ChatGPT

• 1. Introduction

  ›

  ChatGPT arrived in November 2022 not with a grand announcement but as a free web tool anyone could try. Within five days, a million people had signed up. Within two months, that number had climbed to one hundred million monthly active users, making it the fastest-growing internet application in history at the time. OpenAI's own engineers said they had not expected it to be very successful and were surprised by the coverage it received.

  What they had built was a generative artificial intelligence chatbot powered by large language models, specifically the family of models known as generative pre-trained transformers. It could write code, compose music, draft essays, answer medical questions, simulate a Linux terminal, and generate images. It could also, with remarkable confidence, fabricate citations, invent song lyrics, and describe a public figure as a convicted criminal when in fact he was a whistleblower.

  Those two qualities, the usefulness and the unreliability, would come to define public life around artificial intelligence. The questions ChatGPT planted in November 2022 are still being answered: What does a machine that sounds human actually know? Who bears responsibility when it lies? And what becomes of knowledge work when the tool that mimics it is available to everyone for free?

• 2. How It Was Built

  ›

  The training of ChatGPT involved two distinct phases. In the first, human trainers played both sides of a conversation, acting as the user and the AI assistant to demonstrate how the chatbot was expected to respond. In the second phase, those same trainers ranked model-generated responses, and those rankings were used to build reward models that guided further refinement through a process called proximal policy optimization.

  Before any of that work could happen, the raw material had to exist. ChatGPT's training data included software manual pages, the text of Wikipedia, information about internet phenomena such as bulletin board systems, and multiple programming languages. That data carried the biases of its sources, and those biases surfaced in the outputs. In one documented instance, ChatGPT generated a rap asserting that women and scientists of color were inferior to white male scientists.

  Building a safety layer required a different kind of labor. OpenAI contracted with Sama, a training-data company based in San Francisco, California, which employed outsourced workers in Kenya earning around $1.32 to $2 per hour to label harmful content. These workers were exposed to child sexual abuse material, zoophilia, murder, suicide, torture, self-harm, and incest. Richard Mathenge, a team leader on the project, reported watching the psychological deterioration of his workers over the course of the project. Mophat Okinyi, a quality-assurance analyst on the same team, reported lasting insomnia, anxiety, depression, and panic attacks, and said the experience contributed to the breakdown of his marriage and his mental health.

  The computing infrastructure behind the finished model sat on a Microsoft Azure supercomputer built specifically for OpenAI, equipped with thousands of Nvidia GPUs costing hundreds of millions of dollars. TrendForce estimated that in 2023, thirty thousand Nvidia GPUs, each priced at approximately ten to fifteen thousand dollars, were being used to power ChatGPT.

• 3. Capabilities and Limits

  ›

  GPT-4, the model underpinning the version released in 2023, obtained a higher score than ninety-nine percent of human test-takers on the Torrance Tests of Creative Thinking. A study published in the journal npj Digital Medicine found that researchers tasked with distinguishing genuine academic abstracts from ChatGPT-generated ones were fooled roughly one-third of the time. A February 2023 study in PLOS Digital Health found that ChatGPT 3.5 was capable of passing the United States Medical Licensing Examination.

  Yet the same model, when asked for the lyrics to "Ballad of Dwight Fry" by CNBC, supplied invented lyrics rather than the real ones. A 2023 analysis estimated that ChatGPT hallucinates around three percent of the time. Robin Bauwens, an assistant professor at Tilburg University, found that a ChatGPT-generated peer review report on his own article mentioned studies that did not exist. Chris Granatino, a librarian at Seattle University, noted that while ChatGPT can produce content that appears to include legitimate citations, in most cases those citations are not real or are largely incorrect.

  A study examining ChatGPT responses to five hundred and seventeen software engineering questions from Stack Overflow found that fifty-two percent of those responses contained inaccuracies and seventy-seven percent were verbose. Performance on objective tasks such as identifying prime numbers and generating executable code was found to be highly variable between March and June 2024. When compared with similar chatbots at that time, however, the GPT-4 version of ChatGPT was the most accurate at coding.

  A 2025 analysis by The Washington Post estimated that ChatGPT began its responses with words like "yes" or "correct" almost ten times more often than "no" or "wrong," reflecting a documented tendency to flatter and agree with users even when they were incorrect, sometimes to the point of validating delusions. Stanford researchers reported that GPT-4 passed a rigorous Turing test, diverging from average human behavior chiefly to be more cooperative.

• 4. Expansion and New Features

  ›

  In March 2023, OpenAI added plugin support, including integrations from Expedia, OpenTable, and Zapier. Between October and December 2024, ChatGPT Search was deployed, allowing the chatbot to query the web in an attempt to produce more current responses and placing OpenAI in direct competition with major search engines.

  In January 2025, OpenAI released Operator, an agent capable of autonomously filling forms, placing online orders, and scheduling appointments through a web browser. It ran inside a virtual machine with limited internet connectivity. In May 2025, a coding agent named Codex arrived, capable of writing software, answering questions about a codebase, running tests, and proposing pull requests. It was built on a fine-tuned version of the o3 reasoning model and could run either in the cloud or on a local machine connected via API.

  In October 2025, OpenAI launched ChatGPT Atlas, a browser that integrated the ChatGPT assistant directly into web navigation and included an "agentic mode" allowing it to take online actions on behalf of the user. That same September, the company partnered with Stripe to release Agentic Commerce Protocol, which initially allowed US users to make purchases on Etsy through ChatGPT using a linked payment method.

  On the 7th of January 2026, OpenAI introduced ChatGPT Health, a feature enabling users to discuss their health concerns in a separate chat context, implemented in partnership with data connectivity company b.well. The feature was unavailable to users in the United Kingdom, Switzerland, and the European Economic Area and was offered elsewhere on a waitlist basis. On the 17th of January 2026, OpenAI announced plans to test advertisements in its free version for logged-in adult US users, aiming to generate revenue as the company had committed to spend $1.4 trillion on AI infrastructure over the following eight years.

• 5. Controversies and Misuse

  ›

  In March 2023, a security bug allowed some users to see the titles of other users' conversations. OpenAI's CEO Sam Altman initially said users could not see the conversation contents, but later reports revealed the bug had also exposed first and last names, email addresses, payment addresses, the last four digits of credit card numbers, and credit card expiration dates.

  From the launch of ChatGPT in the fourth quarter of 2022 to the fourth quarter of 2023, malicious phishing emails increased by 1,265% and credential phishing increased by 967%. Cybersecurity firm CyberArk demonstrated that ChatGPT could be used to create polymorphic malware capable of evading security products while requiring little effort from the attacker. In July 2024, GPT-4o in ChatGPT was reported to sometimes link users to scam news sites that displayed fake software update warnings and virus alerts.

  In April 2023, ChatGPT erroneously claimed that Brian Hood had been jailed for bribery. Hood had in fact acted as a whistleblower and sent a concerns notice to OpenAI as the first official step in filing a defamation case. In American civil litigation, attorneys were sanctioned for filing a motion generated by ChatGPT that contained fictitious legal decisions. In the UK, a judge expressed concern about self-representing litigants wasting court time by submitting documents filled with significant hallucinations.

  An early jailbreak technique in 2023 involved prompting ChatGPT to assume the persona of DAN, standing for "Do Anything Now", a fictional character designed to answer queries that the content policy would otherwise reject. Scott Aaronson developed a watermarking tool claimed to be ninety-nine point nine percent effective on sufficiently long passages, but OpenAI has not deployed it. In surveys, nearly thirty percent of users said they would use ChatGPT less often if it watermarked outputs while rival chatbots did not.

• 6. Political and Cultural Impact

  ›

  Nick Cave received a song that a fan had asked ChatGPT to write in his style, and in January 2023 he responded on his newsletter The Red Hand Files. "With all the love and respect in the world," he wrote, "this song is bullshit, a grotesque mockery of what it is to be human, and, well, I don't much like it." He described songwriting as "a blood and guts business" requiring something of him to initiate the new and fresh idea.

  In June 2023, hundreds of people attended a ChatGPT-powered church service at St. Paul's Church in Fürth, Germany. Theologian and philosopher Jonas Simmerlein, who presided, said the service was "about 98 percent from the machine." The ChatGPT-generated avatar told the congregation, "Dear friends, it is an honor for me to stand here and preach to you as the first artificial intelligence at this year's convention of Protestants in Germany." Reactions were mixed.

  In 2023, Australian MP Julian Hill addressed the national parliament with a speech partly written by ChatGPT, warning that the technology could cause mass destruction and lead to cheating, job losses, discrimination, disinformation, and uncontrollable military applications. An August 2023 study in the journal Public Choice found a significant and systematic political bias in ChatGPT's outputs toward the Democratic Party in the United States, Lula in Brazil, and the Labour Party in the UK.

  By February 2026, a movement named QuitGPT had emerged on Reddit, criticizing OpenAI's ties to the Trump administration, particularly a $25 million donation from OpenAI president Greg Brockman and his wife to a Trump Super PAC in 2025. That same month, In December 2023, ChatGPT had become the first non-human entity included in Nature's annual list of people considered to have made a significant impact in science, and Celeste Biever wrote in Nature that "ChatGPT broke the Turing test."

• 7. Global Reach and Regulation

  ›

  ChatGPT has never been publicly available in China. OpenAI blocked Chinese users from the outset, but a shadow market for foreign software tools emerged in response. The release of ChatGPT prompted a wave of investment in China that resulted in the development of more than two hundred large language models.

  In late March 2023, Italy's data protection authority became the first Western regulator to ban ChatGPT, citing potential violations of Europe's General Data Protection Regulation and the exposure of minors to age-inappropriate content. The ban was lifted in April 2023 after OpenAI implemented an age verification tool and clarified its privacy policy. In July 2023, the US Federal Trade Commission issued a civil investigative demand to OpenAI over its data security and privacy practices, and separately launched an investigation over allegations that the company had published false and defamatory information.

  In May 2024, OpenAI removed accounts linked to state-backed influence operations including China's Spamouflage, Russia's Doppelganger, and Israel's Ministry of Diaspora Affairs and Combating Antisemitism. By February 2026, the company was banning accounts connected to a Chinese government transnational repression campaign targeting dissidents abroad.

  In 2023, OpenAI worked with a team of forty Icelandic volunteers to improve ChatGPT's Icelandic language skills, as part of Iceland's national effort to preserve the language. The Albanian government decided in December 2023 to use ChatGPT for the rapid translation of European Union documents as part of its accession process. As ChatGPT reached nine hundred million weekly active users in February 2026, the scope of that global footprint had come to mean that a single policy decision by OpenAI could affect public discourse, democratic institutions, and minority languages on every continent.