Hybrid Intelligence: 4 Types of Intellectual Labor

This is Part 1 of a series, taken from my predictions and realizations about AI as Chief Design Officer for IBM Watson way back in 2014. Nine years later, it's fun to see where I got things right or wrong.

In the physical realm, there are a few different types of power sources that civilization has tapped over the course of history: humans, animals, consumptive/chemical (e.g. burning wood/coal/oil), and renewable/physics (waterwheels, windmills, solar, geothermal, etc). They all have their strengths and weaknesses, but collectively we’ve been able to do much with them. There may be a better taxonomy out there, but my focus here is actually on the realm of intellectual labor, which I divide into Experts, Crowds, Algorithms, and Machine Learning.

An Expert is someone with extensive experience who knows the rules of a domain or profession. To borrow from an old joke, they charge $1 for the hammer strike and $99 for knowing where to strike. One single expert may provide all the expertise and knowledge needed for even a complex task.

A Crowd could be an actual group of people providing assistance in real-time—perhaps via API calls—or it could simply be a digital exhaust trail from people’s past activities where the resulting data is mined for insights and direction. Recommendation engines and user reviews, as well as services like Amazon’s Mechanical Turk fall into this category. Things in this category are relatively low-skill but make up for it sometimes with volume because sometimes quantity creates quality.

Algorithms have been responsible for software “eating the world” and are essentially rules-based intelligence codified into computer systems. They are both the output of and a sort of digital peer of, Experts. Generally speaking, algorithms are usually deterministic and debuggable: outputs are predictably consistent if inputs are consistent, and if there is a problem, an expert can ascertain where the rules were incorrect and then directly correct issues.

Machine Learning—a subset of artificial intelligence—has been around for decades but has had occasional leaps forward. The definition of artificial intelligence is a moving target based more on how things are perceived than the underlying tech. There is a saying in the AI community that “it’s only called AI until it works.” For our purposes, I’m defining ML as systems that are trained instead of programmed. They are usually trained with large quantities of data, whereas humans are adept at training with “sparse” data and can often make inferences and extrapolations from pretty small datasets.

Hybrid Intelligence

We’ve already seen in the definitions themselves hints of the complex interdependencies and synergies between the four types of intellectual labor. This is key. With hybrid intelligence, the whole is greater than the sum of its parts. Not only does each type have strengths and weaknesses, they also tend to increase the efficacy of the other forms: e.g. Experts use data from Crowds and Algorithms to train ML models. We are talking too much about artificial intelligence by itself instead of as part of hybrid intelligence. Why does this matter?

Anthropomorphization & Wishful Thinking

First, by recognizing the whole of HI, we have a better framework for practical gains and less magical thinking about AI/ML as human-like—or worse—some sort of god-like intelligence. This is problematic not only because it’s inaccurate. Overpromising and under-delivering in the past caused private and government funding to dry up, leading to “AI Winters”. Similarly, this kind of talk also makes AI seem too powerful (and clearly alien and untrustworthy) causing a backlash like the AI community saw after the portrayal of HAL in the Kubrick movie, 2001.

Augmentation vs Competition

By not expanding the scope to HI and treating all intellectual labor as one thing, the narrative around AI becomes obsessed (both doubters and believers) with “AI replacing humans”. As we’ve hopefully established above, that is not a very sophisticated or accurate way of looking at things. ML (and algorithms) are better at replacing certain tasks as opposed to entire jobs and progress will be better served by remaining cognizant of that.

The Right Tool for the Job

The AI hype leads workers to fear AI and also leads business owners to think they can replace workers en masse with ML. It is a fact that human jobs have historically been eliminated or reduced when automated processes are cheaper than equivalent human labor. The part often missed is that these labor types are rarely equivalent. So we have the potential for chaos and backlash against AI when cost-cutters lay off workers that they will later need to rehire. Instead, we should include hybrid intelligence in the conversation and work on getting the right combinations and interactions of all four types of intellectual labor.