How Trustworthy is AI?

According to Wolfgang Hildesheim, Director Watson, Data Science & Artifical Intelligence Leader at IBM in the DACH region, Tech By Light (TBL) has interviewed him on policy and investment recommendations for trustworthy AI. Here is his answers to the question on ethical guidelines for trustworthy AI.

TBL: Mr. Hildesheim, according to a current PwC study, only six percent of companies in Germany use artificial intelligence or are currently implementing AI systems. What are possible braking factors?
Wolfgang Hildesheim: At the moment, companies are still separating many hurdles from the advantages of artificial intelligence (AI)to use for yourself. Braking factors are often insufficient experience or knowledge to successfully anchor AI in the company. Data silos and missing tools for the development of AI models are responsible on the technical side. However, there are also concerns about the trustworthiness of AI. Those responsible in the companies are concerned about whether they can trust AI and whether the new technologies also comply with all legal regulations. Another key question is whether and how the decisions of self-learning machines can be understood and checked.AI Developers are working hard to address questions on the safety of artificial intelligence, and must come up with mature trustworthy AI applications.

TBL: For which industries is AI generally best suited and why? Aside from chat bots and digital assistants, what are the concrete application scenarios?
Hildesheim: When we speak of Artificial Intelligence – or from the English term Artificial Intelligence (AI) – we mean functionalities that serve to help people to cope better with their daily tasks in working life. It’s about absorbing and understanding information, drawing conclusions, learning and interacting. This can support our decision-making, and additional knowledge can be gained from data and existing knowledge can be used by millions of people. Processes also become more efficient because simple steps are automated.

In addition, the interaction between man and machine can be simplified and personalized. The application possibilities are therefore almost unlimited: in research and development as well as in teaching and business, in supporting experts and in the optimization of processes, products and services, as intelligent assistance and mobility services and in many areas of the healthcare system. In this sense, AI rather means augmented intelligence, i.e. the expansion of human intelligence. It’s about “man with machine”, not “man against machine”.

The ability of intelligent AI systems such as IBM Watson is to recognize patterns in large amounts of data and to make intelligent predictions where human cognitive abilities are no longer sufficient. Specific application scenarios:

Accounting: Due to a new accounting standard, leasing contracts in companies have to be analyzed and evaluated with regard to the changed accounting rules. Together with KPMG, we have developed a solution in which Watson independently evaluates leasing contracts according to certain attributes and prepares them for further processing.

Legal advice: Lawyers spend up to 80 percent of their time on problems that they or their colleagues have already solved. The legal-tech companies 123recht.de and Frag-einen-Anwalt.de use Watson’s Artificial Intelligence to provide affordable legal advice. AI helps to find the right answer to repeating legal questions.

Manufacturing: AI-based systems can also support the maintenance technician, who should ideally solve a spontaneous technical problem in a manufacturing plant as quickly as possible or who also needs help with regular maintenance activities. The use of AI, image recognition, natural language dialogue and augmented reality techniques make the technician’s job easier and ensure that his tasks are carried out correctly and effectively.

TBL: How can you recognize a provider’s true AI skills?
Hildesheim:The topic “AI” is a mega trend and there are also a variety of philosophical, legal and social aspects discussed in the media, some of which are currently not of practical relevance. Hollywood in particular does a lot to ensure that expectations of AI are too high and that there are unfounded fears. Therefore, a good provider of AI solutions can be recognized by the fact that they can quickly develop the first simple AI applications in small steps – within weeks – and switch them “live”. All of these applications are “narrow AI” with low risk and small automation that humans control completely. Interested parties should let the provider show references and other successful customer projects and try them out personally. In this way, everyone can easily convince themselves

TBL: What are classic pitfalls when setting up AI projects in everyday business? What are the stumbling blocks in the implementation?
Hildesheim: Expectations are often too high. The most successful are the organizations that develop new AI skills in small steps and then go live step by step. It is important to optimize and expand the functions in close cooperation with the users: The solution must work to the satisfaction of the users. Typically, the first AI applications become public cloud environments developed with publicly available data. The AI ​​applications can then be expanded to include “real transactions” or the processing of your own information. Private cloud technologies are used in our own data center. This combination of public and private, also known as hybrid cloud, should ideally be used in the order described.

TBL: What role do big data play for AI systems?
Hildesheim: An AI system is only as good as the data with which it is trained. Data is the linchpin. Therefore, the two technologies go well together: AI can help find trends and patterns in big data that would otherwise be misinterpreted or completely undetected. It turns big data into smart data. At the same time, AI can only tap its full potential with big data, because it has to deal with extremely diverse data.

In fairness, however, it must also be said that many AI applications today do not need a lot of data, but that even good results can be achieved with just a little training data. In this respect, the question suggests that AI is always complex and always requires a lot of data. That’s not the case. With AI for word processing in particular, the right data is more important than the amount.

TBL: Keyword “colleague AI”: How do employees see the use of artificial intelligence? Are your concerns well founded? Where will humans have advantages over algorithms in the future?
Hildesheim: Our central guiding principle in the development and handling of AI is to support people with our technology in their work as best as possible, not to replace it. AI can provide relief and support, especially for repetitive, less demanding tasks, but they take a lot of time. If an AI takes on these tasks, there is more time for creative, complex questions.

TBL: What are the risks of AI systems and how can you minimize them?
Hildesheim:An example: When lending or pre-selecting in the application process, the learning machines refer to statistical models from a variety of data sources and parameters. The challenge here: If the data contain unconscious prejudices, stereotypes and old-fashioned role models, these are not only adopted by the learning algorithms, but additionally reinforced. If AI systems are not optimized with solid and diverse data sets, the accuracy can suffer, the results can be distorted and thus the fairness also suffer. That is why we, other AI developers and the research community as a whole have to pay close attention to what data is used for training purposes.

To transparent disclosure of the decision-making processes of AI also means that any prediction, any model version and any training data are recorded and stored – and companies to audit secure adherence to compliance guidelines and the DSGVO supported. This is particularly relevant in heavily regulated industries such as finance and healthcare, or in data-intensive and data-sensitive industries such as the automotive or pharmaceutical industries, where compliance with the GDPR and other comprehensive regulations pose significant obstacles to the widespread use of AI.

TBL: Keyword “AI governance”: How loud is the call for legal regulation of AI? To what extent do you think there should be laws and regulations regulating AI (keywords “ethics”, “security” and “data protection”?
Hildesheim:The principle of “man with machine” is one of the cornerstones of the discussion on “ethics in artificial intelligence”. The High Level Expert Group used by the European Union on the subject of “AI” published one of the world’s first government-initiated guidelines for the development and implementation of ethics in AI at the beginning of April 2019. A white paper on the topic of “AI” is currently being prepared, which also addresses the risks of technology. The topic of “ethics” is of course also reflected in the work of the Bundestag’s “Artificial Intelligence” commission. IBM is represented both in the expert committee that advises the EU Commission and in the Enquete Commission of the Bundestag. It is particularly important to us as a technology company that not only the political, but also the ethical framework are addressed. Because if we bring new technologies – here AI – to market maturity, we also have to create the conditions for the responsible use of these technologies.

We share this goal of responsible and public service-oriented development and use of AI with the political authorities and pursue this principle as a separate corporate principle. We already published guidelines on this in 2017 – see, for example, “Transparency and Trust in the Cognitive Era” or “IBM’s Principles for Trust and Transparency”. Our view is that AI solutions – to complement people – always serve a specific purpose and that purpose should also be made transparent. Furthermore, it is essential for the professional use of AI that the data used – if not publicly available – belong to its originator, ie our customers and users, as well as the knowledge gained from it. And it has to be ensured as much as possible.

I myself am a member of the steering group for the development of a “Standardization Roadmap AI” for Germany, which on behalf of the Federal Government details the ethical, technical and economic framework conditions in order to lead Germany successfully into the 21st century in the area of ​​”AI”.

TBL: How important are you to the AI ​​observatory, which is currently being set up, to be opened in spring 2020 at the Federal Ministry of Labor (BMAS)?
Hildesheim: We welcome the fact that the Ministry of Labor is also concerned with the use of AI in the world of work. Together with Verdi and the University of Maastricht, we initiated a study that deals with the topic. The study is also supported by the BMAS. There is definitely a lot in common when it comes to the topic of “AI and the world of work”, which is why we like to bring our experience in the field to the AI ​​observatory.

TBL: Why is AI important these days and what should it be able to do in 2020?
Hildesheim: The core competencies of AI technologies are speech recognition and dialogue skills. In addition, they have the ability to process enormous amounts of structured and unstructured data such as images or written records, thereby recognizing patterns and establishing correlations. On the basis of these properties, these technologies can provide information and advice for the optimization of production processes, maintenance and repairs, they can sound the alarm when specific machine problems occur, they can find errors or they can also detect and ward off attacks from cyberspace.