AI encompasses a wide range of tasks and techniques. However, from an industrial perspective, it’s essential to remember that the final application and its ROI are what truly matter. From this viewpoint, the final stage is the most important: decision-making. It’s at the moment when we apply all the conclusions drawn from the data (via AI) to reality that true gains can be achieved. This step is often neglected in AI communications for two main reasons:
- Because it is at the end of the chain, and the digital revolution in organizations has often not yet reached the maturity necessary to even ask the question, and players more easily project themselves into the “understanding” AI (Machine Learning, Deep Learning) as it continues the trend of what was called “Business Intelligence” a few years ago.
- Because it sometimes quite brutally encroaches on human territory. Indeed, if AI is confined to providing increasingly advanced dashboards with more precise predictive indicators, humans remain in control. But when it starts to suggest decisions, that’s when problems can begin…
While the first point is being addressed over time, the second is a real issue that deserves reflection, primarily from an ethical perspective but also from a technical one. What are we trying to achieve when we want a machine to make a decision instead of a human? Is it always desirable? Is it necessarily “better”?
Human vs. Machine: Fundamental Differences
When developing such systems, it’s natural to question the validity of the approach. From an application perspective, one can easily consider that AI is particularly relevant where humans are not, and vice versa. To illustrate this, we can attempt to present the “types” of problems that require decision-making based on their ease for a human and for a machine. This classification, loosely inspired by complexity theory, is “unpretentious,” and there could be much debate about it. The idea here is to offer an interesting angle for reflection on the subject.
When considering all the problems humans can think about to make a decision (or just to find a solution), we find that it’s usually computational complexity that limits us. For example, choosing the cheapest mortgage is straightforward, and with the same information, a machine will give the same result. But as the number of alternatives increases, requiring more and more calculations on large data sets, it becomes harder to retain everything and have an overview for decision-making. No financial director, however talented, can compute the financial balance of a global company in their head to validate or not an investment decision. This task is impossible for a human without machine assistance, which involves simply performing sums, averages, etc. These are simple calculations but in large quantities.
Humans are skilled at abstract thinking, using concepts and “patterns” intuitively in daily life. Conversely, it’s effortless for us to learn to play a game like poker, chess, or any video game and make “intelligent” choices from the first game (not to be a grandmaster, just at the average level of any human). Or to learn to navigate (walking, driving, etc.) while paying attention to our environment. These are particularly complex tasks for a machine, which has no ability to intuitively understand the right strategy to adopt, even for games with simple rules like Go. In extreme cases, it’s completely natural for us to question the right course of action in an unknown situation, considering a multitude of aspects: ethics, danger, short-term/long-term benefits, the pleasure we get versus the harm caused to others, etc. We are almost effortlessly capable of discerning power dynamics, friendships, hatred, domination in a social group, and adapting our behavior accordingly. These tasks are so inaccessible to a machine that we don’t know if they will ever be capable of them. Here, we touch on what is closest to our “consciousness,” which we cannot define and whose possession by a machine remains unknown (see the debate on strong AI, which is beyond the scope of this article).
Problems Difficult for Both Machine and Human: The Challenge of Working Together
Recent AI applications communicated to the public, sometimes in the unnecessarily spectacular angle of “the machine surpassing its creator,” often fall into the category of problems easy for humans and difficult for machines: image recognition, autonomous driving, text comprehension. These applications are impressive, sometimes troubling, as they encroach on what was thought to be uniquely human. In these cases, the AI project generally aims for scalability: automating a simple task for humans on a large scale. We can debate the necessity of these applications (do we need autonomous cars?), but the fact remains that it doesn’t take away from what makes human life interesting and that only we can practice. The debate can become tense when addressing applications that are complex for both humans and machines.
These are the hidden, little-known problems that are always difficult: complex, combinatorial problems where the number of possibilities is such that even a machine cannot completely compensate with its computing power. These problems require considering so many aspects that even with an overview and conceptual thinking, it’s difficult to make the right decision. These are problems that require extensive training for a human (medical or mechanical diagnosis, mathematical proof, quality control of an automotive part, optimal planning of a single delivery route) and are also difficult for a machine to perform. When these problems become too large (e.g., managing a global parcel transport network), it is no longer possible to operate them without intense division of labor. On these problems, two strong components intervene:
- Complexity, combinatorial aspects: what is particularly difficult for humans.
- Strategy, relational aspects: what is particularly difficult for machines.
Organizing and optimizing parcel distribution in a delivery center requires considering an almost infinite number of possibilities that change daily, with a significant impact on performance and service quality. But it also involves considering environmental impact, driver and client sensitivities, thinking about the future, and risk management. Machines are far better than humans at the first point but inept at the second. Therefore, on these applications, it’s essential to deeply consider the role machines can and should play within organizations entirely reliant on human sensitivity.
What place for Humans in AI solutions for industry? Discover our article on the subject!