The famous Three Laws of Asimov

If you have ever had any contact with robotics beyond setting up your “smart” vacuum cleaner, or if you are just a geek, you have probably heard of the so-called “Three Laws of Robotics’ or “Asimov’s Three Laws”, right? If not, let me quickly explain what they are.

In 1942, in a science fiction short story titled Runaround, the Russian-American writer and biochemist Isaac Asimov introduced  three principles designed to govern the  robots existence and behavior in the world, ensuring that androids would not negatively impact people or society. These “laws” would be internally encoded into their “positronic brains”. Eventually, this story was included in the book “I, Robot”, alongside other stories exploring the same theme. Below are the three Laws of Robotics, in Portuguese, translated by Aline Storto Pereira:

The first: a robot may not injure a human being or, through inaction, allow a human being to come to harm.

[…] The second […]: a robot must obey orders given to it by human beings, except where such orders would conflict with the First Law.

[…] the third: a robot must protect its own existence, as long as such protection does not conflict with the First or Second Laws. (ASIMOV, 2014)[1]

Looking at these laws very quickly and superficially, without thinking too much about the technical or ethical details, it seems that they are well-structured and make total sense, right? A robot (which, for the purposes of this text, we will also consider artificial intelligence, broadly) would protect itself, without disobeying human beings and, most importantly, without harming anyone .

This apparent harmony and false logic of the laws seem to have captivated some people who, to their great surprise, have used them as a basis for argumentation to discuss  regulation, impact, and application of artificial intelligence systems, both here in Brazil and abroad. However, there are several problems with this, and the objective here is to list some of them. Four, to be exact.

There is also a ‘zeroth law’ that is essentially the first, but involves humanity instead of the human beings. We do not need to discuss it, as these three alone are already enough to discuss, and the problems involved in using them in AI debates are the same as when considering the zeroth law.

3 Laws, 4 Problems

The Problem of Fictionality

The first and clearest problem that arises when taking these laws as a basis for real discussions of AI in society today is the fact that they are fiction. When Asimov created them as a writer, they served as a narrative device, an element that drives the story and creates conflicts for the characters – not as a theoretical foundation for future technology.

In addition, if you are reading this text, forgive me for the spoiler, but not even in the short stories of ‘I, Robot do these principles work! Basically, the stories in the book revolve around ambiguities, conflicts and uncertainties that these laws have between themselves and how the characters use their cunning to take advantage of this and solve the challenge that arises.

But, okay, some technologies first appear in the human imagination and in fiction before coming to reality (just ask Jules Verne), so perhaps this would happen with Asimov’s laws…

If it were not for some technical issues that are worth exploring.

The Symbolic Problem

From a technical perspective, for these laws to be accessed and used as a guiding framework, a source of information or an action classifier, they would need to be programmatically encoded in logical commands or through a mathematical function, but there is a problem at the root of this: they were written in English.

The so-called Natural Language, spoken or signed by human beings, is very different from programming languages, used to create algorithms, computer programs and web pages. In order to program, there must be no ambiguity or vagueness. No algorithm can be ambiguous (although it can be random or pseudo-random, but this is not the scope of this text to discuss). However, human language has a lot of ambiguity and vague terms that would cause problems for machines.

As Rob Miles says in the video “Why Asimov’s Laws of Robotics Don’t Work”, we know what a “human being” is not because we have a defined set of characteristics and properties that satisfy the category “human being”, but because we learn using some generic association mechanism, our learning. How can we define what a human is for a robot? What characteristics make up a person? Is someone who has died still human? And is a fetus a human being? To this day, the topic is a subject of debate. An example of this long-standing debate, with relevant political aspects, is the discussions about a woman’s right to abort, and the rights she has over her own body.

The body is a relevant piece of data for analyzing the third law. What would be the “very existence” of a robot that must be protected? If you argue that it is the integrity of the hardware, then removing a battery or replacing a part could be threats, but if it refers to  only the software, then a robot could, through neglect, not update itself, for example. A simple reflection, therefore, with a slight exercise of imagination, shows how vague these words, these symbols, are.

Nowadays, artificial intelligence has become synonymous with machine learning, so it would be possible to try to imitate people and, instead of defining these terms formally and mathematically, have the AI ​​learn them statistically. The problem is that there are also complications in this approach.

The Statistical Problem

When we talk about machine learning algorithms, such as neural networks, decision forests or support vector machines, it is important to understand that these models need data to learn. Furthermore, “learning”, in this case, is optimizing a function such that, given an input, an output is produced with the purpose of classifying that input, grouping it, and/or modifying the environment in which the system operates.

Regarding Asimov’s laws, to make them feasible we would need a large set of data so that we could classify an action as something that breaks or respects the laws (following their order of precedence). This requirement  is already susceptible to the problems we have today when we talk about artificial intelligence systems. How would this data be collected? How would it be represented? And what about the biases and impacts of these on people?

There are many cases of people being harmed by systems that make decisions based on AI models. Facial recognition, credit granting, recommendation systems, in all these fields we see news about how algorithms can be racist, misogynistic, transphobic (because technology is not neutral, as Clarissa Mendes explains very well in her text Artificial Intelligence: Is the Machine Neutral, and Humans the Ones Who Corrupt It?).

In addition, it would be necessary to define some kind of mathematical function that could calculate how much a person or the “very existence” of the robot is being harmed (remember the symbolic problem?), which in itself is a very complicated task.

The Ethical Problem

Finally, moving away from the technical aspects of computing, which were only presented here without much depth, there is a danger in using these fictitious principles in current and real debates about AI, particularly when it comes to  ethics.

Reducing the ethics of artificial intelligence (in robotics and beyond) to these laws disregards centuries of ethical research and debate, ignores the various schools of thought that should be considered in these discussions, and makes them superficial, favoring  the development of these technologies with little or no criticality. The topic becomes even more relevant in the context of creating regulatory standards for AI around the world. So, for all these reasons, do not mention Asimov’s laws in real discussions.

[1] ASIMOV, Isaac. Eu, Robô. Translated by Aline Storto Pereira. 1st ed. Editora Aleph, 2014.

Lunara Santana

Pessoa pesquisadora do IP.rec nas áreas de Inteligência Artificial e Regulação de Plataformas Digitais, possui graduação em Ciência da Computação pela Universidade Federal de Pernambuco (UFPE) e tem experiência nas áreas de aprendizagem de máquina, processamento de linguagem natural, recuperação de informação e engenharia de linguagens de programação.