← All posts

Human vs. AI: which has better input data to work with?

I recently wrote about the “input data” humans use for learning as well as the data AI learns from. This takes the lessons from these two discussions and shows the implications this has on artificial intelligence—its capabilities, limitations, and how to think about the intelligence in artificial intelligence.

We’ll see that generally human minds start with much more advanced and well-suited data for learning. On the other hand, computers either need their input data to be carefully curated by humans or for the information being learned to be fully contained in a massive, homogeneous dataset far different than how humans learn.

Artificial intelligence is capable of amazing things. When learning about the world, or how to operate in it, humans are much smarter. Further, humans also start out with a much richer set of data to learn from.

AI systems are not intelligent in the way we think of human intelligence, but they can accomplish impressive feats in the right circumstances.

Recap of the “input data” humans learn from

Humans are conscious, and learning is an active process driven by consciousness. From sense organs to the brain and subconscious, data about the external world is collected and processed into a very advanced form for consciousness: an awareness of objects along with their actions, attributes, and relations to one another.

Perception is not of individual stimuli like a photon of light hitting the retina. Humans perceive objects. We don’t think about all the many little bits of sensation and group them together into objects. We just perceive the objects directly.

By the time these perceptions make their way to consciousness, they’re already at a pretty high level of abstraction. And as we learn facts and skills, we automatize the process and push that work down to our subconscious, freeing up precious attention for still higher-level learnings.

Recap of the input data AI systems learn from

In many cases, AI systems start with carefully-curated data, with humans identifying the most relevant factors and the algorithms figuring out how to combine them together into an optimal decision.

In limited cases, such as recognizing objects from images, where the knowledge to be learned is entirely contained in the data without needing to identify relationships to outside data, computers can learn from very low-level data far more detailed and less processed than what humans use. There is an additional caveat: it requires a massive amount of data for each of these systems.

Human perception vs. robotic sensors

While most AI applications are narrow, focused systems doing things like spam filtering or movie recommendations, when most people think of artificial intelligence they think of what they see in movies: robots that understand the world and know how to act within it.

In this area, the comparison is clear: the human mind works with much more advanced data than what a computer can get from an image sensor. We perceive objects, not disconnected pixels of color. Our brain and subconscious fuse external stimuli together into coherent entities which we’re aware of as a whole. Since learning happens in our conscious, this is merely an input to that process.

Computers, on the other hand, have much more work to do. To answer something simple like whether there’s a cat in an image, they don’t start with objects like we do. They start with images composed of tiny pixels, with each having a different color.

Artificial intelligence algorithms have to figure out which pixels go together in one object, and then figure out what that object is. It’s a much harder problem than any thinking we’re consciously aware of. As humans much of this is taken care of for us without needing to explicitly learn it. Our learning happens on a stronger base.

It’s impressive that AI can work with such primitive data. When it works, it’s not just reproducing decisions we consciously make, but also all the processing that goes on in our brain and subconscious to make us aware of the things in the world around us.

Most AI works with smaller, human-curated data

While working with images and audio has recently become possible, most AI is fed by carefully-selected datasets where a human has chosen the most relevant data for the computer, and it’s always at the right layer of abstraction.

This requires less data, but more human effort.

When we’re willing to put in the effort, it works great. Its nearest analog for humans is learning something in school, where an instructor carefully chooses examples and arguments to help you reach a new conclusion.

With AI, it often takes weeks or months to learn one new thing, and the idea is to apply that thing over and over, millions or billions of times.

In school we learn many things in a single class, and much more in a week or month. Nonetheless, it makes sense to spend a seemingly ridiculous amount of time teaching a computer how to do one simple thing.

Since it can repeat the process quickly and essentially infinitely once learned, it actually saves time over having a human do it every time. More than that, it makes products like online search engines possible.

It wouldn’t make any economic sense to have a human answering every search engine entered into Google. Even if it takes inordinate effort to learn simple models, it enables products we wouldn’t have otherwise.

So, computers require much more complex and expensive data, but the programs are so valuable that it’s still worth it.

AI advantage: scale of data, speed

I’ve been talking about the quality of the data used for learning, for which humans are unquestionably better. That said, computers also have a couple major advantages related to the input data, even if quality isn’t one of them.

First, computers can handle way more data than humans. Some AI systems are trained on many billions of records of data. Computers can go through massive numbers of records much quicker than a human could, and learn a little bit from each one.

Humans don’t need billions of examples before they learn something. That’s good for us! But, since computers have to work with such inferior data, being able to use much more data can help with the added difficulty.

Human advantage: integration and automatization

Humans have an integrated understanding of reality, and all our knowledge is connected. We can just glance at an email message and see if it’s spam or not. We can recall whether we recognize the sender, if we signed up for anything like this, whether this looks like other spam we’ve seen in the past, etc.

Computers, on the other hand, only have the limited set of data we provide them. What they have may be a robust set of data, but it’s limited to what humans intentionally collect and input to the system. And nobody has figured out how to effectively represent knowledge in a computer or make distant connections like humans easily do.

A further advantage for humans is the ability to automatize what we’ve learned. When you were learning to count or to ride a bike, it was hard and required concentration. Once you’ve learned it, it becomes second-nature and you can build on these skills. These (and many others) enable us to learn calculus or how to drive a car.

Computers generally are taught one specific task, with a lot of the framing for the problem encoded as assumptions in the model that’s built. But they don’t learn from previous learnings.

Recently, there have been some advances in the direction of enabling AI to learn from its previous steps. For example, a model of human language can be used on a range of tasks related to processing text. This, however, is a slightly more general version of the same thing.

Conclusion

Learning seems relatively easy to us as humans in part because we start with rich, structured data that’s much more advanced than anything available to computers without humans spoon-feeding it to them.

When computers learn, it’s a much harder problem. And computers don’t have tools anywhere nearly as advanced as the human mind.

Computers are also able to work with data in some ways that humans can’t, particularly in working with massive datasets. This enables some products that are enormously valuable. They are valuable not because the computer is smarter than the human programming it, but because they take something easy for a human and find a way to do it at scale without wasting a human’s time.

Humans are far more advanced than computers, and it’s a difference of kind rather than degree. But we should all be thankful for the value brought by computers enabling us humans to spend time on the things we care about.