Notes from AI Ethics by Mark Coeckelbergh
Great book! Very understandable for me, a computer illiterate. It also challenged my anti-AI bias in a couple of ways. It provides a quick history of AI and contemporary uses of it. Coeckelbergh goes from the symbolic AI and expert systems of the 80s to the connectionist neural networks of now. I had vibed out what AI, software, and algorithms were before, but this cleared it up:
AI is technologically simulated intelligence. While we usually mean human intelligence, that doesn't necessarily have to be the bar. AI is software, and software is made up of algorithms. Algorithms are a set of instructions that transform input data into output data.
Machine learning, what we think of as "AI" nowadays, is
"A machine or software that can automatically learn: not in the way humans learn, but based on a computational and statistical process. Feeding on data, learning algorithms can detect patterns or rules in the data and make predictions for future data."
Machine learning and the subfield deep learning often use neural networks. Neural networks are,
"...based on the idea that instead of representing higher cognitive functions, we need to build interconnected networks based on simple units. Proponents claim that this is similar to how the human brain works: cognition emerges from interactions between simple processing units, called “neurons” (which, however, are not like biological neurons). Many interconnected neurons are used."
A previously popular type of AI, symbolic AI, is often based on decision trees. Decision trees require that decisions or rules be made based on data ahead of time, and that those rules are coded into the system by programmers. With machine learning, the rules come from patterns recognized by the AI and are not coded ahead of time. They are unknown to us, a "black box."
AI is already used in many facets of our lives, including in ways we might not immediately expect. I knew about recommendation systems and video game NPCs, but I didn't know that autopilot and self-driving cars used AI too. There's also the different subfields of AI like, "machine learning, computer vision, natural language processing, expert systems, evolutionary computation, and so on." These subfields interact with these facets in various ways that understandably aren't covered in full.
Coeckelbergh separates the actual issues with AI from the "Frankenstein complex" that imagines every AI to be a supergenius Terminator-in-waiting. In reality,
"There seems to be a real risk that in the near future the systems will not be smart enough and that we will insufficiently understand their ethical and societal implications and nevertheless use them widely."
There's also the ethical points of how much moral agency AI has, whether we should be the ones to prescribe it moral status at all, how good AI is at simulating human connection, how low the bar is for using AI maliciously, the inscrutability of AI to the public, explainability, and energy consumption.
I learned about Bostrom's paperclip maximizer and Searle's Chinese room thought experiments from this book too! More to read for later :]
