Ethical use of AI in Kela benefit handling

Ethical use of AI in Kela benefit handling

• 1.
  Ethical use ofAI in Kela benefit handling 10.12.2019 Kela Conference on Social Security Riku Sarlin Chief Architect Kela, Benefit Services Development @RSarln
• 2.
  Contents • Automated decision-makingin Kela – a short history • Practical, legal and ethical challenges of AI in administrative decision- making • Ethical principles of Kela in applying AI • Ethical use of AI in Kela benefit handling 10.12.2019 Riku Sarlin | Ethical use of AI in Kela benefit handling2
• 3.
  Automated decision-making inKela – a short history • Automated decision-making has been in the heart of Kela’s activities since the dawn of computer age – since the beginning of 1960’s • Traditional use-cases of automated decision-making include • handling automated cost-index related changes in benefits • automated handling of address changes • automated payment of benefits • The majority of automation has been done with batch processing. Still very much in use! • The last 10 years have brought two additional means of automation: process automation and robotic process automation. These are in use, but not extensively. • All of these methods use procedural logic to handle benefits. Actually, the same logic and same calculations are used in manual handling, too! • GDPR Act 22 is quite strict on decision-making based solely on automated processing 10.12.2019 Riku Sarlin | Ethical use of AI in Kela benefit handling3
• 4.
  Practical, legal andethical challenges of AI in administrative decision-making • Practical: • Constant changes in legislation – including cost index changes – make it difficult to keep AI models up-to- date. In short, each time legislation changes, the model must change. But - we do not have new data in all cases at hand. • Some benefit applications include a plenty of open-ended questions. It is not easy to handle these automatically with or without AI. This must change for comprehensive automation to be possible. • Legal and ethical • It is possible to interpret GDPR in such a way that it must be possible to “open” the decision taken by the algorithm. In many AI algorithms, this is quite challenging. In comparison, the decisions taken by procedural logic are relatively easy to “open”. • An administrative decision must, according to Finnish legislation, be “reasoned”, and include references to facts taken into account, the documents considered and sections of law applied. With AI, this reasoning is actually a challenging problem. With procedural logic, this is relatively easy. • AI models are based on statistical inference, that is, previous cases . Many benefit require "individual consideration". This is not a problem-free relation. 10.12.2019 Riku Sarlin | Ethical use of AI in Kela benefit handling4
• 5.
  Ethical principles ofKela in applying AI • Responsibility • We take into account economical and ethical sustainability of our implementations. We use information and insights gained in a responsible way, and require the same from other parties, too. • Transparency • Kela utilizes AI in an open and transparent way. We tell our customers in a clear and understandable way where, why and how AI has been utilized. We openly tell the algorithms used as well as input and output data of the algorithms. We describe the data used for teaching and testing our AI models. • Human-centered approach • We utilize AI and data to improve well-being of citizens. We let our customer participate in development of our AI solutions. We respect peoples’ fundamental rights and protect privacy in all applications of AI. 10.12.2019 Riku Sarlin | Ethical use of AI in Kela benefit handling5
• 6.
  Ethical use ofAI in benefit handling • The challenges presented are difficult to overcome without major changes in Finnish general and social security legislation. Practical challenges would still remain. • It is important to note that some challenges are common to all forms of automation, not just AI • In this situation, AI currently has a supportive role in benefit handling in Kela. It has no role in actual administrative decision-making. Current uses include: • Automated classification of inbound (scanned) paper documents • Turning documents received from e-service channel “the right way up” (not yet in production, technically ready) • Chatbots - a few are in production. Currently only “general advice” is provided, since no personal data is handled. • It is likely that this “supportive role” will stay the same in near future. Services based on natural language processing (NLP) are likely to improve and expand quite much in the next couple of years. 10.12.2019 Riku Sarlin | Ethical use of AI in Kela benefit handling6
• 7.
  You can sendfurther questions in email! Riku Sarlin Chief Architect Kela, Benefit Services Development Riku.sarlin@kela.fi @RSarln