Three years ago today, a simple web form quietly opened on OpenAI’s website. It asked you to type a message into a box and press Enter. That was all. No big feature list. No pricing page. Just “ChatGPT” and a blank field.

Within days, millions of people had tried it. Sam Altman tweeted that it had crossed one million users in less than a week. What happened next is now part of AI history. On The AI Navigator timeline, November 30, 2022 sits next to milestones like AlexNet in 2012 and the first GPT models. It marks the moment when large language models stopped being a research curiosity and started to feel like a public utility.

What interests me today is not the launch itself, but the first reactions in the days that followed. The initial shock, the early metaphors people used, the things they worried about, and the things they completely missed.

Three years later, as we work with multimodal, agentic versions of ChatGPT that can browse, see, listen, and act, those first impressions say a lot about how humans respond to new intelligence.

In early December 2022, one of the most widely read tech commentators told a small family story. It was a Wednesday night. His daughter was preparing for a European history exercise called “The Trial of Napoleon” and needed to argue in the voice of Thomas Hobbes. He typed her homework question into this new tool called ChatGPT. The answer came back sounding polished, confident, and well sourced. It was also completely wrong about Hobbes.

That story crystallized two reactions that dominated the first week.

First, panic and fascination around school and homework. It did not take long for people to realize that a free, fluent AI that can write passable essays on demand might break the take-home assignment as we knew it. Teachers began asking how they would tell the difference between a student’s own writing and the output of this system. Parents and students tried it on everything from history essays to college application drafts. The tone was half amused, half anxious. A new phrase started to float around: “the end of homework as we know it.”

Second, a new mental model for computers. In that same essay, the writer drew a line between calculators and ChatGPT. A calculator is deterministic. If you enter 2+2, you always get 4. ChatGPT is probabilistic. It generates the most likely next words based on patterns in its training data. Sometimes those words are exactly right. Sometimes they sound right but are not. The conclusion in that first week was subtle and important: this is not a machine you “trust” like a calculator. It is something you collaborate with, then correct.

From that logic came one of the earliest and most enduring ideas about how to work with AI.

Instead of picturing AI as a black box that you interrogate until it coughs up the perfect answer, those early voices described what they called a “sandwich workflow.” A human starts with an idea and writes a prompt. The AI returns a few options. Then the human edits, combines, cuts, and rewrites. Prompt, generate, edit. Repeat. The emphasis moves from writing the very first draft to critiquing and refining options.

Looking back from 2025, that “sandwich” framing feels almost prophetic. It is the basic pattern that now shows up everywhere, from brainstorming copy to prototyping code, from planning trips to designing lesson plans. It also aligns perfectly with what many people on The AI Navigator have discovered through experience. The skill is not just in what you ask the model to do. It is in how you review, question, and shape what comes back.

Education was not the only early front line. Within a week, Stack Overflow had to put in place a temporary ban on answers generated by ChatGPT. The problem was not that the AI never wrote correct code. The problem was the combination of fluent style, non-trivial error rate, and extremely low cost of generating huge volumes of answers. It was suddenly easy to flood a community with plausible but wrong responses that required an expert to sort out.

This tension still exists today, only at a much larger scale. The AI timeline is now full of milestones that build on that first release: GPT-4, GPT-4o, GPT-5, GPT-5.1, multimodal models, and agents that can click around the web for you. Yet one of the central questions is still the same as it was three years ago. What do we want humans to spend their time on when most “first drafts” are cheap, instant, and often good enough?

In parallel with the homework and coding debates, a different group of writers took a more technical route. In late December 2022, one researcher used ChatGPT to analyze a problem in attribution, essentially asking whether this new system could reason about causes and effects the way humans do. The conclusion at the time was nuanced. ChatGPT could mimic the language of causal explanations and often gave the right verdict in simple games of chance. But its reasoning was inconsistent, and regenerating an answer could flip a correct explanation into a wrong one.

In other words, the earliest serious tests were already poking at issues that occupy researchers and practitioners today: reliability, reproducibility, and the boundary between pattern recognition and understanding.” On our historical timeline, these experiments sit close to the “research preview” label on the original ChatGPT announcement, and they highlight just how quickly people started treating the tool as something worth stress-testing, not just playing with.

One more theme from that first week deserves a place in the history books. The realization that the real innovation was not only the model itself, but the product wrapper around it.

By December 2022, OpenAI already offered an API that developers could call programmatically. If you were willing to read documentation, wire up billing, and build an interface, you could access the underlying models. That was powerful, but it reached a relatively small group of people. ChatGPT changed the equation by wrapping a tuned model in a familiar chat interface, making it free to try, and encouraging open-ended exploration. Write a poem. Fix my code. Summarize this article. Help with my resume.

One commentator described this as the moment when OpenAI shifted from being mostly an infrastructure provider to being a consumer product company, alongside image tools like Midjourney. Looking back from 2025, that observation aged well. Today, hundreds of millions of people use ChatGPT every week, and product choices like chat history, voice mode, app integrations, and built-in agents shape how the world experiences AI as much as the raw model weights do.

So if we stand on the timeline today and look back exactly three years, what did those first reactions get right, and where did they miss?

They were right that writing and homework would change. Schools and universities are now experimenting with “AI-first” assignments where using tools like ChatGPT is required, and the assessment focuses on critique, fact-checking, and synthesis. Early talk of banning ChatGPT has largely given way to grappling with how to teach students to operate in a world where AI is part of the default toolset.

They were right that human skills would shift from generating text to editing and verifying it. The editor’s mindset is now part of how professionals in law, software, marketing, design, and research work. You see it in everyday routines: generate three options, pick one, adjust tone, check facts, then ship. That pattern was visible within the first week, and it has only become more pronounced as models have grown stronger.

They were right that AI would put pressure on communities whose value depends on high quality answers. Stack Overflow’s early struggles pointed to a broader challenge. Whenever it is easier to generate plausible content than to verify it, curation and trust need new approaches. That question is still unsolved at web scale.

On the other hand, some things were underestimated.

Few early reactions fully anticipated the pace at which AI would become multimodal. Back then, ChatGPT was text in and text out. Today, the same family of models can see images, listen and speak, parse long PDFs, and operate browsers as agents that perform multi-step tasks. The conceptual jump from chatbot to universal interface for complex work was not yet obvious.

Most commentators in that first week were focused on text. Very few predicted how quickly generative AI would permeate audio, video, and interactive tools, or how soon entire workflows, from shopping research to financial analysis, would be wrapped inside conversational experiences.

They also did not fully see how social the change would be. The early homework stories were about individual students trying to get an edge. Three years later, we see entire teams building shared prompt libraries, developing internal house style for AI-assisted work, and even creating new roles around AI operations and governance.

So what do we do with all of this, here and now?

One purpose of a history timeline is to make these shifts visible. November 30, 2022 is not just the day ChatGPT launched. It is the week people realized that language models had crossed a usability threshold. It is the week educators began to argue about AI in the classroom. The week coders saw Stack Overflow start to wobble. The week technical writers tested the limits of probabilistic answers. The week everyday users saw, maybe for the first time, that a computer could talk back in a way that felt uncannily human.

Another purpose is more personal. History is not just what the big voices wrote. It is what you did.

So I have two invitations for you.

First, take a moment to think about how your own relationship with AI has changed since that first week.

Are you still in experiment and play mode, or has ChatGPT become part of your daily workflow? Do you use it mostly for writing, or also for coding, strategy, research, planning, or creative work? Where has your skepticism increased, and where has your trust grown?

Second, remember your own first ChatGPT memory. Maybe it was the first essay it wrote for you, the first bug it helped you track down, the first time it completely hallucinated something and you had to catch it. Maybe it was years later, when a voice-based version spoke back through your phone. Whatever that moment was, I’d love to hear it.

Three years from now, when we look back again, those individual stories will be just as important as the big milestones on the AI Navigator timeline. They are the human part of this history, and they remind us that every date on the timeline is really a collection of first encounters between people and a new kind of tool.