Part II: Algorithms’ impact on human behaviour
In this part, we address the following research questions:
— How do algorithms, when exploited in different applications, affect human cognitive capabilities?
— How do algorithms have the potential to modify the way humans make decisions based on them (e.g. influence of recommendations, personalization)?
— Which are the suitable strategies for effective human-algorithm interaction?
7 Artificial Intelligence: an interesting leverage point to rethink humans' relations to machines…and to
themselves.
Nicole Dewandre
Joint Research Centre, European Commission
7.1 Statement
The understanding and effect of the expression "Artificial Intelligence" are strongly conditioned by the implicit assumption that intelligence is –ideally-THE specific human (male) feature. "Artificial Intelligence" is spoken of, either from a creator's perspective,
i.e. with pride or fascination, or from a slave's perspective, i.e. with fear and resentment. In my contribution, I shall challenge both the creator's and the slave's perspectives and invite to a more agnostic and human-centric approach to AI.
- "Intelligence" is much more easily granted to artefacts than to humans. Ex1: A fridge is deemed to be smart when it sends a signal to inform about a lack of milk. When a man or a woman scrutinizes the fridge to check if there is milk, this is not considered a smart task. Ex2: Public lighting is deemed to be smart if it adapts to the type of user (pedestrian, car, bicycle). If a man or women was posted on each street to turn the lights on according to the type of user, this would not be qualified as a high-skilled job! In fact, artificial intelligence is granted to artefacts reacting to their environment, when human intelligence is, instead, granted to behaviours gaming the environment, in order to reach an objective or materialise an intention. Mere reactivity, when it comes to humans, is not considered as intelligence, but rather as weakness. This leads to rethinking human intelligence, and the role of intelligence in characterising humanness.
- In some places, we need to be reminded we interact with humans and not with machines. For example, a sticker encouraging saying "Hi" before asking for a ticket reveals that the default solution might have become to get a ticket from a machine instead of from the hands of another man or woman. In the same vein, interacting on websites, humans are asked to demonstrate to machines that they are humans, for example, through CAPTCHA, to be able to pursue the interaction. This reveals a generalisation of the fact that human-machine relationships are more and more positioning machine in the active role and humans in the passive or reactive mode. This way to see things occults the fact that machines and artefacts are owned and developed by agents, corporate or humans. So, instead of considering the human-machine interactions, the focus should be on human (owner/developer)-machine-human (user) interactions. This leads to rethinking the radical changes in the way (smart) artefacts mediate human relations.
7.2 Position regarding the research questions
How do algorithms have the potential to modify the way humans make
decisions based on them (e.g. influence of recommendations, personalization)?
As the way humans make decisions is highly conditioned by their environment, and the environment being more and more pervaded by connected artefacts and algorithms, it is obvious that algorithms will impact the way humans make decisions. I would be cautious using the term "modify" as if there was a "before" and an "after" algorithms in the way humans make decisions. This question has to be addressed with the baseline of actual decision-making (partly contingent, depending from partial information, based on some random or serendipity) and not idealised decision-making (based on perfect, unbiased information).
Which are the suitable strategies for effective human-algorithm interaction?
As suitable strategies might depend from addressing the users' perspective or the owner's perspective, I shall deliberately answer this question from the user's perspective. From the users' perspective, it is essential to protect the attentional sphere of the users. Potentially, connected machines and AI could cannibalise human attention to a point that human attention is totally "sucked up" by machines, and humans are prevented from directing their attention according to their own desire or to each other. I recommend focussing on the situation of users faced with multiple systems instead of thinking of each application separately. Besides protecting the vulnerability of our attentional spheres, it is also essential to re-create the conditions enabling trust, and making sure that fooling each other is not a winning strategy. Instead of addressing these issues through control, I would recommend a minima- recreating the conditions allowing each of us to know if he or she is in an environment which "recognizes" him or her, and what is the impact of this recognition. For example, when I look for a good or service on the internet, is the price which is offered the same that anybody else would be offered? And if not, what is the impact of the environment adapting to me (in this example, with dynamic pricing)?
7.3 Challenges
- Critical approach of intelligence: what it is; its role in characterising humanness; the articulation between intelligence, knowledge and power.
- What is the effect of AI on human relations, and on the relation between humans and their environment (mix of artefacts and nature)?
- What does AI not change? In other words, what are the invariants with the past?
References
— Cohen, J. (2012). Configuring The Networked Self, Yale University Press.
— Floridi, L. (Ed.), The Onlife Manifesto, Springer, 2015.
— Pasquale, F. (2015). The Black Box Society, Harvard University Press.
— Ganascia, J.G. (2017). Le Mythe de la Singularité. Faut-il craindre l'intelligence articficielle? Seuil.
8 Algorithmic discrimination
Carlos Castillo¹
Universitat Pompeu Fabra
8.1 Statement
Statistical group discrimination is disadvantageous differential treatment against socially salient groups based on statistically relevant facts (Lieppert-Rasmussen, 2013). In this definition, a group is socially salient if membership is important to the structure of social interactions across a wide range of social contexts; this includes in particular categories that are protected by law, such as individuals with disabilities, and groups defined by properties that are a matter of anti-discrimination law, such as gender, age, religion, national origin, etc. Algorithms based on statistical learning can engage in statistical group discrimination, if we understand statistically relevant facts as any information derived from training data — indeed, there are many examples of this (Hajian et al.
2016). Hence, the interaction between humans and algorithms should be one in which the human is able to not only understand but also challenge algorithmic decisions. The "FAT" framework (Fairness, Accountability, and Transparency) has been advanced in recent years as a set of characteristics to which algorithms should adhere.