Switzerland is a distinctly federalist country with a pronounced division of powers. Therefore, technical innovations in the public sector are often first developed in the cantons.
One example of this is the introduction of an electronic identity (eID). At the federal level, the legislative process required to introduce the eID has not yet been completed, whereas an officially confirmed electronic identity is already in operation in one canton. In 2017, and as part of Switzerland’s cantonal eGovernment strategy, Schaffhausen canton became the first to introduce a digital identity for residents. Using this eID, citizens can apply online for a fishing permit, calculate tax liabilities of a real estate profit or a capital statement, or request a tax declaration extension, among other things.
Also, an adult account can be set up at the Child and Adult Protection Authority, and doctors can apply for credit for patients who are hospitalized outside their district. Another example, that began as part of a pilot project in the same canton in September 2019, enables residents to order extracts (via smartphone) from the debt collection register. These services are in a constant state of expansion (Schaffhauser 2020). Although the eID itself is not an ADM process, it is an essential prerequisite for access to digital government services, and therefore, could also facilitate access to automated procedures, e.g., in the tax field. The fact that a single canton has progressed further down this path than the Federal Government is typical for Switzerland.
Direct democracy is another defining element of the Swiss state. For example, the legislative process for a national eID has not yet been completed because a referendum is going to be held on the corresponding parliamentary bill (eID – Referendum 2020). Those who have asked for the referendum do not fundamentally oppose an official eID, but they want to prevent private companies from issuing the eID and managing sensitive private data.
Another element that must be taken into account is the good economic situation in Switzerland. This allows great progress to be made in individual areas, such as automated decisions used in medicine, and in many areas of research. Although there is no central AI or ADM strategy in Switzerland, due to the distinct federal structure and the departmental division of responsibilities at the federal level, sectoral research is conducted at a globally competitive level.
A catalog of ADM cases
Cancer diagnoses and treatments
At the moment, Switzerland is investigating the use of automated decisions in medicine, which is why ADM has been developed further in the healthcare sector than in other domains. Today, more than 200 different types of cancer are known and almost 120 drugs are available to treat them. Every year, numerous cancer diagnoses are made, and, as each tumor has its own particular profile with gene mutations that help the tumor grow, this creates problems for oncologists. However, once they have made the diagnosis and determined the possible gene mutation, they have to study the ever-growing medical literature in order to select the most effective treatment.
This is why the Geneva University Hospitals are the first hospitals in Europe to use the IBM Watson Health tool, Watson for Genomics®, to better find therapeutic options and suggest treatment for cancer patients. The doctors still examine the gene mutations and describe where and how many of them occur, but Watson for Genomics® can use this information to search a database of about three million publications. The program then creates a report classifying genetic alterations in the patient’s tumor and providing associated relevant therapies and clinical trials.
Until now, cancer doctors have had to do this work themselves – with the risk that they might overlook a possible treatment method. Now the computer program can take over the research, but oncologists still have to carefully check the literature list that the program produces, and then they can decide on a treatment method. As a result, Watson for Genomics® saves a great deal of time in analysis and provides important additional information. In Geneva, the report from this ADM-tool is used during the preparation of the Tumor Board, where physicians take note of the treatments proposed by Watson for Genomics® and discuss them in plenary to jointly develop a treatment strategy for each patient (Schwerzmann/Arroyo 2019).
ADM is also in use at the University Hospital of Zurich, as it is particularly suitable for repetitive tasks, predominantly in radiology, and pathology, and, therefore, it is used to calculate breast density. During mammography, a computer algorithm automatically analyzes X-ray images and classifies the breast tissue into A, B, C or D (an internationally recognized grid for risk analysis). By analyzing risk based on breast density, the algorithm greatly assists physicians in assessing a patient’s breast cancer risk, since breast density is one of the most important risk factors in breast cancer. This use of ADM to analyze medical images is now standard practice at the University Hospital of Zurich. Furthermore, research into advanced algorithms for the interpretation of ultrasound images is ongoing (Lindner 2019).
Having said this, more than one-third of breast cancers are overlooked in mammography screening examinations, which is why research is being carried out on how ADM can support the interpretation of ultrasound (US) images. The interpretation of US breast images contrasts sharply with standard digital mammography – which is largely observer-dependent and requires well-trained and experienced radiologists. For this reason, a spin-off of the University Hospital of Zurich has researched how ADM may support and standardize US imaging. In doing so, the human decision process was simulated according to the breast imaging, reporting, and data system. This technique is highly accurate, and, in the future, this algorithm may be used to mimic human decision-making, and become the standard for the detection, highlighting, and classification of US breast lesions (Ciritisis a.o. 2019 p. 5458–5468).
Chatbot at the Social Insurance Institution
To simplify, and support administrative communication, certain cantons also use so-called chatbots. In particular, a chatbot that was tested in 2018 at the “Sozialversicherungsansalt des Kantons St. Gallens” (Social Insurance Institution in the canton of St. Gallen, SVA St. Gallen). The SVA St. Gallen is a center of excellence for all kinds of social insurance, including a premium reduction for health insurance. Health insurance is compulsory in Switzerland and covers every resident in the event of illness, maternity, and accidents, and offers everyone the same range of benefits. It is funded by the contributions (premiums) of citizens. The premiums vary according to the insurer, and depend on an individual’s place of residence, type of insurance needed, and it is not related to income level. However, thanks to subsidies from the cantons (premium reduction), citizens on a low income, children, and young adults in full-time education or training, often pay reduced premiums. The cantons decide who is entitled to a reduction (FOPH 2020).
Towards the end of each year, the SVA St. Gallen receives approximately 80,000 applications for premium reductions. To reduce the workload associated with this concentrated flood of requests, they tested a chatbot via Facebook Messenger. The object of this pilot project was to offer customers an alternative method of communication. The first digital administrative assistant was designed to provide customers with automatic answers to the most important questions regarding premium reductions. For example: what are premium reductions and how can a claim be made? Can I claim a premium reduction? Are there any special cases, and how should I proceed? How is premium reduction calculated and paid out? Also, if it was indicated, the chatbot could refer customers to other services offered by the SVA St. Gallen, including the premium reduction calculator and the interactive registration form. While the chatbot does not make the final decision to grant premium reductions, it can still reduce the number of requests as it can inform unauthorized citizens of the likely failure of their request. It also performs a significant role in disseminating information (Ringeisen/Bertolosi-Lehr/Demaj 2018 S.51-65).
Due to the positive feedback from this first test run, the chatbot was integrated into the SVA St. Gallen’s website in 2019 and there is a plan to gradually expand the chatbot to include other insurance products covered by the SVA St. Gallen. It is possible that the chatbot will also be used for services related to contributions to Old-Age and Survivor’s Insurance, Disability Insurance, and Income Compensation Insurance (IPV-Chatbot 2020).
The Swiss Execution of Penal Sentences and Justice is based on a system of levels. According to this system, inmates are generally granted increasing amounts of freedom as the duration of their imprisonment continues. This makes it a collaborative process between executive authorities, penal institutions, therapy providers, and probation services. Of course, the risk of escape and recidivism are decisive factors when it comes to granting these greater freedoms.
In recent years, and in response to convicted felons committing several tragic acts of violence and sex offenses, the ROS (Risk-Oriented Sanctioning) was introduced. The primary objective of ROS is to prevent recidivism by harmonizing the execution of sentences and measures across the various levels of enforcement with a consistent focus on recidivism prevention and reintegration. ROS divided the work with offenders into four stages: triage, assessment, planning, and progress. During triage, cases are classified according to their need for recidivism risk assessment. Based on this classification, a differentiated individual case analysis is carried out during the assessment stage. During the planning stage, these results are developed into an individual execution plan for the sanction of the corresponding offender, which is continuously reviewed during the progress stage (ROSNET 2020).
Triage plays a decisive role at the beginning of this process – both for the offender and in terms of ADM, as it is performed by an ADM-tool called the Fall-Screening-Tool (Case Screening Tool, FaST). FaST automatically divides all cases into classes A, B, and C. Class A signifies that there is no need for assessment, class B equates to a general risk of further delinquency, and class C corresponds to the risk of violent or sexual delinquency.
This classification is determined by using criminal records and is based on general statistical risk factors including age, violent offenses committed before the age of 18, youth attorney entries, number of previous convictions, offense category, sentences, polymorphic delinquency, offense-free time after release, and domestic violence. If risk factors are met that, according to scientific findings, have a specific connection with violent or sexual offenses, then a C classification is made. If risk factors are met that have a specific connection with general delinquency, then a B classification is applied. If no risk factors, or hardly any, are found, then an A classification is made. Therefore, the classification consists of items (risk factors) in a closed answer format, each of which has a different weighting (points). If a risk factor is given, the relevant points are included in the total value. For the overall result, the weighted and confirmed items are added to a value, leading to the A, B or C classification that, in turn, serves as a basis to decide if further assessment is necessary (stage 2).
This classification is carried out fully automatically by the ADM-application. However, it is important to note that this is not a risk analysis, but a way of screening out the cases with increased assessment needs (Treuhardt/Kröger 2018 p. 24-32).
Nevertheless, the triage classification has an effect on how those responsible at a particular institution make decisions and which assessments are to be made. This also determines the offender’s so-called ‘problem profile’ regarding how the execution of sentences and measures are planned (stage 3). In particular, this planning defines any possible facilitation of enforcement, such as open enforcement, day release employment, or external accommodation. Furthermore, no ADM applications are apparent in the other stages of ROS. FaST is, therefore, only used during the triage stage.
In some cantons, in particular in Basel-Landschaft, Aargau, and Zurich, the police use software to help prevent criminal offenses. They rely on the commercial software package “PRECOBS” (Pre-Crime Observation System), which is solely used for the prognosis of domestic burglaries. This relatively common crime has been well researched scientifically, and police authorities usually have a solid database regarding the spatial and temporal distribution of burglaries as well as crime characteristics. Furthermore, these offenses indicate a professional perpetrator and thus show an above-average probability of subsequent offenses. In addition, corresponding prognosis models can be created using relatively few data points. PRECOBS is, therefore, based on the assumption that burglars strike several times within a short period if they have already been successful in a certain location.
The software is used to search for certain patterns in the police reports on burglaries, such as how the perpetrators proceed and when and where they strike. Subsequently, PRECOBS creates a forecast for areas where there is an increased risk of burglary in the next 72 hours. Thereupon, the police send targeted patrols to the area. PRECOBS thus generates forecasts on the basis of primarily entered decisions and it does not use machine learning methods. Although there are plans to extend PRECOBS in the future to include other offenses (such as car theft or pickpocketing) and consequently create new functionalities, it should be noted that the use of predictive policing in Switzerland is currently limited to a relatively small and clearly defined area of preventive police work (Blur 2017, Leese 2018 p. 57-72).
At the federal level, ADM is expected to be particularly present at the Federal Customs Administration (FCA), since this department is already highly automated. Accordingly, the assessment of customs declarations is largely determined electronically. The assessment procedure can be divided into four steps: summary examination procedure, acceptance of a customs declaration, verification, and inspection, followed by an assessment decision.
The summary examination procedure represents a plausibility check and is carried out directly by the system used in the case of electronic customs declarations. Once the electronic plausibility check has been completed, the data processing system automatically adds the acceptance date and time to the electronic customs declaration, meaning the customs declaration has been accepted. Up to this point, the procedure runs without any human intervention by the authorities.
However, the customs office may subsequently carry out a full or random inspection and verification of the declared goods. To this end, the computerized system carries out a selection based on a risk analysis. The final stage of the procedure is when the assessment decision is issued. It is not known whether or not this assessment decision can already be issued without any human intervention. However, the DaziT program will clarify this uncertainty.
The DaziT program is a federal measure to digitize all customs processes by 2026 to simplify and accelerate border crossings. Border authorities’ customer relations relating to the movement of goods and people will be fundamentally redesigned. Customers who behave correctly should be able to complete their formalities digitally, independent of time and place. While the exact implementation of the DaziT program is still at the planning stage, the revision of the Customs Act that correlates to DaziT is included in the revision of the Federal Act on Data Protection.
This is explained in more detail below, and should serve to clarify the previously mentioned uncertainty regarding the automated customs assessment procedure: In the future, the FCA will also be explicitly entitled to issue fully automated customs assessments, meaning there will be no human intervention for the entire customs clearance procedure. Thus, the determination of the customs duty will be decided fully automatically. In contrast, human contact is intended to be concentrated only on checking suspicious goods and people (EZV 2020).
Accident and Military Insurance
Throughout the revision of the Data Protection Act (explained in more detail below), it was decided that the accident and military insurance companies will be entitled to automatically process personal data. It is not clear what automated activities the insurance companies intend to use in the future. However, they may, for example, use algorithms to evaluate policyholder’s medical reports. Through this fully automated system, premiums could be calculated, and benefit claim decisions made and coordinated with other social benefits. It is planned that these bodies will be authorized to issue automated decisions.
Automatic Vehicle Recognition
In recent years, both politicians and the public have become concerned with the use of automatic systems, such as cameras that capture vehicle license plates, read them using optical character recognition, and compare them with a database. This technology can be used for various purposes, but at the moment in Switzerland it is only used to a limited extent. At the federal level, the system for automatic vehicle detection and traffic monitoring is only used as a tactical tool depending on the situation and risk assessments, as well as economic considerations, and only at state borders (parlament.ch 2020). The half-canton of Basel-Landschaft has enacted a legal basis for the automatic recording of vehicle registration plates and alignment with corresponding databases (EJPD 2019).
Allocation of primary school pupils
Another algorithm that has been developed, but is not yet in use, is designed to allocate primary school pupils. International studies indicate that social and ethnic segregation between urban schools is increasing. This is problematic, as the social and ethnic composition of schools has a demonstrable effect on the performance of pupils, regardless of their background. In no other OECD country are these so-called ‘compositional effects’ as pronounced as in Switzerland. The different composition of schools is mainly due to segregation between residential quarters and the corresponding school catchment areas. Therefore, the Centre for Democracy Aarau proposed not only to mix pupils according to their social and linguistic origin but also when defining catchment areas, so that the highest possible level of mixing between schools can be achieved. To optimize this process, a novel, detailed algorithm was developed that could be used in the future to support school allocation and classroom planning. The algorithm was trained to reconstruct the school catchment areas and to survey the social composition at individual schools using the census data of first to third graders in the canton of Zurich. Traffic load data, the network of pavements and footpaths, underpasses and overpasses, were also taken into account. This data could be used to calculate where pupils need to be placed to mix the classes more. At the same time, the capacity of school buildings will not be exceeded and the length of time spent traveling to school will remain reasonable (ZDA 2019).
Policy, oversight and public debate
Switzerland’s federal structure as the prevailing circumstance
When reporting on policy in Switzerland, the prevailing federal structure must be emphasized. It is a situation that has already been reflected upon in the previously mentioned ADM examples. Switzerland is a federal state, consisting of 26 highly autonomous Member States (cantons), which in turn grant their municipalities extensive leeway. As a result, the political and public debate on ADM depends greatly on the corresponding government, which cannot be described exhaustively in this report. Furthermore, this fragmentation on policy, regulation, and in research, introduces the risk of working in parallel on overlapping issues, which is also why the confederation strives for advanced coordination as stated below. However, the Federal Government has full responsibility over certain relevant legal fields and political governance, which legally binds all the governments in Switzerland, and thus impacts the entire population. Hence, set out below are those parts of the current federal political debate.
At the moment, the role of ADM in society, generally referred to as AI, is predominantly treated as part of a larger discussion on digitization. The Federal Government does not have a specific strategy concerning AI or ADM, but in recent years it launched a “Digital Switzerland Strategy”, where all aspects regarding AI will be integrated. More generally, the national legal framework concerning digitization will be adjusted simultaneously through the revision of the Federal Act on Data Protection (FADP).
In 2018, and against a background of increasing digitization beyond government services, the confederation launched the “Digital Switzerland Strategy”. One focus of this is on current developments in AI (BAKOM 2020). Responsible for the strategy, especially its coordination and implementation, is the “Interdepartmental Digital Switzerland Coordination Group” (Digital Switzerland ICG) with its management unit “Digital Switzerland Business Office” (Digital Switzerland 2020).
As part of the Digital Switzerland Strategy, the Federal Council set up a working group on the subject of AI and commissioned it to submit a report to the Federal Council on the challenges associated with AI. The report was acknowledged by the Federal Council in December 2019 (SBFI 2020). Alongside discussing the most central challenges of AI – those being traceability and systematic errors in data or algorithms – the report details a concrete need for action. It is recognized that all challenges, including this need for action, depend strongly on the subject area in question, which is why the report examined 17 subject areas in greater depth, such as AI in the fields of healthcare, administration, and justice (SBFI 2020 b).
In principle, the challenges posed by AI in Switzerland have, according to the report, already been largely recognized and addressed in various policy areas. Nevertheless, the interdepartmental report identifies a certain need for action which is why the Federal Council has decided on four measures: In the area of international law and on the use of AI in public opinion-forming and decision-making, additional reports will be commissioned for in-depth clarification. Further on, ways of improving coordination, relating to the use of AI in the Federal Administration, will be examined.
In particular, the creation of a competence network, with a special focus on technical aspects of the application of AI in the federal administration, will be examined. Finally, AI-relevant policy will be taken into account as an essential component of the “Digital Switzerland” strategy. In this context, the Federal Council has decided that interdepartmental work should continue and that strategic guidelines for the confederation should be developed by spring 2020 (SBFI 2020 c).
In addition, at its meeting on 13 May 2020, the Federal Council decided to create a national Data Science Competence Center. The Federal Statistical Office (FSO) will establish this interdisciplinary center on 1 January 2021. The new center will support the federal administration in implementing projects in the field of data science. To this end, the transfer of knowledge within the Federal Administration as well as the exchange with scientific circles, research institutes, and the bodies responsible for practical application will be promoted. In particular, the center of excellence will contribute to the production of transparent information while taking data protection into account. The reasoning behind the new center is backed up by a statement from the Federal Council, which said that data science is becoming increasingly important, not least in public administration. According to the Federal Council data science includes “intelligent” calculations (algorithms) so that certain complex tasks can be automated (Bundesrat 2020).
Since the Federal Data Protection Act is no longer up to date, due to rapid technological developments, the Federal Council intends to adapt the FADP to these changed technological and social conditions, and in particular, improve the transparency of data processing and strengthen the self-determination of data subjects over their data. At the same time, this total revision should allow Switzerland to ratify the revised Council of Europe Data Protection Convention ETS 108 and to adopt the Directive (EU) 680/2016 on data protection in the area of criminal prosecution, which it is obliged to do due to the Schengen Agreement. In addition, the revision should bring Swiss data protection legislation as a whole closer to the requirements of Regulation (EU) 2016/679 of the European Parliament and of the Council of 27 April 2016 on the protection of people regarding the processing of personal data and on the free movement of such data, and repealing Directive 95/46/EC (GDPR). The revision is currently being discussed in parliament (EJPD 2020).
While the total revision of the existing federal law on data protection will evidently revise the entire Act with all its various aspects, one new provision is of particular interest in regard to ADM. In the case of so-called “automated individual decisions”, there should be an obligation to inform the data subject if the decision has legal consequences or significant effects. The data subject may also request that the decision is reviewed by a human or that he or she should be informed of the logic on which the decision is based. Thereby, a differentiated regulation is provided for decisions by federal bodies. Correspondingly, even though federal bodies must also mark the automated individual decision as such, the possibility that the data subject can request a review by a human may be restricted by other federal laws. In contrast to the EU’s GDPR, there is neither a prohibition of an automated decision, nor is there a claim not to be subject to such a decision (SBFI 2019).
Civil Society, Academia and other Organizations
In addition, a number of fora in Switzerland research, discuss, and work on digital transformation and its opportunities, challenges, needs, and ethics. Most of them address this on a general level, however some address ADM or AI specifically.
Switzerland has a number of well-known and long-established research centers regarding AI technology. These include the Swiss AI Lab IDSIA in Lugano (SUPSI 2020) and the IDIAP Research Institute in Martigny (Idiap 2020) as well as the research centers of the Swiss Federal Institute of Technology in Lausanne (EPFL)(EPFL 2020), and Zurich (EPFL)(EPFL 2020). In addition, private initiatives such as the Swiss Group of Artificial Intelligence and Cognitive Science (SGAICO) complement the academic initiatives, by bringing together researchers and users, promoting knowledge transfer, confidence building, and interdisciplinarity (SGAICO 2020).
Government Research Funding
The confederation also addresses the topic of AI via research funding. For example, the Federal Government invests in two national research programs via the Swiss National Science Foundation (SNSF) (SNF 2020). Firstly, the National Research Programme 77 “Digital Transformation” (NRP 77) (NRP77 2020) and, secondly, the National Research Programme 75 “Big Data” (NRP 75) (NRP 75 2020). The former examines the interdependencies and concrete effects of digital transformation in Switzerland, and focuses on education and learning, ethics, trustworthiness, governance, the economy, and the labor market (NFP 77 2020). The latter aims to provide the scientific basis for an effective and appropriate use of large amounts of data. Accordingly, the research projects examine questions of the social impact of information technology and address concrete applications (SNF 2020).
Another institute working in this area is the Foundation for Technology Assessment (TA-Swiss). TA-Swiss is a center of excellence of the Swiss Academies of Arts and Sciences, whose mandate is laid down in the Federal Law on Research. It is an advisory body, financed by the public sector, and it has commissioned various studies on AI. The most pertinent of these is a study published on 15 April 2020 on the use of AI in different areas (consumption, work, education, research, media, public administration, and justice). According to the study, a separate law on the use of AI is not considered to be effective. Nevertheless, citizens, consumers, and employees in their dealings with the state, companies or their employer should be informed as transparently as possible about the use of AI. When public institutions or companies use AI they should do so according to clear rules, in an understandable and transparent manner (Christen, M. et al. 2020).
Digital Society Initiative
The Digital Society Initiative was launched in 2016. It is a center of excellence at the University of Zurich for critical reflection on all aspects of the digital society. Its goal is to reflect on and help shape the digitization of society, democracy, science, communication, and the economy. In addition, it aims to critically reflect and shape the current change in thinking brought about by digitization in a future-oriented manner and to position the University of Zurich nationally and internationally as a center of excellence for the critical reflection of all aspects of digital society (UZH 2020).
The Digitale Gesellschaft (Digital Society) is a non-profit society and broad-based association for citizen and consumer protection in the digital age. Since 2011, it has been working as a civil society organization for a sustainable, democratic and free public sphere and it aims to defend fundamental rights in a digitally networked world (Digitale Gesellschaft 2020).
Several other Swiss organizations are also worth a mention. These organizations concentrate on digitization in general, especially in an economic context, e.g., Swiss Telecommunication Association (asut 2020), digitalswitzerland (Castle 2020), Swiss Data Alliance, and Swiss Fintech Innovations.
ADM is used in various parts of the public sector in Switzerland, but these tend not to be in a centralized or comprehensive manner. Only a few cantons use ADM in police work, for example, and the systems used vary. The advantage of such an approach is that cantons or the federal government can benefit from the experience of other cantons. The drawback is that efficiency losses may occur.
There are selective legal foundations, but no uniform ADM law or e-government law or anything similar. Neither is there a specific AI or ADM strategy, but recently attention has been paid to greater coordination, both between different departments at the federal level and between the Federal Government and the cantons. Machine learning methods are not used in state activities in the narrower sense, e.g., in police work or the criminal justice system, as far as can be seen.
Also, at that same level, ADM is used or discussed selectively, but not in a comprehensive manner. In the wider public sector, ADM is used more often and more widely. A good example is a deployment in the Swiss health system, where the Geneva University Hospital became the first hospital in Europe to use ADM to suggest treatments for cancer patients.