The evolution of human-AI collaboration

One of the most common questions I hear from people just getting started with AI is how humans and AI will coexist. I often tell them to look to history for answers. While human-AI collaboration might seem like a recent innovation, it’s been evolving for a while now—even before HAL and Data were helping humans explore new worlds.

The main types of AI systems

Let’s start by getting familiar with the kinds of AI systems out there and the way humans have used them over the years.

There are three main types:

• Knowledge-based systems produce conclusions based on expert, machine-followable rules.

• Learning systems improve performance by learning from data and feedback.

• Generative systems create new content based on patterns in training data.

Knowledge-based systems

MYCIN was an early example of a knowledge-based system. Developed in the 1970s, MYCIN could diagnose bacterial infections and recommend treatments based on “rules” established by medical experts. Although it wasn’t used in practice, MYCIN made expert knowledge accessible and provided reliable outputs. 

Knowledge-based systems are great for automating decision-making. That’s why they’re still used today for diagnostic support in healthcare, customer support chatbots, and legal expert systems. 

But knowledge-based systems need very structured data inputs, which means they’re not very flexible, and they can’t adapt to new situations without human help. Human oversight is usually needed to make sure unanticipated cases don’t lead to mistakes.

Learning systems

In 1959, Arthur Samuel developed one of the first learning systems: a checkers-playing program that improved at the game by playing with itself. 

Today, learning systems are used in recommendation engines, fraud detection, and personalized marketing. 

They’re more adaptable than knowledge-based systems, but a learning system’s outputs depend on its training data. If the data is biased, the output will be too. And if a change occurs in the data’s profile, the system might lose effectiveness over time. Integrating learning systems involves monitoring all decisions the system makes to ensure outcomes are both accurate and fair.

Generative systems

The first generative systems were chatbots. ELIZA, created in 1961, was an early chatbot that answered people with engaging, human-sounding responses.

As you can imagine, the sophistication of ELIZA’s outputs now pales in comparison with newer generative AI. OpenAI’s GPT-4o, Anthropic’s Claude Sonnet 3.5, and Google’s Gemini 1.5 can create sophisticated text, images, music, and code—and many other outputs. These AIs are super versatile and can handle all sorts of tasks. 

The tricky part is making sure what they create is high-quality and matches your request. As a result, the outputs can need a lot of human oversight—so much so that the interaction might be better thought of as co-creation.

The paradigms for human-AI collaboration are evolving

As these AI systems have evolved, so have the ways we work with them. 

Agent systems

Take the development of agent-based systems, for example. Early AI systems were mostly autonomous, meaning they operated without constant human oversight.

That paved the way for agent systems: groups of autonomous entities that work together to solve a common problem. Think of a smart home system, where separate agents (say, a thermostat, air conditioner, and humidifier) work together to keep the temperature in your home comfortable.

What’s exciting about agent-based systems is they’re modular, flexible, and scalable. You can add new agents or modify existing ones without disrupting the whole system. That’s why they’re being used in robotics, supply chain management, and traffic control. And they’re becoming more relevant with GPT bots and interconnected systems performing various tasks together.

HITL systems

Whereas agent-based systems are largely set-it-and-forget-it, other evolving systems need more human involvement—like human-in-the-loop (HITL) systems. In this system, AI makes an initial decision or recommendation, which humans then review and approve. We get the AI’s speed and efficiency and human expertise and accountability. HITL systems have become common in medical diagnosis, where AI flags potential issues for healthcare professionals to examine further.

Co-creation with AI

Another way our collaboration with AI has evolved? We’ve been co-creating with it more. You see this all the time in generative AI systems, where people let AI help them in creative processes. AI offers suggestions and generates ideas, which humans can refine and build on. (I like to call these tools “first-draft machines.”) Co-creation is gaining ground in content creation, design, and problem-solving, where the mix of human skill and AI capabilities leads to better results.

AI as a personal assistant

AI isn’t just about getting tasks done, though. It can also help you personally. In the past, if my friend and I used an AI system, both of us would’ve had the same experience. Today, AI can adapt to our individual preferences—which has opened the floodgates for human-AI collaboration.

For starters, AI can:

• Share specific, tailored feedback about your writing or code

• Analyze your speech and behavior during meetings

• Be a debate partner when you’re working out a robust strategy

• Organize your messy notes or diary entries into tidy, accessible formats

And now that tools like ChatGPT can remember past chats, they can learn your preferences and priorities, personalizing all that feedback even further. 

What comes next?

The history of human-AI collaboration has a long story to tell, and emerging trends suggest where we might go from here.

Multimodal systems that can process and integrate information from various sources—like text, images, and audio—are becoming table-stakes capabilities for many of the frontier models. These systems will enable us to do even more different kinds of tasks with AI, like asking a question about an image, or analyzing audio and video. We’ll also see more multi-agent systems, where multiple AIs work together and leverage these capabilities.

In my opinion, one of the most exciting developments is the expansion of AI’s context and memory. An increasing context window means we can tackle even larger, more ambitious projects with AI. Future AI tools will be able to understand and assist with more and more complex tasks by remembering more about us and our past interactions. With an enhanced memory, AI will be more personalized and effective, better supporting our long-term goals.

As for where you fit into all this? Different AI systems are better suited for different things, so consider the types of systems out there and the ways we’ve used them over the years. That should help guide you toward the right AI for you.

And keep an eye on the future. Today’s AI systems are more interactive, adaptive, and personalized than ever before, and they’re enhancing our human capabilities in ways we could only dream of a few decades ago. Looking ahead, I can’t help but get excited about the potential for even more meaningful collaborations between humans and AI. I can’t wait to see what we’ll achieve together.

Related reading: