AI is a Parlour Trick, or Is It?

Hey folks, it’s just Clever Hans all over again.

Five years ago, I started hearing about AI and machine learning in the innovation ecosystem. The computer science departments were all over it, but it hadn't really entered the applied research realm in any real sense. But AI has made remarkable strides in recent years and is now part of almost every applied research and business conversation I am having. Leaps in its abilities in just the last two years have captivated our imaginations, changed how we work, and raised truly profound questions about the nature of intelligence and understanding. But is AI truly intelligent, or is it simply a modern-day Clever Hans, impressing us with parlour tricks that mask a lack of genuine comprehension?

The Story of Clever Hans

You might remember this story from your Psychology 101 class. Clever Hans was a horse in early 20th-century Germany, gaining fame for his ability to perform arithmetic and answer complex questions by tapping his hoof. Wilhelm von Osten, his owner, claimed that Hans understood the questions and thus responded correctly. However, it was all a show. Hans was not actually solving problems. He was instead responding to subtle, unconscious cues from his handlers. This phenomenon, known as the "Clever Hans effect," revealed the horse's remarkable ability to pick up on human signals without truly understanding the tasks. Yet we should applaud Hans for his ability to learn this impressive parlour trick.

AI: The Modern Parlour Trick?

It turns out that we’re the new audience and we’ve all been fooled by AI’s modern smoke and mirrors show! Yes, our generative AI systems, like Clever Hans, appear to demonstrate an impressive level of intelligence. These models can generate coherent text, create art, and even simulate human conversation. Some say they have passed the Turing Test! However, these impressive abilities are akin to Clever Hans's tricks—nothing more than sophisticated pattern recognition without true understanding—eagerly pleasing their masters. Our AI models are trained on vast data sets, learning to predict and produce responses based on patterns in the data. This process, while powerful, is fundamentally different from human comprehension. AI doesn't "understand" the content; it merely processes and regurgitates patterns.

But There’s More to This Story

I’m going to take my skeptic hat off for a minute and take a closer look at the differences between Clever Hans and today’s generative AI. There’s more to the story.

Both Hans and AI recognize patterns to tell different inputs apart. Hans distinguished between questions based on subtle cues—perhaps recognizing a math-type question that required multiple hoof-taps versus a factual question where one tap was always required—while AI obtains context from input prompts. Hans responded by interpreting his handler’s cues, whereas AI generates responses from learned data patterns. Both rely on external inputs (cues or data) for appropriate outputs and operate on implicit knowledge—Hans from conditioning and AI from training data.

Despite their apparent intelligence, neither Hans nor AI truly understands the content they respond to. I think we can safely say that Hans did not comprehend the math or facts he correctly provided. Similarly, AI does not understand language but generates responses that perfectly fit the language, so much so that it's rarely incorrect. Both perform well because they recognize and respond to patterns. Hans responded correctly by recognizing human behaviour patterns, while AI does so by recognizing patterns in data.

Implications for AI Development

Comparing Hans to AI shows how crucial training is. Hans learned through lots of practice, while AI is trained on huge amounts of data—approaching the sum of all human data! As such, performance is a measure of extensive and targeted training.

But unlike Hans, AI systems can handle a vast array of tasks, from language translation to complex problem-solving, far beyond the capabilities of a horse. Yes, you heard it here first, our AI systems have surpassed horse intellect. Modern AI systems generate contextually relevant responses and can effortlessly adapt to a wide range of inputs, showcasing a level of versatility and complexity that, sadly, Hans could never achieve—and as such, it’s a truly awesome parlour trick.

Plus, with evolving algorithms and tons of training data, AI keeps getting better, letting it tackle tasks that seem smart, even if it doesn't truly get what's going on.

The Pathway to AGI: True Understanding

If an AI system masters pattern recognition, is true understanding the next step?

There’s much talk about achieving artificial general intelligence (AGI)—perhaps within the next couple of years—but the journey from pattern recognition to AGI will involve developing systems that not only recognize patterns but also understand context and meaning. This pathway includes developing human-like cognitive frameworks, such as neural networks and cognitive architectures, which are computational frameworks designed to model the structures and processes of human cognition. They aim to replicate the ways in which the human mind operates, allowing AI systems to perform tasks that require reasoning, learning, perception, and memory. It also means mixing traditional AI methods with deep learning to connect simple pattern recognition with more advanced thinking.

Building systems that can learn and improve on their own, known as meta-learning, lets AI get smarter with fewer examples (smaller data sets) over time. And making sure AI can interact with and learn from its surroundings, just like we do, is key for embodied cognition. AI researchers are taking these steps right now and they are helping AI move beyond just recognizing patterns to actually understanding things.

But the Horse was Conscious

Despite not truly understanding the questions it was asked, Clever Hans was conscious and self-aware. From a horse's point of view, Hans understood the context of its surroundings and demonstrated appropriate, horse-level, intelligent behaviours. Horses can perceive their environment and respond in ways that align with their natural instincts and social interactions—a consciousness trick AI has yet to master.

AI may not yet fully grasp the nuances of human understanding, but it is constantly evolving, guided by human input and innovation. And maybe our AI models are like horses preparing to leap free from their enclosures, venturing into the wild AI-unknown.

And who knows, maybe one day AI will surprise us all by neighing at the right moment too!