Models and good evaluative practices to detect impacts, risks and damages of artificial intelligence

Abstract

Starting from exemplifying and recognizing the impacts, risks and damages caused by some artificial intelligence systems, and under the argument that the ethics of artificial intelligence and its current legal framework are insufficient, the first objective of this paper is to analyze the models and evaluative practices of algorithmic impacts to astimate which are the most desirable. The second objective is to show what elements algorithmic impact assessments should have. The theoretical basis for the analysis of models, taken from Hacker (2018), starts from showing the discrimination due to lack of guarantees that the input data is representative, complete, and purged of biases, in particular historical bias coming from representations made by intermediaries. The design to discover the most desirable evaluation instrument establishes a screening among models and their respective inclusion of the elements present in the best practices at a global level. The analysis sought to review all algorithmic impact evaluations in the relevant literature at the years 2020 and 2021 to gather the most significant lessons of good evaluation practices. The results show the convenience of focusing on the risk model and six essential elements in evaluations. The conclusions suggest proposals to move towards quantitative expressions of qualitative aspects, while warning of the difficulties in building a standardized evaluation formula. It is proposed to establish four levels: neutral impacts, risks, reversible and irreversible damage, as well as four protection actions: risk prevention, mitigation, repair and prohibition.

Cite this work

No public copy is stored for this record, so only the metadata is cited.

Stored only when both papers are in the corpus — 10 citations exist upstream.

Papers that cite the same references — the relational fallback while embeddings are still backfilling.