This answer is closely linked to what I explained in the previous point about bias. If we start from the premise that AI is created by humans and trained with data generated by humans, the logical conclusion is that the technology inherits our own flaws. Humans always have biases (conscious or unconscious), and when we pour our history into a model, we are teaching it those same prejudices. Therefore, it is a mistake to think that AI will solve inequalities on its own; in fact, there is a risk that it will exacerbate them. If we train a hiring algorithm with data from a company that has been discriminating against certain groups for 20 years, the AI will learn that this is the “pattern for success” and will automatically replicate it on a large scale. The danger is that this creates a kind of “black box” where discrimination becomes invisible. Before, you could blame the recruiter; now they hide behind the excuse that “the algorithm decided it”. Thus, far from closing social or economic gaps, AI can act as an amplifier that makes structural inequalities more difficult to detect and eradicate.
Since AI systems access existing and past data, it could, for example, application process puts certain groups at a disadvantage if previous recruitment dates were one-sided. In the case of lending, people from certain social or social backgrounds could be regional groups. In law enforcement, too, there is a risk of that AI is more likely to classify certain populations as "conspicuous" when the training data is skewed. As a result, existing injustices can be seemingly objectively and be continued on technical grounds.
There is a high risk that AI will exacerbate existing inequalities if systems are trained with biased data or are inadequately monitored. Women, certain ethnic groups or people from disadvantaged neighbourhoods may be systematically disadvantaged because historical prejudices and structural discrimination are perpetuated in the training data. The situation is further exacerbated by non- transparent "black box" decisions, which are difficult for those affected to challenge if they do not understand why they have been denied a job, a loan or a service, for example.
AI is always a reflection of humanity, and the user. I believe existing inequalities could persist, only now, they can be mostly automated. This, in turn, would be very convenient for those people who reinforce inequalities, by stripping them of the need to see faces, hear voices, and deal with the heavy weight of conscience. Furthermore, it will give them an excuse for whatever unjust behavior, by just blaming the model for it.
Just as it can amplify these disparities, it can also level them out. The trend suggests that these areas may soon become a battleground for AI models, since what can be used to create inequality can be countered by another model. The inequality here is that someone might be tricked into conducting (for example) a live interview with a real person, while the AI is the one evaluating them. But it is not impossible, for example, that after an objective AI evaluation, a person might be hired whom a human HR representative would not have hired—perhaps due to some sudden personal dislike unrelated to the candidate’s qualifications, or simply due to the examiner’s incompetence. Here, too, however, the rule applies: AI is an assistant, not a replacement. !
