High system accuracy is usually one of the design goals of AI systems. Many AI systems require accurate and reliable training data to achieve the best results. When processing personal data, keeping it up to date and correcting wrong inputs is also a legal requirement.[1] The data subject can also demand the rectification of inaccurate personal data.[2] AI systems should therefore be designed with the need for retraining in mind, during which data may not only be added – but also removed (see “Right to rectification” within Part II section “Data subject’s rights” of these Guidelines and “Fairness, diversity and non-discrimination”within this Part III on AI, as well as, “Lawfulness, fairness and transparency principle” within Part II section “Principles”).
In addition, the output of an AI system should not only be a result, but also a measure of how confident the system is that the result is correct. Such a measure is not only a technical indicator of the system’s accuracy, but also a valuable indication of whether human intervention may be required (see the “Accuracy principle” section in the “Principles”within Part II of these Guidelines).
References
1Article 5(1)(d) of the GDPR. ↑
2Article 16 of the GDPR. ↑