Logic and foundations of artificial intelligence and society’s reactions to maximize benefits and mitigate harm

Kaufman, Dora

doi:10.4013/fsu.2024.251.10

Acessibilidade / Reportar erro

Brasil

Español English

sumário « anterior atual seguinte »

Sumário

Dossier • Filos. Unisinos 25 (1) • 2024 • https://doi.org/10.4013/fsu.2024.251.10 copy

Logic and foundations of artificial intelligence and society’s reactions to maximize benefits and mitigate harm

Lógica e fundamentos da inteligência artificial e as reações sociais para maximizar benefícios e mitigar danos

Authorship SCIMAGO INSTITUTIONS RANKINGS

ABSTRACT

Artificial intelligence is a general-purpose technology (GPT), term given to technologies that shape an entire era and reorient innovations by reconfiguring the economy’s logic and functioning and bringing in new business models. AI offers unprecedented opportunities and risks. The benefits of AI are extraordinary, as are its potential harms. Potential damage does not have the same degree of problematization, since the intensity and extent of the damage varies according to the domain and the object of application. To address the scale of this challenge, regulation is necessary but not sufficient. Standards, unwritten codes of compliance and arbitration procedures, supervision and auditability, AI governance, international agreements, compliance with current local and global standards and laws. All of this needs to be integrated. Society seems to have no alternative to facing the challenges of at least mitigating the damage already identified and trying to predict future damage in advance. The purpose of this article is to encourage reflection regarding the main initiatives that are available to society to protect its citizens and organizations from the potential harm caused by AI models, vis-à-vis the technology’s own limits to act in ethical and legal compliance.

Keywords:
artificial intelligence; opportunities and risks; regulation

Sources of bias	Explanation
When generating data	Discrimination in data production is present both in the predominance of users from developed countries who have greater access to technology and social networks, which results in a database biased by the light-skinned racial biotype, and in the failure to disaggregate data by gender and/or the treatment of men as “standard human beings”.
In the choice of developers	During the development of a DLNN model, the initial task for computer scientists is to identify the problem to be solved by the system, in which situation and for what purpose the system will be used. The second step is to translate the problem to be solved into variables that can be observed and manipulated (feature engineering process). These are the things that define, for example, which research terms will be used to collect the data, the number of hidden layers and the number of nodes in each layer. Identifying the influence of human subjectivity on the design and configuration of the AI algorithm is no simple task, nor is it possible to eliminate it even if it is identified. The Alan Turing Institute points out that one of the critical problems that allows systemic biases to infiltrate data is the attitude of algorithm developers and designers, who do not prioritize actions to identify and correct potentially discriminatory imbalances in demographic and phenotypic representation. The institute attributes these biases to the complacency of technology producers, who are generally part of the dominant group and therefore exempt from the adverse effects of discriminatory results.
In the database	Bias occurs if the reference data is less demographically diverse than the target population, in other words, if the database contains few or no examples of a specific sub-population by ethnicity and/or gender. The difference between controlled environments (laboratories) and uncontrolled environments (the real world) also has the potential to produce biased results; for example, on the street, cameras can capture images in low resolution, the angle captured of the face and the lighting can make it hard to extract facial features or even distort them, causing errors in facial recognition.
In the data labeling process	Creating a training database means sampling an almost infinitely complex and varied world, and fixing it into taxonomies made up of classifications. [...] Maintaining uniformity in the manual classification of large data sets is a challenge, which becomes almost impossible when it entails classifying images of people; there are countless classification categories, including race, age, nationality, profession, economic status, behavior, character and even morality. Structuring a taxonomy in order to classify images of people with the logic used for objects generates countless distortions and, consequently, biases.
In the training data of the algorithms	It is deemed that there is a bias in the database when the system displays a systematic error in the result (“statistical bias” or “algorithmic discrimination”). Strictly speaking, any dataset could be impartial for the performance of a given task. However, potentially there is a risk that, if it is used for a different task, it will be biased towards that second task.

Universidade do Vale do Rio dos Sinos - UNISINOS Av. Unisinos, 950 - São Leopoldo - Rio Grande do Sul / Brasil , cep: 93022-750 , +55 (51) 3591-1122 - São Leopoldo - RS - Brazil
E-mail: deniscs@unisinos.br

Acompanhe os números deste periódico no seu leitor de RSS

[1] Email:
kaufman1955@gmail.com