A few weeks ago, newspapers reported on a Facebook artificial intelligence (AI) system labelling a video of black men as ‘primates’. To those working in the field of ethics of AI the case sounds both familiar and disheartening. It is familiar as it reminds us of the case reported in 2015 of a Google image recognition AI system which labelled images of Afro-Americans as images of chimpanzees. It is disheartening as after almost seven years one would hope that some progress had been made in addressing unintended and unwanted outcomes of AI systems.

The case of the Facebook AI system and the many cases of AI perpetrating human bias and discrimination show that there is still work to do at the intersection of ethics, technology and society to address these issues. The question is what the best approach is to ensure that AI systems behave respecting the fundamental values of our societies.

Solutions based on ethics-by-design approaches work only in part, especially when considering online learning AI models which can change their behaviour by adapting to their environment. Learning systems are unpredictable and ethical values that may be embedded in the model at design stage may be altered by the model itself after interacting with the environment and lead to new unintended (possibly unwanted) behaviours.

Because of this unpredictability, trust in AI is also not a good idea. We should not trust AI to make decisions and perform tasks that have a serious impact on individual, societal and environmental life. Not trusting AI is not tantamount to not using it, trust implies the delegation of a task and the lack of control or supervision as to how the delegated task is performed (Taddeo 2010): a sort of “trust and forget” dynamics (Taddeo 2017). Cases like the Facebook, Google teach us that these dynamics can lead to severe ethical problems, which if left unaddressed may hinder individual rights, justice, and social justice, as the nefarious cases of COMPAS) and the Allegheny Family Screening Tool have shown.

AI technologies have a huge potential to improve individual lives and to address societal and environmental issues (Cowls et al. 2021), but this potential cannot be leveraged without a strong focus on the ethical risks that AI poses and solutions to mitigate these risks. Two elements are key to this end: a more interpretable technology (Rudin 2019) and appropriate auditing mechanisms to identify ethical problems in the life cycle of AI systems (Mökander and Floridi 2021).

More interpretable AI systems are systems whose behaviour can be investigated to understand what determined it and to identify, and correct, the source of unwanted outcomes. Auditing mechanisms are a form of monitoring over the entire design-development-deployment cycle of AI technologies. It is what Mökander et al. call:

“Ethics-based auditing is a governance mechanism that can be used by organisations that design and deploy AI systems to control or influence the behaviour of AI systems. Operationally, ethics-based auditing is characterised by a structured process by which an entity's behaviour is assessed for consistency with relevant principles or norms” (Mökander and Floridi 2021, 324; emphasis added).

AI technologies are a double-edged sword. On the one side, they have a huge potential to improve human and environmental conditions, as shown by the many cases of AI projects supporting SDGs oriented projects (Taddeo and Floridi 2018). On the other side, the risk eroding the fundamental values of our societies (Yang et al. 2018). “Trust and forget” (Taddeo 2017) and ethics-by-design approaches to AI do not mitigate these ethical risks. To achieve this result, it is crucial to develop appropriate forms of monitoring of AI systems and foster innovation that facilitate interpretability of AI models, so as to make AI processes reliable.