A Very Brief History of Chatbots
"It looks like you're writing a letter. Would you like help?"
You're sitting in front of your bulky monitor. You’ve dialed up to connect to the Internet. You're trying to navigate the labyrinth of Microsoft Word, wrestling with stubborn formatting errors as your deadline looms. Out of nowhere, Clippy springs into action, eager to assist you. Its timing was usually off, its help often unneeded, and yet when you actually needed assistance, it was conspicuously absent.
This notoriously bad chatbot greeted millions of Windows users with more interruptions than assistance throughout the late '90s and early 2000s. The small, animated paperclip with googly eyes and an endearing enthusiasm for administrative task was ineffective… at best.
Despite its shortcomings, Clippy was a harbinger of what was to come—a primitive example of a future where machines could understand and interact with us. Clippy was a promise of what was just over the horizon, a technology whose time had not yet come.
I want to talk about what happened before and after Clippy today. It’s a really interesting story, and it might tell us something about where we’re headed next.
The Pre-Clippy Era
Before there was Clippy, before computers found a place in nearly every home and office, a ground-breaking experiment in conversational agents set the stage for our modern understanding of chatbots. The year was 1966, and the project was ELIZA.
ELIZA, developed at MIT by Joseph Weizenbaum, was a primitive form of a chatbot designed to simulate human conversation. While its capabilities were elementary by today's standards, ELIZA demonstrated a fascinating concept: a machine could interact with human users in a conversational manner, using pre-programmed scripts to mimic responsive dialogue.
ELIZA's most famous script, DOCTOR, emulated a psychotherapist's non-directional responses to patient input. Despite its simple design, it often fooled users into thinking they were interacting with a real human, thus challenging the prevalent notion of what machines could do.
Some people started getting really excited. Could ELIZA pass a Turing test? Maybe this was it!
Wishful thinking. ELIZA’s capacity to understand context was almost non-existent, and its responses were heavily reliant on pre-set scripts. It could not learn from past interactions or understand nuanced or complex inputs.
Nevertheless, as you can see from the above video, ELIZA instilled a feeling that it was humanlike more than 50 years ago.
The Rise of More Advanced Chatbots
Technological progress tends to move very slowly at first, but as the previous discovery is often used to make additional discoveries, progress tends to pick up speed over time. Naturally, the progress of chatbots started out standing still, so it took a couple of decades before a much more sophisticated approach to machine-human interaction made its debut. Enter ALICE, or Artificial Linguistic Internet Computer Entity.
Developed by Richard Wallace in the mid-1990s, ALICE was a crucial leap forward for chatbot technology. Wallace leveraged a novel approach called AIML, or Artificial Intelligence Markup Language, which allowed ALICE to carry on more realistic and dynamic conversations. Although still heavily scripted, ALICE's ability to respond to a variety of inputs marked a significant departure from ELIZA's relatively rigid scripts.
Incidentally, ALICE is the inspiration for the film Her.
Then, in the early 2000s, came SmarterChild. Launched on AOL Instant Messenger and MSN Messenger, SmarterChild represented a shift toward utility-focused bots. It could deliver the weather forecast, provide stock quotes, and even play games, demonstrating that chatbots could do more than just mimic conversation—they could provide practical information and services, too.
However, despite these advancements, chatbots like ALICE and SmarterChild were still fundamentally limited. They could not understand complex language, they couldn't recognize context, and they certainly couldn't learn from their interactions. They followed predefined scripts and rules, and if a user strayed from these scripts, the bot often stumbled.
Nonetheless, these were significant steps on the long road to creating genuinely conversational bots. The stage was set for a new era of AI-powered chatbots, although it would take another decade or so for that vision to start becoming a reality.
As we crossed the threshold into the 2010s, advancements in machine learning and natural language processing set the stage for a new wave of chatbots. Now, instead of following rigid scripts, these chatbots could learn from user interactions, making them far more adaptable and dynamic.
IBM's Watson, which famously won Jeopardy! in 2011, showcased the potential of these new AI-powered systems. Watson was not a chatbot in the traditional sense. However, it demonstrated the capacity of AI to understand and generate human language on a level never seen before.
Ken Jennings’s final response in that famous last game granted him the correct answer, but he added the commentary:
I for one welcome our new computer overlords.
This era also saw the rise of virtual assistants like Apple's Siri (2011), Google's Assistant (2012), and Amazon's Alexa (2014). These conversational AI brought the technology directly into the hands of everyday consumers. They could respond to a variety of commands, assist with scheduling and setting reminders, and even crack a joke if asked. While these assistants each had their quirks and limitations, their release marked a significant step forward for conversational AI.
Notably, these virtual assistants were products of some of the biggest tech companies in the world, signaling the beginning of a new race in the realm of AI. These new breeds of AI-powered chatbots were starting to understand the context and semantics of human language, taking us into a whole new realm of technological possibilities. This wasn't just about winning at Jeopardy or chess anymore; it was about bringing sophisticated AI into our daily lives.
And, of course, it was about making money.
Some time in November of last year, I was introduced to ChatGPT. I saw that you could tap into a vast and powerful system that could help you sort and organize data. I also saw a truly conversational chatbot for the first time ever, capable of remembering context and responding accordingly.
Sure, GPT-3 was limited in capacity. It was easy to “trick” or “break”, and it “hallucinated” (made stuff up), kind of like a human being. But I saw clearly, immediately, that this was revolutionary.
Up until November of 2022, I understood that, through the power of a carefully crafted search (usually Google), you could figure out almost anything by tapping into all collective human knowledge. You could research the same way I thumbed through the World Book Encyclopedia as a kid. But not only could you find the sort of stuff that you would find in an encyclopedia; the Internet let you find out almost anything else about people, places, or things in real time.
In November of 2022, I saw that you could now do this instantly. GPT-3 had already looked at everything on the Internet (an exaggeration, but closer to reality than you might at first think). Not only this, but it understood context. It understood how to look for specific answers to nuanced questions, at least to a surprisingly good degree, completely blowing away everything that came before it.
Then came GPT-4.
The upgrade in efficacy and power was tenfold, easily.
Now, GPT-4 isn’t indistinguishable from a real person, and it still makes mistakes… but in many ways, it is vastly more useful than a human assistant. When connected to the Internet (as Open AI’s paid plan allows, and which Bing does for free to a limited degree), GPT-4 can look stuff up in real time and give you answers about specific web pages.
It’s not amazing at this yet, but it can even look at multiple websites to give you an answer. It’s like having an executive assistant who’s around 24/7, never off the clock, and who moves at the speed of light.
They're still fundamentally pattern-matching machines, not conscious entities, but it does beg the question: what’s the difference? I’m not convinced anyone can answer this question today.
GPT 5, 6, and 7
As we look towards the future, it's clear that the chatbot journey has just begun. We've moved from rudimentary scripts to sophisticated AI-powered systems that can understand context, provide valuable services, and interact in a humanlike manner. But there's still a long way to go.
Today's chatbots, even advanced ones like GPT-4, are still limited in many ways.
However, rapid advancements in AI and machine learning suggest that these barriers might not be insurmountable. Efforts are underway to improve the conversational abilities of these systems, their understanding of context, and their ability to learn and adapt over time.
Further, chatbots are now writing code themselves. They can often solve problems with code that humans can’t. If we reach a point where a “machine” can improve its own code better than a human can, that moment is often called “the singularity”, a moment in which we can no longer predict what will happen. We’d no longer hold the title of “world’s smartest.”
Meanwhile, chatbots are set to become more integrated into our everyday lives, assisting us in tasks ranging from scheduling appointments, to helping us shop, to guiding us through complex data analyses. For many people out there, chatbots are already becoming our personal assistants, our tutors, our therapists, or even our companions.
Alongside the possibilities, there is deep danger for humanity. Sure, the robot overlord thing could happen, but in the immediate and medium future, we have considerably disruptive threats to society.
The idea of AI companionship probably gives you the willies; companionship isn’t the sort of annoying task we’re looking to outsource. If you’ve seen 2001: A Space Odyssey, this awareness has been reinforced into your psyche. Even if you haven’t, if you’ve watched any movies or TV where AI is mentioned, the dystopian side of AI “friendship” is likely well ingrained.
Just as with any technological innovation, chatbots come with a blend of promise and peril.
But if there's one thing history has shown us, it's that technology doesn't stop advancing. So whether we're ready or not, the chatbots of the future are coming.
Knowing a little of the early history—from ELIZA to Clippy—can help you see the continuum we’re on. Iterative improvements began slowly (30 years from ELIZA to ALICE!), but then picked up steam gradually, as each previous version helped to create the next version.
Now, generative AI helps us code better LLMs, which help us code better LLMs…
It's an exciting time to be alive.