There is often talk of cognitive bias and of bias artificial intelligence, but what exactly are and how do they influence our mind? In short, to have a bias It means having one preference or a prejudice systematic towards something or someone. There are different types of bias, but in this article we focus on cognitive biaslinked to the functioning of the human mind, and Algorithmic Biaswhich emerge in the AI systems when they learn from distorted data, and how these two levels influence each other. The cognitive bias are mental shortcuts: automatic mechanisms useful to speed up the understanding of the world and decide quickly. These mechanisms condition our world view and have a impact on development and use of “intelligent” technologies. And not only that: a recent study has shown that theprolonged interaction with to which present Bias lead to a Increase of our own prejudices.
The cognitive bias are mental shortcuts
THE cognitive bias They are mental shortcuts that our mind adopts daily to elaborate and interpret the large amount of information to which we are exposed. These evolutionary mechanisms allow us to make quick and efficient decisions in a complex and overload of stimuli environment, helping us to save time and energy. However, precisely because they simplify, they can lead us to systematic evaluation errors.
There are hundreds of cognitive bias, but, in general, they respond to Four great needs of our mind:
- select relevant information;
- give a meaning the information collected;
- act quickly;
- decide What is important to remember.
For example, when we have to manage too many information, we tend to select and memorize only those that confirm what we already think: this is called “confirmation bias“. Another very common bias is thehaloso we tend to think that a beautiful person aesthetically is also intelligent, or that an authoritative person in an area is also competent in other fields. As for our interaction with technology, however, a very relevant bias is theAutomation Bias, That is, our tendency to trust more decisions taken by automated systems (such as AI) rather than humans. This is particularly dangerous when AI also contains its own bias.
Bias in artificial intelligence depend on the data
In the case of Algorithmic Bias of the AI, we are not talking about mental shortcuts, but of systematic preferences – or structural discrimination – which emerge in the data with which AI is trained. The AI models learn by identifying patterns and regularity in the data: if these data contain prejudices towards a specific category, the algorithm learns them e reproduces themamplifying them.
For example, in 2017 it was discovered that many facial recognition systems they did not recognize the faces of black womensimply because they had been trained on dataset full of images of white men. In the same period, it emerged that the algorithms for the selection of Amazon staff tended to penalize female CVs and select almost only men. This is because the data used for training presented a great imbalance between men and women.
An extremely recent example comes from United Kingdom: the Ministry of Justice has started the development of a predictive system to identify people “at risk” to commit murders. The project uses data from police forces, justice and health services to process predictive profiles also on subjects who have never committed crimes. The criticisms raised by experts and associations for civil rights concern the concrete risk that these systems end up aiming vulnerable subjects, already discriminated against, belonging to minorities or fragile socio-economic contexts. In this case, AI does not limit itself to reproducing inequality: the institutionalizes.
Bias in the Ai worsen our prejudices
The problem does not stop inside the AI. According to a study by the University College London (UCL), When we interact for a long time with Ai who present Bias, we tend to adopt them too. In this experiment, over 1,200 participants have made a series of decisions by helping either by an algorithm or other human being. Those who interacted with a sexist who then tended to underestimate women and overestimate the competence of men. This effect was more marked than those who received the same suggestions from another human being. This is a direct consequence of theAutomation Bias: many people attribute greater accuracy and impartiality to automated systems, underestimating their impact on their judgments and therefore resulting more vulnerable to their influence. This effect is particularly dangerous for the most impressed subjects, such as children, who can internalize the bias of the AIs more ease.
Fortunately, the study also shows an encouraging aspect: this effect also works in the opposite direction. THE’Interaction with an AIA without or almost bias, improves human judgments. For this reason, those who develop and distribute these tools have one fundamental ethical responsibility: not only avoid perpetuating existing bias, but also preventing their amplification
