What are the ethical boundaries of Artificial Intelligence?

A movie recommendation on Netflix, an album suggestion on Spotify, and the rental price on 5º Andar: what do these three things have in common? They are all governed by algorithms.

Algorithms and other mechanisms linked to artificial intelligence (AI) are increasingly present in our lives, defining our choices, guiding our tastes, and curating what we hear, read, watch, buy, and listen to. But beyond recommending that new Beyoncé song, algorithms are already responsible for determining your credit score and, in some places, even the likelihood of someone committing a crime. And it is precisely in these sensitive areas that the question arises: to what extent is a decision made by an AI valid? What are its ethical limits?

Technology is a human creation. It may seem like an obvious statement, but it is necessary to emphasize this aspect inherent to any technology. The technology used to build AI is not neutral: it is human-oriented and, if it does not observe clear principles, it can reproduce patterns of prejudice and discrimination that are essentially human. More than just not being neutral, technology is also not impartial—which is why we must think about the ethical and moral implications that decisions made by AI can have on people's lives.

We have as an example the COMPAS (Correctional Offender Management Profiling for Alternative Sanctions), a software used by the justice system in the state of Wisconsin, in the United States, to calculate the risk level of criminals. Other U.S. states also use similar software to help in the decision-making process for judicial proceedings. In this case, the system uses data on family, educational, and professional history to create a "score" for defendants.

Race is not a criterion for the algorithm, therefore, all individuals are "equal" and answer the same questions. But that is precisely where the distortions begin: the U.S. social and economic context makes the system more rigorous and points to a higher possibility of Black individuals reoffending than white people, something that is not observable in reality (as shown in this article by Pública), but is explainable by the social context in which Black people live in the U.S.

In other words, in the United States, a Black individual is more likely to live in a poor area and have relatives with records in the prison system, but this does not mean that because of this family history and economic condition they are more prone to committing crimes than a white subject who does not have this family history, but has committed more serious offenses or has a much more extensive criminal career. The algorithm is not intentionally racist, but the decisions made by it are.

‍

‍

Another example of how artificial intelligence mechanisms can directly influence the lives of thousands of people is the credit score calculation, a procedure performed by the two largest credit bureaus operating in Brazil, Serasa Experian and Boa Vista SCPC. The score is a rating that determines a consumer's chances of getting a loan, financing, or even store credit. It is possible to check your score on the Serasa Experian website, but, more importantly, the service makes it clear which criteria are evaluated for the calculation.

This is a principle of transparency that is not always followed by other banking institutions. We know that an internal rating exists, but we do not know which criteria are taken into account for the analysis and what data is used to obtain this "internal score." Data such as education level, race, and address can be used improperly and, in this way, generate distortions and prejudiced conclusions about the consumer. And, in these cases, the lack of transparency in the process prevents us from knowing how we are evaluated, which leaves us at the mercy of these hidden systems.

‍

‍

The social context must be taken into account when we use AI to make decisions that directly influence people's lives. And, above all, these processes must be transparent, auditable, reliable, and, if possible, validated by human oversight. However, for this to happen, we face two major challenges: the first is how to translate a set of codes for the general public, and the second is how to prevent algorithms from making erroneous decisions, but if they do, how to assess and correct these distortions?

Thus, it becomes necessary to discuss and create a governance policy and best practices directed at AI, which is similar to what is being developed with the recently approved General Data Protection Law (LGPD). But for this to happen, society and political actors must also turn their attention to the possible implications of how this data will be used and what the consequences of decisions made by algorithms and AI will be.

A good example comes from the European Union, which this year released a document with guidelines for the creation and development of human-centric artificial intelligence technologies. The document, which has no legal status, is the result of a commission created specifically to guide the development of artificial intelligence in the bloc. The motto that guides the guidelines is that the development of artificial intelligence technology must be "human-centric." Saying this means making explicit the political and social character of this set of codes, since they will be applied in society and, therefore, must respect human rights and cannot commit discrimination or prejudice.

Humanizing technology and taking it out of the "neutral" field seems to be a good path for the development of fairer AI and algorithms. This must be taken into account from the very construction of these technologies, which also implies the diversity of the teams responsible, but that is a subject for another article. The report in question signals seven points as being essential for building reliable and human-centric AI technologies:

·Human agency and oversight

· Technical robustness and safety

· Privacy and data governance

· Transparency

· Diversity, non-discrimination, and fairness

· Societal and environmental well-being

· Accountability

These seven guidelines are core principles aimed at building AI mechanisms that are socially responsible, ethical, and respectful of human dignity. While it is impossible to calculate the impact that new technologies will have on our lives, we must, at the very least, consider ethical and responsible models. Perhaps in the near future, we could create a kind of universal code, similar to Asimov’s Laws of Robotics, directed exclusively at AI.