Questão 11 Albert Einstein 2027 · Vestibular 2027 · Prova I: Interpretação de texto

Read the text to answer questions from 11 to 14.

1 If artificial intelligence (AI) systems shape decisions that affect people’s lives, they should do so impartially. But guaranteeing the fairness¹ of AI systems is far more complex than it might sound. What counts as fair in one domain may be inappropriate in another. In criminal justice, fairness may prioritise avoiding disproportionate harm to particular communities. In education, it may focus on equal opportunities for all. In finance, it often involves balancing access to credit with risk analysis.

A deeper issue lies in the data itself. AI systems learn from historical datasets that reflect past decisions, institutional practices, and social inequalities. When a model is trained to replicate observed outcomes, it may reproduce existing injustices under the appearance of objectivity.

Let’s take a look at hiring decisions. If a company historically promoted fewer women into senior roles, an AI system trained to predict “successful” job candidates may learn patterns that favour characteristics more common among men, even if gender is not explicitly included as an input. A fundamental question is whether AI systems mirror the world as it was, or attempt to correct for known injustices. The idea of fairness is also complicated by how it is assessed. Many assessments examine a single protected attribute, such as gender or race, in isolation. While common, this approach can obscure how discrimination operates in practice. An automated hiring system might appear fair when comparing men and women overall, and fair when comparing ethnic groups overall, yet it might also consistently disadvantage older women from minority backgrounds.

5 Societies change, demographics shift and language evolves. A system that performs satisfactorily today may produce unfair outcomes tomorrow. In particular, recent advances in large language models (LLMs)² add further complexity. Unlike traditional systems that make specific predictions, LLMs generate language based on vast collections of historical text which, inevitably, contain imbalances. Fairness is therefore an ongoing responsibility requiring monitoring, and a willingness to revise or withdraw systems when harms emerge. Together, challenges suggest that fairness in AI is not a purely technical problem but a moving target shaped by social values and historical context.

(Michael Mayowa Farayola. https://theconversation.com, 12.05.2026. Adapted.)

1 fairness: the quality of being impartial and free from bias.

2 large language models (LLMs): AI systems trained on massive amounts of text data to understand and generate human-like language.