPT’s use cases include: 

• Content creation: ChatGPT can help compose emails, outline and draft articles, script podcasts and videos, create social media posts and more. 

• Summarization: ChatGPT can process and summarize lengthy articles and reports in the user’s wanted format. Users can also have it create simple explanations of complex topics. 

• Web searches: With access to search engines since October 2024, ChatGPT can include real-time search results in its generated text and audio responses. 

• Job applications: With access to search engines since October 2024, ChatGPT can research companies, find relevant job opportunities and tailor resumes and cover letters to specific roles. 

• Business planning: Users can ask ChatGPT to behave as a business analyst and conduct market research, then create business plans and reports. 

• SEO: ChatGPT can return a list of relevant keywords for a query. Users can then ask it to outline and draft the corresponding article or webpage copy. 

• E-commerce: ChatGPT can draft product descriptions for use on e-commerce websites. 

• Language translation: ChatGPT supports over 80 languages and can also provide machine translation between them. 

• Content marketing: ChatGPT can repurpose content from one channel to another, such as by generating a series of social media posts from an article or white paper. 

• Advice: Users can describe a real-world or hypothetical situation and ask ChatGPT for advice on what to do or how to respond. ChatGPT can be instructed to behave in a specific way, such as a therapist or coach. However, because ChatGPT is a machine learning algorithm and not a trained human professional, its responses should never be regarded as genuine expertise—especially regarding high-stakes topics such as law or healthcare. 

• Create images: One of ChatGPT’s new features is the ability to create AI-generated images. While ChatGPT could previously create images with OpenAI’s DALL-E model, it can now do so on its own. 

• Voice conversations: ChatGPT’s voice mode allows it to interact more naturally with users. Natural language understanding (NLU) and natural language generation (NLG) capabilities lead to AI-powered conversational experiences. OpenAI currently limits the number of voice conversations users can have with ChatGPT. 

• Coding and code debugging: Programmers and developers can ask ChatGPT to review code and potentially find errors or provide missing code snippets. 

In 2023, OpenAI introduced the option to create custom GPTs: chatbots based on ChatGPT that users can modify for specific purposes. Rather than prompt ChatGPT each time with thorough instructions on how to behave, users can fine-tune a custom GPT for a specific use case. 

The advantages of custom GPTs include the ability to upload knowledge files to the GPT builder and share the GPT with other users. With custom GPTs, users can create their own AI systems that add functions beyond the default capabilities of stock ChatGPT. The ability to create custom GPTs is restricted to premium users, though free users can use them.

ChatGPT is available with a free OpenAI account or at several paid tiers. As of publishing, ChatGPT plans include: 

• ChatGPT Free: Free users default to GPT4o-mini with limited access to GPT-4o and GPT-o3 mini, a newer small model released in January 2025. Users can have ChatGPT search the web to find and include real-time data in responses. Voice mode and file uploads are limited. 

• ChatGPT Plus: This tier brings increased usage limits as well as access to OpenAI’s “deep research” and reasoning models, voice mode and a preview of GPT-4.5. Users gain the option to create custom GPTs and have limited access to Sora, OpenAI’s video generation model. 

• ChatGPT Pro: ChatGPT Pro provides unlimited access to all reasoning models and GPT-4o in addition to advanced voice mode. Users get increased limits for video and screen sharing, deep research models and Sora video generation. ChatGPT Pro users can also access o1 “pro mode,” which OpenAI describes as offering “more compute for the best answers to the hardest questions.”²

• ChatGPT Teams:  Teams is the first of two workplace plans for ChatGPT. Organizations can connect internal Google Drive storage for personalized outputs while providing all Plus features to their employees, with higher message limits for GPT-4o. 

• ChatGPT Enterprise: The organization-wide ChatGPT plan offers a greater context window for GPT-4o, allowing users to create larger input sequences with more content and data. Access to OpenAI’s reasoning models is included, while Enterprise data is excluded from model training (OpenAI reserves the right to train models on user data by default). 

• ChatGPT Edu: ChatGPT’s education-focused plan is analogous to ChatGPT Enterprise, but for colleges and universities. This includes the training exemption for input data given by ChatGPT Edu users. 

ChatGPT, and generative AI as a whole, is impressive technology. Users can generate content and automate routine work that might otherwise take up time. But some find drawbacks in the software. The limitations of ChatGPT include: 

• Reliability: Out of ChatGPT’s 400 million weekly active users, only 15.5 million are paying subscribers.³ Some users report a decline in performance since the start of 2025, citing diminished intelligence⁴, memory failures⁵ and shorter replies⁶ among other concerns. 

• Accuracy: ChatGPT does not cite sources when answering questions. Users can prompt ChatGPT to give sources, but those sources are not verified. It’s best to verify all information provided by ChatGPT before use. 

• Transparency: ChatGPT, like all of OpenAI’s products, is closed-source, meaning that its inner workings cannot be verified or examined by third parties. ChatGPT also does not provide citations in its responses. 

• Hallucinations: All generative AI tools are prone to hallucinations or confabulations in which they generate outputs that does not correspond with real-world outcomes. This is due to the underlying algorithm identifying patterns in its training data that do not exist in real life.

  Generative AI models do not know things in the way that humans do. They are highly sophisticated pattern recognition tools, and the best they can do in any situation is make an educated guess. 

• Reasoning: A 2025 study⁷ found that ChatGPT sometimes falls victim to the same logical fallacies that humans do. It is subject to biases, which makes sense considering that it is trained on human-created content. It struggles to perform as well with subjective reasoning tasks as it does with straightforward logic challenges. ChatGPT is also prone to overestimating its own intelligence—another humanlike quality. 

While ChatGPT has enjoyed considerable popularity since its initial launch, the chatbot and its parent company have sparked multiple ethical concerns, such as: 

• Copyright infringement: OpenAI has been sued by multiple publishers, including The New York Times and The Chicago Tribune, for training its models on copyrighted content. In response, OpenAI published an open letter⁸ in which it argued that allowing it and other American AI developers is crucial to overcoming Chinese competition. 

• Plagiarism: AI content detectors have emerged in the wake of generative AI apps such as ChatGPT, but they are not always reliable. Users can have ChatGPT generate an essay and modify it enough to bypass detection. ChatGPT itself can also output copyrighted content in its responses, the use of which also constitutes plagiarism. 

• Potential for misuse: Like any generative AI tool, ChatGPT can be turned toward malicious ends. AI-created deepfakes and misinformation have been deployed around the world in an attempt to influence election results.⁹

• Privacy violations: With a few exceptions, such as its Teams and Enterprise plans, OpenAI collects input data and uses it to further train its models. If users submit prompts that include confidential or sensitive information, that data can potentially appear in outputs given to other users. 

• Environmental impacts: A ChatGPT query uses 10 times as much electricity as an internet search. As more users opt to search with ChatGPT over traditional search engines,¹⁰ this energy demand continues to grow. OpenAI’s new reasoning model, o3, can consume over USD 1,000 of compute power per task.¹¹ A 2021 paper estimated that the training process for GPT-3 produced approximately 552 tons of carbon dioxide¹²—and OpenAI’s models have grown more sophisticated since then.