Across myriad industries there is a recognition that AI is an emerging technology that has a real place at the heart and centre of many different business processes. AI explainability is therefore an important part of any business plan where AI is planned to be used. With increased use of AI within business, we will see an increased use of personal data, and this is where it is important to have AI explainability in a transparent and simple fashion. AI systems rely on data sets in order to be as efficient as possible. The longer an AI system is in place, the greater the amount of data it will process and the more efficient and effective it will become in theory.
There are 5 techniques and processes that each business should go through to consider AI explainability as part of GDPR protocols, a vital part of data protection regulations where personal data is being collected and processed.
1. Data protection Impact Assessment (DPIA)
Whenever there is a project where data processing could lead to a high risk to data subjects, there is a need to conduct a DPIA to assess and minimise risks. The use of AI processing should always mean that a DPIA is triggered. This is due to the volume of data sets that are relied upon for an algorithm to be effective. A DPIA should be conducted before any personal data is processed. Alongside a DPIA, an Algorithm Impact Assessment (AIA) must also be conducted. An AIA will go deeper into the use of AI within a project than a DPIA, assessing the potential impact of every single stage of a project.
2. Take care of individual’s rights
Within both UK and EU GDPR there is a stipulation that data subjects have the right to be informed about the reason their personal data is processed and how it is processed. This must be demonstrated in a clear way, providing explanations that are simple to understand for the data subject, stressing the clear processes involved in a completely transparent way. The key components of this AI explainability are transparency and accountability. There must be a strong justification for the reasons behind collecting personal data. You must explain how the data will be used, where it will be stored, and what will happen to the data when it is no longer pertinent data and must be removed in a secure manner.
3. Removal of discrimination bias
There are a couple of ways in which discrimination can take place within AI technology. The first is that there is intentional discrimination built into the algorithm for decision making. This could be seen where a certain group of specific people (say women, people over a certain age etc) are discriminated against and a negative decision made regardless of any other factors. Unintentional discrimination bias could come to a head where an original data set has only a certain type of information within it, and the AI has not yet learned to deal with other factors and information, with weighted prejudice against a different type of data that what has previously been present. So, for instance, women against men in a loan application process where previously there was only male data present.
4. Implementation of Article 22
Article 22 of the GDPR is intended to protect individuals from any adverse decisions that affect them that are solely made by AI and automated means. These decisions are often made using algorithms without any or little human involvement. There are different degrees of AI decision-making capabilities, with some not fully autonomous. With the rise in AI systems and their use to make decisions across a whole host of industries and sectors it is vital that Article 22 is adhered to at all times.
5. Meaningful human review
Last, the implementation of Article 22 relies heavily on meaningful human review wherever there is a decision that could have a legal or adverse impact on the data subject. As AI and machine learning use increases this will become more important. It is not enough to just rubber stamp a decision made by an algorithm or AI system. Instead, it is a requirement to have a meaningful review of any data and decision that is made by AI, taking all the information and data, and making a decision based on that. AI can be useful in assistance but should never be the sole means to decide a person’s fate, under any circumstance.
By looking at these five techniques and processes as a way of ensuring you have explainable AI benefits that data subjects can clearly understand, you’ll put transparency front and centre of your organisation. Data protection is an important part of any business in the modern age. As a greater number of companies utilise the benefits of AI and automation within many different processes and decision-making arenas, there is a real requirement for clear thinking when it comes to AI explainability. Data subjects must understand the how and the why when discussing the use of AI and their personal data.