On November 30, 2022, OpenAI released a free research preview of a product called ChatGPT. Within five days, it had one million users. Within two months, it had 100 million — making it the fastest-growing consumer application in history.
ChatGPT was not the most powerful AI system ever built. It was not even the most capable model OpenAI had. But it was something more important: it was the first AI system that ordinary people could use, understand, and be amazed by. ChatGPT was AI's iPhone moment — the product that took a technology from the realm of specialists and put it in everyone's hands.
Why ChatGPT Was Different
Large language models had existed for years before ChatGPT. GPT-3 had been available through an API since 2020. Google had internal language models that were arguably more capable. Researchers and developers had been using these systems to build applications for months.
But using GPT-3 required technical knowledge. You needed to understand API calls, prompt engineering, and the quirks of language model behavior. The general public had no way to interact with these systems directly.
ChatGPT changed that with a simple interface: a text box. Type a question, get an answer. Ask a follow-up, get a follow-up. The conversation felt natural — more like talking to a knowledgeable friend than interacting with software.
The underlying technology was GPT-3.5, a model that had been fine-tuned using a technique called Reinforcement Learning from Human Feedback (RLHF). Human trainers had conversations with the model, rated its responses, and the model was adjusted to produce responses that humans preferred. This process made the model more helpful, more conversational, and less likely to produce harmful or nonsensical output.
RLHF was the crucial ingredient that made ChatGPT feel different from previous AI systems. Raw language models are powerful but erratic — they might answer your question, or they might go off on a tangent, or they might produce something offensive. RLHF smoothed out these rough edges, producing a system that behaved more like a helpful assistant and less like an unpredictable text generator.
The World Reacts
The reaction to ChatGPT was immediate and overwhelming. Students used it to write essays. Programmers used it to debug code. Writers used it for brainstorming. Lawyers used it to draft contracts. Marketers used it to generate copy. Teachers panicked about cheating. Journalists wrote breathless articles about the end of human creativity.
Within weeks, every major technology company was scrambling to respond. Google, which had been cautious about deploying its language models publicly, declared a "code red" — reportedly the first time the company's founders had issued such an alert since the early days of Google Search. Microsoft, which had invested billions in OpenAI, rushed to integrate ChatGPT into its Bing search engine. The announcement that Microsoft would challenge Google's search dominance with AI sent shockwaves through Silicon Valley.
Google responded by announcing Bard (later renamed Gemini), its own conversational AI. The launch was rushed, and Bard's first public demonstration included a factual error — it incorrectly stated that the James Webb Space Telescope took the first pictures of exoplanets. Google's stock dropped $100 billion in a single day. The incident illustrated both the intense pressure companies felt to keep up and the risks of deploying AI systems that could confidently state falsehoods.
The Education Crisis
ChatGPT's impact on education was immediate and contentious. Students could now generate essays, solve problem sets, and write code by typing their assignments into a chat window. Teachers who had spent years developing writing assignments found their assessments suddenly compromised.
The initial reaction from many educational institutions was to ban ChatGPT. New York City public schools blocked access to the system on school networks. Universities added AI use to their academic integrity policies. Some professors returned to handwritten, in-class exams.
But a counter-narrative quickly emerged. Some educators argued that trying to ban AI was futile — like trying to ban calculators in the 1970s. The better approach, they suggested, was to integrate AI into education. Teach students how to use it effectively, how to evaluate its outputs critically, and how to do the thinking that AI could not do for them.
The debate revealed a deeper question: what is education for? If AI can write a competent essay on the causes of World War I, what is the value of asking students to write one? The answer, many educators concluded, was that the value was never in the essay itself. It was in the thinking required to write it. The challenge was designing assignments that developed thinking skills and could not be short-circuited by AI.
GPT-4 and the Capability Jump
In March 2023, OpenAI released GPT-4, and the capabilities of large language models took another dramatic leap.
GPT-4 could pass the bar exam with a score in the 90th percentile. It could score in the 88th percentile on the LSAT, the 80th percentile on the GRE quantitative section, and the 99th percentile on the Biology Olympiad. It could analyze images, not just text. It could write working software from detailed specifications. It could explain its reasoning in a way that was often clear and logical.
These results shocked even AI researchers. A system trained on next-word prediction — the same basic approach that had powered every GPT model — was now outperforming most humans on standardized tests designed to measure sophisticated reasoning.
GPT-4 was not perfect. It still hallucinated. It still struggled with certain types of mathematical reasoning. It could be tricked into producing harmful content with cleverly designed prompts. But its baseline capabilities were so far beyond what had existed a year earlier that the gap was hard to comprehend.
Anthropic released Claude 2 in the same period, demonstrating similarly impressive capabilities with a focus on being helpful, harmless, and honest. The competition was pushing all companies to improve rapidly.
The Alignment Problem Goes Mainstream
ChatGPT and GPT-4 brought a previously academic concern to mainstream attention: the alignment problem.
The alignment problem, in its simplest form, is this: how do you ensure that an AI system does what you actually want, rather than what you literally asked for? This sounds simple, but it is profoundly difficult.
A language model optimized to be helpful might provide dangerous information to anyone who asks nicely. A model optimized to avoid harm might refuse to answer legitimate questions. A model optimized to be truthful might be brutally honest in situations where diplomacy is called for. Balancing these objectives — helpfulness, harmlessness, honesty — is a challenge that no one has fully solved.
The concern extended beyond current systems to hypothetical future ones. If AI systems continued to become more capable, at what point might they become difficult to control? If a system is smarter than the humans overseeing it, how do you ensure it follows human values? These questions, which had been the province of philosophers and science fiction writers, were suddenly being discussed in congressional hearings and corporate board rooms.
In March 2023, a group of prominent AI researchers and technology leaders signed an open letter calling for a six-month pause in the training of AI systems more powerful than GPT-4. The letter cited "profound risks to society and humanity." The pause never happened — competitive pressures were too strong — but the letter signaled that even insiders were concerned about the pace of progress.
The Economic Disruption
ChatGPT triggered immediate economic anxiety. If AI could write, code, analyze, and create, what would happen to the people whose jobs involved those skills?
The impact on the knowledge economy was tangible and rapid. Companies began using AI to draft marketing copy, customer service responses, legal documents, and software code. Freelance writers, translators, and illustrators reported declining demand for their services. Coding bootcamps began teaching "AI-augmented development." Law firms explored using AI for document review and contract analysis.
But the picture was more nuanced than the headlines suggested. AI did not replace most workers outright. Instead, it augmented them — making individual workers more productive but potentially reducing the number of workers needed. A marketing team that had employed five copywriters might now need three, with each using AI to produce more content. A software team might ship features faster, not because AI wrote the code, but because AI helped developers debug, document, and test more efficiently.
The economists' consensus — tentative and contested — was that AI would follow the pattern of previous automation waves: disrupting specific tasks more than entire jobs, creating new kinds of work while eliminating old kinds, and ultimately increasing overall productivity. But the speed of the transition was unprecedented, and the workers most affected were knowledge workers who had previously felt insulated from automation.
The Year Everything Changed
Looking back, the period from November 2022 to the end of 2023 was arguably the most transformative year in the history of artificial intelligence. In twelve months, AI went from a technology known mainly to researchers and developers to a mainstream cultural phenomenon discussed at dinner tables, debated in legislatures, and used by hundreds of millions of people.
The technology itself continued to advance at a pace that surprised even its creators. Models became more capable, more reliable, and more useful with each iteration. But the most significant change was cultural: for the first time, the general public had direct experience with AI systems that seemed genuinely intelligent. Whether they actually were intelligent — and what "intelligent" even meant in this context — became the subject of the most widespread public debate about technology since the invention of the internet.
The next chapter would bring something even more transformative: AI systems that did not just answer questions and generate text, but that could take actions in the world — browsing the web, writing and executing code, using tools, and pursuing complex goals with minimal human supervision. The age of AI agents was about to begin.