Designing Autonomous AI

Chapter 1. Sometimes Machines Make Bad Decisions

Why do humans so frequently manage, supervise, step in to assist, or override automated systems? What are the qualities of human decision-making that automated systems cannot replicate? Comparing these strengths and weaknesses helps us design useful AI and avoid thinking that AI is a magical cure-all or complete hype.

I took a trip to Australia and while I was there, a resources company asked for help with their aluminum production process. The company uses an expert system to make automated decisions during the process.

Aluminum is made in large tubs called cells (see Figure 1-1). Aluminum is smelted using a process called electrolysis. In aluminum electrolysis, alumina powder is injected into a cell full of cryolite. Electricity passes through the cryolite, and the resulting reaction produces aluminum. Here’s the problem: it’s almost impossible to tell exactly what’s going on in the cell because the temperature is high enough and the cryolite is corrosive enough to destroy sensors. So experts and automated systems rely on changes in the electrical properties of the cell for feedback. Voltage is how hard the electricity is pushing through the system. Current (measured in amps) is the amount of electricity that flows through the system, and resistance is the amount of opposition the system provides to the flow of electricity.

So standard practice for controlling alumina cells (across the industry, not just at this particular company) ...