POLAND Trends and Developments Contributed by: Barbara Kiełtyka, Jakub Gładkowski and Małgorzata Kiełtyka, Kieltyka Gladkowski KG Legal
Legal Tendencies The current European legal framework for AI systems in healthcare has emerged from a tendency towards the creation of both comprehensive sectoral regula - tions, as exemplified by the latest European Health Data Space (EHDS) regulation, and cross-cutting regulations, which comprehensively regulate the intro - duction to the market, supervision and responsibilities of AI system manufacturers across all industry sec - tors, as exemplified by Regulation (EU) 2024/1689 (AI Act). As a result, AI healthcare systems in Europe operate in a multi-layered legal environment, and this trend is set to continue. Training Algorithms in Healthcare AI Working with AI algorithms begins with collecting and preparing the appropriate data to be used to train models. AI components are then used to analyse this data, recognise patterns and make decisions based on the information collected. Because of this, the transparency, representativeness and adequacy of the data used to train language models in healthcare AI systems become crucial. Already in the vertical large language models, the cre - ators, within the meaning of the AI Act, are focusing on the fact-checking in those language models. This is especially true for horizontal models, ie, multi-agent models of AI systems, which hold promise for break - throughs in medicine and scientific breakthroughs (superintelligence). In practice, the problem is that these AI networks are trained on so-called natural language, and the lan - guage models that search and process online resourc - es treat these resources as a representation of reality. Unfortunately, this information is sometimes false, yet the language models treat it as true. Infecting a lan - guage model with erroneous sources is therefore also one of the major challenges in the healthcare sector, where AI systems provide informational and scientific functions in medicine. For example, a language model can be “fed” false or falsified information about adverse drug effects, false scientific papers, non-existent medical disciplines, or false cases of medical errors in procedures involving a medical device or clinic. Therefore, the feeding of
language models with false information, for example, that a given drug works and has superior effects, or the feeding of an AI system model with information about an allegedly large number of medical malprac - tice lawsuits involving a given drug, can pose a very significant challenge in applying AI systems in the medical sector. AI systems are exposed to services for manipulating internet content on a large scale, so that the fact- checking algorithms verifying real information for net - work training are lost, and positive information about a product or negative information about a service or institution is smuggled into the trained networks. Therefore, from the perspective of the fundamental functionality of AI based on a language model, special regulatory oversight is necessary for the authorisation of AI systems, based on fair competition standards, data processing in the clinical evaluation process, and regulation of the introduction of tools and drugs to the market. The challenge is to create tools that the healthcare sector can use to identify itself online to ensure that the data sources used for training come from real persons, given that 60% of social media content is created by bots for specific purposes. Because of the above, sentiment analysis algorithms used in language models to “excise” emotional over - tones are insufficient to create the required level of objectivity, and this means AI systems in healthcare must be classified as high-risk AI systems. A chal - lenge legislators face when regulating AI is the issue of liability for errors – especially when AI systems oper - ate fully autonomously. Practical Challenges of Liability and Safety Currently, there are two approaches in law to liabil - ity in AI: one that focuses on liability for a “defective product” and one that places the liability on the user making the final decision. Therefore, AI transparency is crucial. Another challenge is the protection of per - sonal data and privacy. The EHDS framework intro - duces a system of electronic medical records, which can also be used for research and innovation. The risk of false data and results must also be considered. Large AI models process vast amounts of information, the reliability of which can be difficult to verify, which can lead to erroneous analyses and clinical decisions.
110 CHAMBERS.COM
Powered by FlippingBook