ÖFIT sets the course for a digital future worth living in politics and administration.

^{Prof. Dr. Peter Parycek © ÖFIT}

AI – a question of definition

The complexity of the topic begins with the definition: at the beginning of the research was rule-based AI, in which reality was precisely described and rules were mapped with the help of high-quality data sets. "This brings us to one of Germany's great weaknesses", knows Peter Parycek, "due to size and the data protection discourses of recent decades, there are few Germany-wide registers that show the approach to the truth. The lack of data makes comparisons with small countries such as Austria or Estonia, in which applications for family allowances or tax returns are automated, pointless. Today, however, there is less talk of automation in AI. "At present, technically we mean above all machine learning, in other words algorithms which learn decision patterns from existing data that they can apply to further cases," explains research assistant Basanta Thapa, “there are only pilot projects in existence in administrations in Germany.” For example, facial recognition at the Berlin Südkreuz station, where surveillance cameras can be used to identify wanted crimi­nals. With "Bobbi", a browser-based chatbot is also in use in Berlin: instead of having to go to the Bürgeramt (citizens’ administrative office), citizens can contact the programme with ques­tions. It has even helped with concerns about SARS-CoV2 and COVID-19. While ”Bobbi“ is still in its infancy, 80 per cent of all inquiries in China during the pandemic were already handled by a chatbot. "To manage systems as large as those in China, but also in India, an enormous use of data and algorithms is required," says Peter Parycek, making this lead in AI applications easy to explain, "we live in different dimensions in Ger­many. We are not Estonia, but compared to China, our cities are villages.”

Fear of the surveillance state

The application of machine learning does not necessarily fail because of technology. Rather, it is a question of (political) necessity and social objective. When high-precision facial recognition systems detect persons who are travelling without a face mask, this is part of everyday coronavirus life in China. In Germany, such an application would hardly be met with approval. It is accepted that Amazon's Alexa controls our everyday life or that Apple relies on facial recognition to unlock the smartphone in the latest iPhone generation. However, "In the public sector, on the other hand, artificial intelligence is discussed in a highly emotional way," Prof. Dr. Parycek knows one of the challenges in this segment, "this is about the sensitive relationship between society and the state." Surveys con­ducted by the ÖFIT show how sensitive this is: Germans see a lot of potential in digital public admini­stration and regularly attest the authorities a high level of satisfaction and trust in their work. But when it comes to handling digital data, that's quickly over: only one quarter of the population trusts the public sector. "Platforms like Facebook were far ahead of the administration," the ÖFIT director was surprised. 79 per cent feel defenceless in the face of unfair decisions; more than half of those surveyed also doubt compliance with data protec­tion and IT security regulations. "In Germany there is great unease about the administra­tion helping to become a surveillance state," Parycek says, naming the main reason for the mistrust, "this fear is even stronger in Germany due to recent history with the GDR.” The queasy feeling is intensified when it comes to the use of artificial intelli­gence: while 46 per cent of citizens see AI as a possible enrichment, almost one-quar­ter perceive it rather as a threat. According to another survey, it is important for 86 per cent whether a human or a machine makes decisions. If the latter is the case, nine out of ten people feel alienated; eight out of ten feel lost. In view of these results, it is hardly surprising that almost two-thirds of the respondents (65 per cent) think that AI should only be used in the administration in a supportive way.

Potential of comprehensible decision-making support

What such decision-making support can look like in practice is shown in the Berlin State Administration Office. There, submitted invoices for aid payments are already being scanned with intelligent software from the US company IBM. If anything abnormal occurs, the system sounds the alarm. "AI analyzes data, on the basis of which administrative or political decisions can then be made more quickly," explains Peter Parycek, who has headed ÖFIT since 2017, "these processes are often not visible because they run in the background.” For him, the greatest potential of machine learning in the public sector lies in this decision-making support. AI could also help case workers to make a better assess­ment of applications for building or housing grants. But for this it is necessary to empower employees to carry out this control function. "Not everyone needs to be trained as a data scientist," the expert placates. But the employee must be able to understand which data the machine is using and where a distortion could lead to a wrong decision. Especially in public administration, the demands on AI systems are high: while the economy does not have to treat all customers equally, but can segment and concentrate on the 80 per cent with the highest profit margin, the state cannot allow itself to do so. In order to prevent bias in the best possible way, the design of the work system and performance measurement in the public sector should be questioned. "If the case worker is not measured by the number of cases per day, but by the quality of the decisions, which are measurable by way of appeals, then we will achieve a good man-machine interaction”, he is convinced.

Design task of politics

If artificial intelligence is used in a comprehensible manner, this could also have a positive effect on acceptance by the public. "Explainable artificial intelligence is a challenge for AI that learns from data," admits Basanta Thapa, "but we need it in an administrative con­text.” Instead of full traceability, it would also be possible to show applicants what would be required for approval in the event of rejection. It would also be conceivable to let the citi­zen decide between AI and a human: an immediate, perhaps more favourable decision by the machine or wait three months until a clerk can take care of it. "My theory is: if citizens see an advantage for themselves, for example, it is more convenient and they do not have to sub­mit a 40-page application, but the data are evaluated automatically, then they are less criti­cal," says the ÖFIT scientist, who sees this as one of the design tasks of politics: "Initial applications should make people's lives better and more comfortable. Surveillance cam­eras that catch parking offenders are not applications that win people's hearts and minds." A pet finder service à la FindeFifi app, on the other hand, can easily do the latter.