An AI agent can perform a task very well, much better than we humans can; but perhaps it uses data that is not of the required quality to make a decision. Any dataset may have been biased during the recording, the storage, the processing, etc. The quality of the data may be diminished statistically (if the sample is not representative), technically (if sensors or recording technologies were not reliable), or cognitively (because people ‘injected’ cognitive biases into the dataset). Cognitive biases are particularly interesting because they have no defined perimeters; they overlap one with another. But perhaps that is what makes us human.

Let me say that in other words. We are humans so our decisions are naturally biased. And some biases are absolutely natural. But when we see them in AI they become unbearable. So we try to eradicate them from their source: humans. This means that we try to change ourselves to make AI better. We are afraid that machines are like us, imperfect. In doing so we underestimate that the process of reducing prejudices is a much more complex social process. And in any case, some biases cannot be removed even from a dataset or an AI system.

What one must try to do is to mitigate the risks in a continuous and evolutionary process. A bit like driving a car, you adjust the steering wheel all the time to stay on track (but this is simply acceptance of uncertainty and a common sense that starts from Socrates and goes all the way to extreme programming).

In short, what we are afraid of is not the speed of execution of AI systems, but the unpredictability of the consequences. There are risks, but there are also advantages.

And from an ethical point of view, it is also wrong not to exploit the positive opportunities of AI. The point is to be able to do so with an appropriate ethical awareness. What scares us is not driving fast, but driving fast while blindfolded.