1. Introduction

AI systems are rapidly developing in public administrations, and we believe that the first systems to be implemented and proposed for use by the public can give precious insight on trends that might generalise in the development of such systems. In this sense, chatbots based on LLMs are a perfect terrain to study, as “pioneer” project used to test such technologies. We believe that the way authorities act, communicate and try to regulate such novel AI systems can have an influence on the public sentiment around these technologies in the trust in the authorities’ actions. Thus, studying the content of the texts supposed to inform the users of such AI systems about their rights allows us to both get insight into the authorities’ concrete view of these technologies and analyse their possible impact on public trust^[1].

In this sense, this study doesn’t approach the technical aspects of the reliability of the AI systems implemented^[2], nor conducts a psychological study of the perception by the users of these systems^[3]. We chose to concentrate on one type of “cue”^[4] of the trustworthiness of public administration AI that are the rules the authorities communicate to the people using their chatbots. To provide a practice-oriented analysis, we decided to limit the scope of this case study to terms and conditions of chatbots present on Swiss public law institutions. These texts are a tool used by the authorities to communicate about the degree accountability^[5] – or in this case lack thereof – they are willing to assume. Although similarities can be found in other countries that implemented similar texts with similar contents, we had to limit ourselves to a national context for practical reasons. This paper analyses the terms and conditions of selected public chatbots and AI systems. Switzerland being a federal system with a high degree of decentralisation, three levels of state coexist, and we searched for examples from each level of Government and each linguistic region. Since there is no definitive list of existing systems^[6], we based our analysis on systems either present on these curated lists, that had had media coverage or that were present in the literature. Since our study focuses on the use by the authorities of such systems, we chose chatbots implemented directly on the homepage of an administration, or the ones present on a page of an administrative body. Since our focus is on the trust the citizens should place in the use of AI by the authorities, we chose only publicly available chatbots^[7]. For our analysis be rooted in public law, we chose systems that were implemented by the authorities on their official webpage or that were reachable by means advertised on their websites^[8]. These systems are in this sense illustrative of a common trend both in the use of such terms and conditions as well as in their wording.

The broader subject of “trust in the public authorities” concern the three levels of government, each in their own sphere of responsibility. For the public, however, the subject isn’t as clear, digitisation of state has been a major factor for the development of “one stop government” that tends to blur the line between the areas of competencies of the different authorities.

This paper chose to focus on the disclaimers present in these terms and conditions, these exclusions of responsibility that wish to exonerate the authorities for the information given. We primarily focus on the rules pertaining to data protection that have already been studied^[9]. We will approach the way these data protection rules differ from the rules surrounding the content produced by the chatbots and the influence of these differences on public trust.

The systems implemented by public administration are constantly evolving, the nature of the terms of conditions allow for their evolution which further complicates the analysis. For example between the beginning of our study in September 2025 and the completion thereof in march 2026, one website underwent a complete overall in January and the chatbot is now unavailable and points to an error^[10] on the webpage of the authority advertising it^[11]. In this sense, our study is example based and qualitative in nature. We argue that the relatively large scope of entities considered, on all level of government, from all regions, allow us to show a trend in the way authorities across Switzerland implement terms and conditions to try to limit their responsibility.

2. Trust

2.1. A definition of Trust

Trust is one of the building blocks of society and law. Each discipline studying trust have different approaches and views the definition of trust, and there still doesn’t seem to be a definitive consensus across scientific fields about this notion^[12].

Trust is often defined as the «willingness of a party to be vulnerable to the actions of another party based on the expectation that the other will perform a particular action important to the trustor, irrespective of the ability to monitor or control that other party»^[13].

This makes trust something that is perceived by people, is not necessarily rational and is influenced by the interaction, context and the relationship between the parties^[14]. The nature of trust places it outside of the direct control of an entity. It is not possible to command trust, but only to behave in such a way to generate trust in one’s counterpart. By behaving in a trustworthy manner, an actor might be able to inspire trust.

From a legal standpoint, the need for trust is both a requirement in the relationship between the actors, and a core principle underlying the whole legal environment^[15]. For rules to be effective there needs to be a shared belief that these rules will be respected and enforced.

2.2. The need to act and to be perceived as trustworthy when providing public services

The relationship between the individual and the authorities is asymmetric and there is a definite verticality and power relationship between the individuals and the authorities^[16]. The individuals have administrative obligations towards the authorities, while authorities are bound by the rule of law and administrative principles^[17].

But trust isn’t outside of the equation. By behaving in a trustworthy manner, the authorities will gain in legitimacy^[18]. The perceived trustworthiness of the authorities might have an influence on the behaviour of the citizens. The development of good governance principles^[19] often refers to the trustworthiness of the authorities to increase their legitimacy or as a justification for the implementation of certain legislation or principles.

Although this assumption is challenged by the literature, transparency of the authorities is considered a «key value for trustworthy governments»^[20]. To create trust in the institutions, empirical studies conducted have mostly focused their attention on the right to ask for transparency the rights to demand access to government documents. These studies couldn’t find a definitive impact of such “rights for transparency” on the trust the individuals put into the government, the successful interaction with a government agency having more of an effect^[21]. Some studies have even found that the knowledge of the existence of a latent right to transparency, knowing one has a right to ask for information from the state, wouldn’t necessarily produce higher trust in the government itself^[22].

Since trust is a building block of constitutional democracies and a big factor in the legitimacy of the government^[23]. Every actor is therefore required to act in good faith^[24] (art. 5 al. 3 Cst.^[25]). Which is then specified by art. 9 Cst. «Every person has the right to be treated by state authorities in good faith and in a non-arbitrary manner». In the German version, this good faith element is presented as a «duty to be trustworthy»^[26]. This prerequisite is also enshrined in the art. 2 of the Civil code^[27] for the relationships between individuals and enterprises.

2.3. The probable impact of digital transformation and AI systems implementation on trust

The implementation of new eGovernment services by the authorities presents a “new” challenge for trust. Indeed, since digital transformation is mostly an iterative process, most digitisation processes go through a trial period before transitioning entirely, and suppressing existing analog or pre-existing digital systems. In this period e their adoption by users might have a particular impact on their future adoption. In the first phase of their implementation, the users need to make the decision to “try out” the new system, to adopt a new way to do a process and transition from a mostly analog process to the new digital one.

For this reason, the public discourse often presents trust as being a core component of the acceptance of eGovernment systems^[28]. Trust also appears as a central principle in strategies on the implementation of AI systems^[29]. Studies have focused on the influence of trust of the citizens in the State for the uptake of such systems^[30], on the trust in the technologies themselves^[31] or even the trust in the use governments would do of these technologies^[32]. Legal studies have discussed the concept of trustworthy AI, its origin and its influence on AI regulation^[33]. Most of these studies seem to have found a positive effect of trust in the adoption of the eGovernment services^[34].

The importance of trust in the implementation of digital technologies is a developing field of study. Some focus on the changes AI could represent in the definition of trust itself^[35], while other research ways to implement certain pieces of technology to foster trust^[36]. While these approaches are either technical or philosophical, our study concentrates on the legal framework and the way the authorities communicate about such systems and the influence these texts could have on trust. We hope in this sense that our paper, coming from a legal background, will inspire more experimental studies for researcher focusing on the philosophical and psychological aspects of trust in eGovernment and AI.

3. The implementation of AI systems by the Swiss authorities

3.1 A quick overview of the adoption of AI systems by public administrations

The evolution of technologies and their adoption by private companies is encouraging public authorities to follow suit to «keep up with the times»^[37]. The adoption of technology is in this sense not a new process^[38], the development of internet and more recently of AI technology empowered Swiss authorities to start a digitisation process of their internal administrative processes wishing for such processes to be «digital first»^[39]. This modernisation isn’t based on competitive mechanisms, as it would be the case in the private sector^[40]. ICT is often seen by public authorities as a tool for achieving effectiveness gains and efficiency gains^[41]. By computerising data collection and processing, the state can provide better services at a lower cost^[42]. The automation of certain computerised processes constitutes a logical next step in this quest. Some researchers even consider that automation is the result of a fundamental dynamic in public law resulting from the desire to limit arbitrariness and to ensure that the law is applied as uniformly as possible, which would thus lead to the development of a “mechanical” law and hence lead to the automation of state processes^[43].

While AI is not yet widespread in the public sector, some sector-specific systems are being developed^[44]. Their adoption by the public authorities follows a similar logic of seeking effectiveness and efficiency^[45]. The hoped-for efficiency gains, and hence improvements in the quality of the public service, are generally aimed at improving output, that is, the service provided to the public. This improvement can be achieved by reducing the risk of errors, by improving service availability as well as the processing speed of cases. Improvements to the public service can also be achieved by changing the processes themselves, or by using more and better data^[46]. Numerous AI applications wish to simplify the work of the agents responsible for managing files within the administration^[47] or envision to help combat fiscal fraud or crime^[48]. Efficiency gains, and therefore potential savings, seem to be expected in payroll; by automating tasks, the authorities hope to be able to do more, and better, with the same amount or even fewer people; thus they hope to reduce the financial burden on the state budget. Few studies have been carried out to date to calculate the real cost-benefit ratio of the implementation of AI systems. This might be explained by the novelty of the phenomenon and the rather prospective aspects of this research^[49]. It is not unlikely that the high development costs of these specialised systems and the additional costs associated with maintaining IT systems might, in many cases, prevent authorities from achieving real efficiency gains.

3.2 The implementation of Chatbots in particular

Following the rapid uptake of generative AI on the market, many administrative bodies took upon themselves to “try out” the implementation of such systems by creating “chatbots” that would help better communicate with the public^[50]. The main reason for implementing chatbots in public administrations is to improve administrative efficiency^[51]. Some researchers believe that chatbots could represent an «exploratory application of AI in such organisations»^[52]. These types of systems are the most common type of artificial intelligence system implemented in public administrations in Switzerland^[53]. These systems facilitate access to government information, as these conversational systems are available to citizens 24 hours a day, whereas physical counters or telephone counters are only available during fixed hours, often office hours^[54]. These chatbots also improve access to information for the authorities by lowering the barriers of access. Indeed, these systems are directly accessible on the authorities’ web page and it is not necessary for a person looking for information to go through multiple steps to find it. The question of the quality of the information provided is open to debate. While citizens seem relatively ready to believe the information provided by these systems^[55], the systems implemented to date have several limitations partly due to the way LLMs operate^[56]. Indeed LLM’s have, for the moment at least, technical limitations and are prone to “hallucinations”, which in layman’s term means that AI might give out wrong or incomplete answers. We appreciate the term of “careless answers” to approach the notions of particularly wrong or dangerous answers^[57]. This notion allows for an adaptation of the threshold for exactitude depending on the sensitivity of the system. In this sense, public chatbots should be held to higher exactitude standards. It is still unclear if the technological development will really be able to guarantee the total absence of errors.

For chatbot systems, more than a public interest, the authorities have an obligation, based in public law, to inform the public about their activities^[58]. On the federal level for instance, the Government is supposed to «inform the general public fully and in good time about its activities» (art. 180 al. 2 Cst.^[59], implemented by the Government and Administration Organisation Act^[60] (art. 10), which embodies for the federal administration an active duty to be transparent. Similar laws were also adopted on cantonal level^[61].

The right to ask for access to documents, often described as a passive transparency principle, is prescribed on the federal level by the Federal Act on Freedom of Information in the Administration (FoIA)^[62]. The same types of obligations of transparency are implemented in cantonal laws^[63]. For the moment, there doesn’t seem to be any chatbots implemented to fulfil this type of transparency. Indeed, the procedural requirements are pretty narrow, and most current laws wouldn’t allow chatbots to fulfil the requirements needed to give access to specific documents.

Chatbots partly embody that “active” transparency principle, in the sense that they give a new way for the administration to fulfil their obligation of information towards the public. These chatbots might also have an organisational function that can be linked with every special procedure they support. The authorities have an obligation to inform the public about the way in which administrative proceedings can be conducted^[64], they might choose chatbots to do so. It might then be less of a generic “information” of the public, and more a tool used to fulfil an obligation resulting from the proceedings themselves. For example, the chatbot established by Basel land^[65] from unemployed people finds its origin in the general competence of the cantonal office for industry, trade and labour to execute the federal law on the matter^[66] (§1 AVLG-BL^[67]) and to operate contact, information and training centres for jobseekers (§11 AVLG-BL). Since this chatbot was implemented notably to inform the individuals about the proceedings needed to grant them a right for unemployment benefice, this technology is at the border between a general duty to actively inform the public and a specific obligation linked to a specific administrative procedure. If most chatbots discussed in this paper are closer to the general duty to inform, the development of AI and the adoption of such systems could represent a shift towards the development of chatbots as means to conduct administrative proceedings.

If multiple types of systems might have been called chatbots by the authorities, starting from mere interactive information panels without any user inputs, the majority of chatbots developed in the last few years use AI and LLMs to process textual inputs from the users and give “answers” to their questions.

The industry predicts the evolution of “virtual assistants” chatbots^[68], that would, based on a conversation with the user automate some real-life processes. We couldn’t find an example of such technologies being implemented on Swiss public websites to date^[69]. Some websites used forms that had components that in some ways are akin to interactive questionnaires, with an array of options to choose from^[70]. The way these forms operate forbids assimilating them to real chatbots. Moreover, such systems weren’t presented as “chatbots” by the authorities, presumably to differentiate between the conversational – and informational – agent and the form being part of an actual administrative proceeding. Although we didn’t find literature on the subject, it doesn’t seem improbable to see AI chatbots soon be implemented into, or maybe even totally replace forms and form like systems to conduct administrative proceedings.

The chatbots implemented by the authorities either use ChatGPT^[71] or LLMs implemented by Deeplink^[72] a swiss company that seems to have developed their own AI system. We also found one vocal assistant that is based on Spitch^[73].

4. The terms and conditions and their influence on Trust

4.1 The use of terms and conditions surrounding public chatbots in Switzerland

Terms and conditions are present either in the chatbot itself, for example in an information checkbox at the top of the chat window, with a message in the chat, or even with a link in the footnotes of the webpage giving access to the chatbot itself.

For example, the chatbot of the migration office of the canton Basel land^[74], has a message that asks the users to consent to the use of data with a link to a pdf document containing data protection information^[75]. Since the chatbot is on the canton’s website, their general terms and conditions^[76], available through a link at the bottom of each page named «disclaimer», also apply to the use of the chatbot. The chatbot of the local police in Morges, in Canton Vaud, only has a link at the bottom of the website named «terms and conditions», which contains both rules for the content and the data protection^[77]. The chatbot ALVA from Canton Basel-Stadt^[78] has an info box that gives a general disclaimer and a link to the «data protection rules»^[79] that also contain disclaimers about the content and rules for the use of the system and asks the user to accept them to start the chat. Some systems have different documents available in their interface, separating the “disclaimer” from the rules surrounding the data processed^[80]. The chatbot of the federal archives^[81] that starts the discussion by a short text informing the user about some data protection rules and a short disclaimer, with links to a pdf document with more detailed rules^[82], and a button asking the user to «accept» these principles. This pdf then sends the user to the data protection rules for the whole archive’s website^[83], which themselves send the user to the terms and conditions for the federal administration website^[84], thus layering the information in multiple interconnected documents. The chatbot of the Swiss army^[85], includes an info logo at the top of the windows that gives access to four pop-up windows, one giving general information about the chatbot, one about the use thereof – which actually contains information about the data provided and the suppression of identifying data and metadata – one titled «disclaimer»^[86], and one «legal notice»^[87].

If most chatbots are implemented on public authorities’ websites, with texts indicating that the services are provided by the authorities themselves, it is worth noting that the chatbot of the municipality of Hochdorf, canton Lucern refers to the chatbot present on their website^[88] as a service provided by a private company^[89] which seem to have set their own terms and conditions in place^[90].

It seems important to note that although most systems presented terms and conditions, we couldn’t find any such provisions for the Vocal assistant of the commune of Kreuzlingen. Although the system «is not permitted to provide legally binding information» according to its implementors^[91], nor the specialised administrative webpage indicating the contact number to the assistant^[92], nor the terms and conditions of the communal webpage^[93] give any information about this system. This is particularly striking considering how detailed these terms and conditions are, and the sensitive nature of the voice recording that are processed through Amazon Cloud Services^[94].

As we exemplified, two types of rules are presented or linked to these chatbots. First these texts contain information or rules concerning the data protection of the users. Second rules concerning the content that is published and the texts that will be provided by the chatbots themselves when “answering” questions and the possible responsibility of the authorities for these texts.

4.2 The content of the terms and conditions surrounding the use of such chatbots

Before the terms and conditions themselves, some chatbots inform their users about their use for example Alva «If you activate me, I will help you find the right information»^[95], the municipal police state, «I will assist you in finding information about our services» before listing different types of information they can provide, in particular «Getting help if you are experiencing domestic violence» or «Obtain a parking permit»^[96]. The Swiss army chatbot asks «How can I help you?»^[97] as a first question in the chat.

These introductory texts also make the user attentive to the limitations of the system; this is for example the case for the BarBot of the federal archives «It is possible that some information may be incorrect. Please check important information»^[98]. The Swiss army chatbot includes a warning at the bottom of the chatbot window: «[t]he Swiss Army chatbot may make mistakes. Please check any important information»^[99]. Alva states «[m]y answers are generated by artificial intelligence and are therefore not always correct»^[100], while the police chatbot indicates, «I am just starting out and learning every day»^[101]. The “HofdererBot” of the Municipality Hochdorf indicates at the bottom of the chatpage «[a]ll information provided without guarantee»^[102].

The Swiss army chatbot is the only one that clearly stated the experimental nature of the system: «[t]he current version is a prototype and is being used to carry out further user testing»^[103], and provided an email address to allow users to provide commentaries and feedbacks.

Some chatbots require the users to accept the terms and conditions by pressing a button titled «accept»^[104], «agreed»^[105] or «activate»^[106] before being able to use the system. While others don’t demand such an explicit consent^[107], the telephonic system of the city of Kreuzlingen doesn’t even inform the caller that the call is recorded.

The structure of these disclaimers varies widely between websites and services, but most of them starts with a general statement about the goal of the terms themselves^[108], or even with the disclaimers themselves^[109].

In comparison to the rules about data protection, the rules concerning the information and texts produced by the chatbot are scarce. If most chatbots include disclaimers or responsibility waivers, these texts sometimes include indications on the way to ask for human help or review^[110]. There are, however, no indications of any rights by the users concerning the content and texts produced by the users, that means that a right for a human review isn’t included, and that these texts don’t provide any possibility to get access to past conversations. In the same sense, if some data protection texts gave information about the laws governing their implementation^[111], the same isn’t true for the disclaimers and responsibility exemption clauses that gave no information about the public law that allowed such exceptions to be put in place.

The content of the disclaimers is always similar, the authorities reject all and every responsibility, both for the information provided, for any actions taken by the individuals based on the texts provided by the chatbots and even for any direct harm that the chatbots could cause.

The canton Basel-Stadt is particularly precautious, concerning Alva they state: «The Canton of Basel-Stadt does not guarantee the accuracy, completeness or timeliness of the information provided by Alva. The information is for general information purposes only and is not legally binding. We ask that you verify the information provided by Alva on the official websites of the canton»^[112], although the same terms and conditions, under disclaimer^[113] indicate that if they take care to «ensure that the information on its websites is accurate and up to date», the content of the website can’t be trusted: «Nevertheless, neither the canton nor individual offices or departments can guarantee the accuracy, reliability or completeness of the information»^[114]. They then include two disclaimers, one specific to Alva: «Alva is based on artificial intelligence and may therefore make mistakes or provide inaccurate information. In particular, answers may appear correct but be incorrect in terms of content. The Canton of Basel-Stadt accepts no liability for the consequences resulting from the use of the information provided by Alva»^[115], which completes the general disclaimer for the whole canton’s website: «The canton accepts no liability for damage or consequential damage resulting from access to its websites or individual parts thereof (such as downloaded documents), their use (or the inability to access or use them) or links to other websites»^[116].

The disclaimer for the chatbot of the federal archives is the most comprehensive we could find. After stating the goal of the chatbot and stating that the chatbot couldn’t replace personal advice, the disclaimer waives every responsibility for the system: «The use of the chatbot and any consequences thereof are your responsibility. The Federal Archives do not guarantee the truthfulness, accuracy, completeness or reliability of the answers provided by the robot. False, inaccurate or invented answers cannot be ruled out. You should always verify the information and sources provided. The Federal Archives are in no way responsible for actions taken by users or their decision not to act on the basis of information provided by the chatbot, nor for any consequences thereof. Any liability claims are expressly excluded»^[117].

The chatbot of the Swiss army, which intends to «serve[..] as a key point of contact for military personnel. It handles standard enquiries immediately and forwards more complex requests to the relevant departments»^[118], indicates in it’s disclaimer that that the system «was developed to provide general information and general details about the Swiss army»^[119]. It then went on to assert that this information doesn’t constitute personalised legal advice and that the users should seek advice towards the hotline of the army. This provision was then followed by the broadest disclaimer found to date: «Use of the chatbot is at your own risk and peril. The Swiss army accepts no liability for the correctness, accuracy, completeness and reliability of the answers provided. False, incorrect and fictitious answers must be expected. Please check each answer, including any sources cited. The Swiss army is not liable if the user takes or fails to take action in reliance on the correctness and completeness of the answers and suffers disadvantages or damages as a result. Any liability claims caused by the use of this website are expressly excluded»^[120]. A similar provision is still present on the terms and conditions of the canton Basel Land which starts with a disclaimer: «The canton makes every effort to ensure the secure operation of the website. Nevertheless, defects such as data loss or falsification, virus attacks, operational interruptions, etc. cannot be ruled out. Access to the website is therefore at your own risk and responsibility»^[121]. Which is then followed by a exclusion of responsibility clause «[t]he publications on the websites of the Canton of Basel-Landschaft are for informational purposes only. The Canton of Basel-Landschaft expressly reserves the right to change or delete all or part of the content at any time without notice, or to discontinue publication temporarily or permanently. Liability claims against the Canton of Basel-Landschaft for material or immaterial damage arising from access to or use or non-use of the published information or from technical faults are rejected»^[122].

In contrast, the municipal police of Morges terms of uses don’t contain a real “disclaimer” but informs the user on the unreliability and dangers of using the service: «[i]n addition, [the association] endeavours to provide accurate and up-to-date information on its website. However, it cannot guarantee the accuracy, reliability or completeness of the information contained therein or referred to therein. Furthermore, system failures such as interruptions, virus attacks or data loss cannot be ruled out»^[123]. The disclaimers convey multiple messages; first the unreliable nature of the information conveyed through the chatbot and the risk for the system to make errors and provide false advice^[124]. This indicates that the authorities are aware that the risks of the systems they implemented aren’t trivial, to the point the users should “expect” them.

Second these texts inform the users that, because of these errors, the users shouldn’t trust these texts and can’t use these texts to make decisions. Since these chatbots were implemented specifically to give information to the users, it is rather surprising to see such warnings surrounding their use. The disclaimers convey the message that the authorities consider the individuals to be the sole responsible for any damage resulting from the use of their systems^[125].

Third, the exclusion of responsibility clauses implies that the authorities explicitly try to exclude the right of the users to hold the authorities accountable for the information their system conveyed or the damage their systems could have caused. If the occurrence of such harms isn’t yet recorded on authorities’ chatbots, it is not entirely impossible for such a chatbot to defame someone or to create harm to the personality of the users or of public figures.

To make an analogy, these disclaimers are like warning signs that would be put up on a critical piece of public service. Such warning signs are often times needed, to warn that a walk way could be dangerous^[126], or an old mountain source be untested and thus, not rated for human consumption. But e would be shocked if the authorities spent money to provide more access to water on a public square, installed a new water fountain and publicised this event as a great development in the provision of such an essential public service, before putting warning signs saying that there are risks for people to poison themselves sometimes, and that each user should test the water before drinking. Information provided to the users by the administration has a real influence on their lives, missing a deadline or failing to fill the proper administrative paperwork could have real consequences for these persons. These authorities, by choosing to provide a “new” public service all while refusing to take any responsibility are at least sending the wrong signals about their trustworthiness and clearly underline the untrustworthiness of the technology they implement.

4.3 The legality of such disclaimers

In a private law systematic, where the website and the person consulting it are on an equal level, terms and conditions are generally considered to be of a contractual nature even if extra-contractual liability may arise from information available online^[127], these terms and conditions are part of the contract that is or will be passed between the user and the website to either access to content, to buy goods or services^[128]. With the recent developments of data protection legislation, most websites even have an explicit consent form for the individuals to accept the data processing made by the website and the use of cookies present on it. We should note that too broad disclaimers are generally forbidden in Swiss private law^[129] which forces private actors to set “reasonable” limitations in their responsibility.

The situation for public authorities is widely different. If a definitive analysis of this question is outside of the bound of the present study, we found no reasons for the relationship between the individuals and the authorities to be placed in the realm of private law, and thus no reason to consider these terms and conditions not to be of a public law nature.

The behaviour of the authorities when publishing or using software to provide public information and administrative services to the public are ruled by public law. Indeed, to determine if the matter is of a public or a private law nature, Swiss doctrine considers a multiplicity of criteria that, combined, assert the nature of the relationship and of the acts adopted in it^[130]. In this case, public actors communicate with privates, to realise public tasks in a clear public interest. There is a clear subordination relationship, and the individuals are in no place to negotiate, the fact that the user disposes of a choice, to use the chatbot or not, doesn’t change the relationship between the user and the state. Which means that these terms and conditions have to be analysed from a public law standpoint.

The principle of legality^[131] will in this sense be of the utmost importance to consider the legal standing of these texts. Such an analysis will need to be conducted for each system separately, which explains why we decided to leave it out of the present study. Indeed, the legality of each system will vary depending on the legal framework in place. The relevant laws should also be examined for the possible delegation of authority allowing the administration to enact such texts and the possible legal exceptions tailored for such systems.

It isn’t simple to define which type of administrative act these terms and conditions are, to the point that it might be easier to state what these texts aren’t.

The terms and conditions are abstract texts, presenting rules and principles applicable to all users, in this sense they can’t be administrative decisions as they miss the needed individuality^[132]. These disclaimers are fixed for all users of the service; there is no individual aspect at play, and thus, no administrative decision.

These terms and conditions can’t be part of an administrative contract, because of the missing bilaterality of the agreement, of the effects and of the missing legal basis authorising such contracts to be passed^[133].

These texts aren’t formal Laws or legislative ordinances since they weren’t approved nor published according to the constitutional and legal obligations in place^[134].

In this sense, we believe these texts constitute administrative ordinances^[135], although these texts should theoretically only be addressed to agents of the administrations and shouldn’t contain any rule of law^[136]. Since these texts aren’t considered as rules of law, they shouldn’t directly deploy any effect on the individuals^[137]. In a particular case, these texts would then only constitute hints about the way the authorities considered their information, but by elements permitting to infer the rights, or lack thereof, of the individuals.

If these terms and conditions, in the hierarchy of norms, really shouldn’t be able to create any rights and obligations for the individuals, we argue that a judge could consider these texts as void and without value for a specific case, thus reinstating the liability the authorities tried to avoid.

It is worth noting that previous studies were already conducted on the rather weak legal standing of such disclaimers in the protection of the trust principle the individuals should be able to put into advice given by the authorities chatbots^[138] and websites^[139].

We decided not to conduct such an analysis since providing definitive answer to this question would require, for each disclaimer an analysis of their wording as well as of the topical federal or cantonal laws. If the principle of the protection of trust in the information given by the authorities is present in multiple procedural laws, special provisions might be added to exempt certain chatbots or webpages from these provisions. The same is true for the clauses exempting the authorities from any liability for harm, each level of Government has put in place a specific legal framework around “state’s liability”^[140], if these laws don’t include exemptions for these systems, there is a high probability that the disclaimers and responsibility exemptions we analysed in this paper have little to no legal standing^[141].

4.3 The influence of these disclaimers on public trust

We argue that by implementing such disclaimers and pointing towards the unreliable nature of such technical systems, the authorities ask their users not to trust the system they implemented. If further studies are needed to evaluate the subjective impact of such terms and conditions on perceived trustworthiness of authorities, we can already affirm that these texts don’t comply with the usual standards of “good faith” actions^[142]. Nor with the federal recommendations on the use of AI by the authorities that specifically demand accountability^[143].

Having “terms and conditions” available with the chatbot itself is a positive element in the relationship between State and citizens. A specific information about the rights of the users, about their redress mechanisms are mechanisms of transparency that profits to the users. If the existence of privacy statement could have a positive impact on the trust of the users, the same effect might be true on the terms and conditions framing the information provided by such AI systems.

Although data protection texts included links to the legal basis governing their implementation, most disclaimers or terms and conditions about the content of the information provided by the chatbots, that we examined in this study presented no link of references to the applicable public laws^[144]. Although their legal standing is questionable at best, a non-experienced user might reasonably^[145] consider the clauses present in these terms and conditions to be binding.

Since most of these texts present clear disclaimers in favour of the authorities, because they affirm that the user is using the service at their own risks^[146]. A reasonable user might, rationally, consider these clauses to be an admission by the authorities of their untrustworthiness.

These broad disclaimers in favour of the authorities can’t, from a rational point of view participate in creating trust in the eGovernment systems. Why use the chatbot function if «False, incorrect and fictitious answers must be expected»^[147]. Why use such a system if the terms and conditions ask the users to «verify the information provided by [the chatbot] on the official websites»^[148]. We believe the systematic use of disclaimers on Swiss public websites might convey the wrong message about the trustworthiness of State as a whole, and in particular about the eGovernment systems that the authorities implement.

The contrast between the content of the data protection rules and the responsibility for the content and answers provided by the chatbots is striking. The former present clear pathways for citizens to exert their rights or to ask questions about their personal data processing. These provisions don’t include the right to ask for a review from the information given by the authorities themselves. Even in texts that mix both types of rules, the provisions for the rights of the users don’t include any rights for the users concerning the texts produced by the chatbots^[149].

5. Possible pathways to foster or restore Trust in the testing and implementation of AI systems

Considering the critics we expressed in this study, we see multiple pathways that could be considered and combined to foster trust in the experimentation and implementation of AI technologies for public services.

First, future projects should be cleared up both from a technical point of view as well as from a legal perspective before such technologies are implemented^[150]. By deciding to test unsafe technologies in a “production” environment, on a public facing website, the authorities take a risk to destroy trust from their users that might be difficult to gain back. Such prerequisites should in our view be generalised with internal and external audits to ensure that the solutions put in place are both safe and functional. As recently said in the examination of the future Swiss eID system, «there will be no second chance to win the trust of the Swiss citizens»^[151], we believe that this affirmation is applicable for every technological solution implemented by any authorities.

Second, authorities should develop clearly indicated and transparent “testing” framework for each new system. These tests would be clearly communicated and apparent for the users and be subject to specific regulation and testing requirements^[152]. For these tests, clear indications would be present on the systems themselves to differentiate these systems still undergoing testing from systems that are “in production”. Moyt of the chatbots we considered in this study are presented as fully functional piece of technology providing public services, the chatbot of the Swiss army indicates it’s prototype status in a separate popup, in the middle of a text on the goal of the service. An indication about the “testing phase” should be present from the start window.

A real testing framework, put in place on the government’s level the new algorithm is developed for, would have multiple advantages. Such systems could be implemented with ad hoc legislation that would define before the start of the experimentation both the duration of and the evaluation^[153] of the test. A legal framework would be put in place that both limit the risks and guarantee the accountability the authorities are taking upon themselves during the testing. A system being implemented for testing purposes should be clearly recognisable by the users, for example, with specific pages and warnings about the system. To better differentiate between testing and implemented technology, the chatbots still in the testing phase probably shouldn’t be present on the landing page of the authorities’ websites. Special mechanism would also be put in place to ensure a possibility for the users to signal errors and bugs, and to make sure their rights are guaranteed. In the case of a chatbot implemented by the authorities, we think for example of a button allowing in a simple way to ask for a human review of the advice given by the machine. The implementation of transparency and evaluation mechanisms would then ensure that the decision process to continue the development to allow for a better implementation of such technologies.

Third, the authorities should take accountability for the systems they implement. From a legal and moral point of view, authorities have a duty to act in good faith^[154], to act in a trustworthy manner. This means that they shouldn’t try to “hide” behind broad disclaimers and responsibility waivers. The validity and legal standing of such disclaimers being questionable at best, their implementation should really be put into question. The probable impact on public trust of such disclaimers far outweighs the protection the authorities could gain from them in their current redaction. The general principles surrounding the protection of trust already provide protection to the authorities for their actions on the web. Disclaimers should in this sense be reserved for systems still in their testing phase, with the necessary provision to ensure the rights of the users.

6. Conclusion

In this paper, we exemplified how the broad use disclaimer in terms and conditions by Swiss authorities in their implementation of AI chatbots is contrary to the duty of the authorities to act in a trustworthy manner. By implementing technology that they know is faulty, by saying that the service implemented shouldn’t be trusted, and by refusing to take accountability for the answers given by these chatbots, the administrations are asking the citizens not to trust these systems.

In this sense, the administrations’ practices surrounding their implementation of terms and conditions on their internet platforms are creating an appearance of untrustworthiness. Such an appearance creates unnecessary risks, both for future systems adoption and a risk for public backlash that could threaten a project that would be technically sound.

Since disclaimers might be unavoidable in certain testing situations, further research is needed to analyse the impact on the users of such disclaimers, their placement and the way they operate.

Although this point isn’t central in the present study, the authorities should, in our opinion be more cautious and question the need to implement such technologies^[155]. AI technologies present whole new sets of challenges, many of which have not yet been solved. From ecological concerns^[156], to unclear social impacts^[157], the unclear impact of AI should, in our opinion, warrant more caution in their implementation by public administration. It is always difficult to know what the future holds, uncertainties about the regulation of AI, the promises for the development of such systems and geopolitical tensions might have an impact on the generalisation of such systems. We encourage the authorities to review their use of these disclaimers and to accept to be accountable for their use of technologies and, thus, to take action to ensure that they continue to maintain a high level of public trust in their use and implementation of technologies.

1. This text was written by the author without the help of LLMs for the redaction or the brainstorming. Deepl was used to streamline the translations, and antidote was used to proofread the text. All links were checked on the 22.03.2026. ↑
2. For a discussion on the matter, see for example E. Toreini, M. Aitken, K. Coopamootoo et. al., The relationship between trust in AI and trustworthy machine learning technologies, Proceedings of the 2020 Conference on Fairness, Accountability, and Transparency, Barcelona Spain, in ACM, FAT* ‘20: Conference on Fairness, Accountability, and Transparency, 27 janvier 2020, p. 272‑283; Also in a broader sense T. M. Lechterman, The Concept of Accountability in AI Ethics and Governance, in J.B. Bullock, Y. Chen, J. Himmelreich et. al. (dir.), The Oxford Handbook of AI Governance, Oxford University Press, Oxford, 2024. ↑
3. For a discussion on the factors influencing trust in the relationship, see for example N. Schlicker, K.. Baum, A. Uhde et. al., How do we assess the trustworthiness of AI? Introducing the trustworthiness assessment model (TrAM), in Comput. Hum. Behav. 170, 2025, p. 2 ss. ↑
4. Multiple cues – information pieces – will be used to make an assessment about the trust to place in an agent E. J. De Visser, M. Cohen, A. Freedy et. al., A Design Methodology for Trust Cue Calibration in Cognitive Agents», in R. Shumaker, S. Lackey (dir.), Virtual, Augmented and Mixed Reality. Designing and Developing Virtual and Augmented Environments, Springer International Publishing, Cham, 2014, p. 251‑262; a concept used by N. Schlicker et. al., How do we assess the trustworthiness of AI?, 3, 7 ss. ↑
5. On the question of accountability as an ethical requirement in the implementation of AI, see for example D. Leslie, Understanding artificial intelligence ethics and safety: A guide for the responsible design and implementation of AI systems in the public sector, The Alan Turing Institute, 2019, p. 23 ss. the authorities in these disclaimers refuse to implement any «remedial accountability», on the topic p. 25 s. The focus is here on the accountability for their implementation of the systems, more than of the accountability that the systems can achieve, for a discussion of the concept of AI accountability and their use see for example M. Busuioc, Accountable Artificial Intelligence: Holding Algorithms to Account) in Public Adm. Rev., 81-5, 2021 pp. ↑
6. Only partial lists exist which are sometimes not updated regularly. The list provided by Algorithm watch dates form 2023 Atlas of Automation Switzerland, AlgorithmWatch CH, online on the website www.algorithmwatch.ch,en,atlas; and the list from the CNAI lists only systems implemented by the federal administration, Projektsteckbriefe Kompetenznetzwerk CNAI, Geschäftsstelle CNAI – OFS, 2026, online on the website http://www.bk.admin.ch. ↑
7. This means purely internal chatbots are excluded from this study, for an internal AI system on the federal level see Office fédéral de l’informatique et de la télécommunication OFIT, RoBIT apporte son soutien en cas de problèmes d’informatique, online on the website http://www.bit.admin.ch. ↑
8. This excludes the AI systems implemented by private entities that were trained on public data «GenevaGPT , online on the website www.geneva.gpt.liip.ch. ↑
9. See in particular P. Glass, Datenschutzrecht für künstliche Intelligenz in der öffentlichen Verwaltung, in Datenschutz Rechtliche Schnittstellen, Dike Verlag, Zürich, 2023, p. 252 for chatbots. ↑
10. The chatbot “Frag Esi”, online on the website https:www.fragesi.ch, totally changed its presentation at the end of 2025 and as of march 2026 gives an “error 403”; An archived version of november 2025 is still available at the internet archive, online on the websitewww.web.archive.org. ↑
11. «Esi – le chatbot de l’ASF», online on the website http://www.esa.admin.ch. ↑
12. For a deep reflection about trust, in particular in law see K. Sutter, Vertrauen im Recht: eine Theorie für den demokratischen Verfassungsstaat, Dike, Zürich, 2020 for the definitions of trust in particular p. 177 ss; or also M. Hartmann, Die Praxis des Vertrauens, Suhrkamp, Berlin, 2022, p. 20; for an analysis of the literature around the implementation of digital technologies by public authorities, see also L. Alzahrani, W. Al-Karaghouli, V. Weerakkody, Analysing the critical factors influencing trust in e-government adoption from citizens’ perspective: A systematic review and a conceptual framework, in Int. Bus. Rev. 26-1, 2017, p. 165 ss. ↑
13. R. C. Mayer, J. H. Davis, F. D. Schoorman, An Integrative Model of Organizational Trust, in Acad. Manage. Rev., 20-3, 1995, p. 712. ↑
14. We refer the reader to the large literature on the different meaning of trust, its impact and the factor that influence it, for example R. Campagna, A. Mislin, K. Dirks, H. A. Elfenbein, The (Mostly) Robust Influence of Initial Trustworthiness Beliefs on Subsequent Behaviors and Perceptions, in SSRN Scholarly Paper, Rochester, NY, Social Science Research Network, 18 février 2021; or D. J. Dekker, D. Krackhardt, P. H. Franses, Dynamic Effects of Trust and Cognitive Social Structures on Information Transfer Relationships, in SSRN Scholarly Paper, Rochester, NY, Social Science Research Network, 3 décembre 2002. ↑
15. K. Sutter, Vertrauen im Recht, p. 411 ss. ↑
16. On the tension between the authorities and individuals, see for example G. Malinverni, M. Hottelier, M. Hertig Randall et. al., Droit constitutionnel suisse Vol. I: L’Etat, 4ème, Stämpfli, Bern, 2021, p. 5 ss.; On the the regulation of the exercise public power by the interdiction of arbitrary actions see for example P. Moor, A. Flückiger, V. Martenet, Droit administratif 1 : les fondements, vol. 1, 2012, p. 892 ss. ↑
17. On the notion of legality see for example C. Grisel Rapin, La légalité, in Les grands principes du droit administratif, Schulthess, 2022, online on the website: www.archive-ouverte.unige.ch. ↑
18. On the difference and interdependencies of the two notions see for example J. Jackson, J. Gau, Carving Up Concepts? Differentiating between Trust and Legitimacy in Public Attitudes Towards Legal Authority, in SSRN Scholarly Paper, Rochester, NY, Social Science Research Network, 20 février 2015. ↑
19. For analysis on the notion in EU law, see for example J. Mendes, Good Administration in EU Law and the European Code of Good Administrative Behaviour, in SSRN Scholarly Paper, Rochester, NY, Social Science Research Network, 2009, p. 4 mentions trust explicitely; or J. Ponce, The Right to Good Administration and the Role of Administrative Law In Promoting Good Government, in SSRN Scholarly Paper, Rochester, NY, Social Science Research Network, 24 janvier 2016. ↑
20. ↑
21. For example, P. Rodríguez et al., «How the exercise of the right to information (RTI) affects trust in political institutions», online on the website www.sciencedirect.com. ↑
22. S. G. Grimmelikhuijsen, S. J. Piotrowski, G. G. Van Ryzin, Latent transparency and trust in government: Unexpected findings from two survey experiments, in Gov. Inf. Q., 37- 1, 2020. ↑
23. K. Sutter, Vertrauen im Recht, p. 226 ss. ↑
24. On this notion F. Bernard, La protection de la bonne foi, in Les grands principes du droit administratif, Schulthess, 2022, online on the website http://www.archive-ouverte.unige.ch. ↑
25. Constitution fédérale de la Confédération suisse (Cost.), du 18 avril 1999, RS 101. ↑
26. «Treu und Glauben» ↑
27. Code civil suisse (CC), du 10 décembre 1907, RS 210. ↑
28. Stratégie suisse de cyberadministration 2020–2023, 2019, pp. 8273-8282. ↑
29. Stratégie Utilisation de systèmes d’IA dans l’administration fédérale, du 21 mars 2025, SB021, p. 6. ↑
30. Kantar, eGovernment MONITOR 2023, Initiative D21 e. V, Technical University of Munich, p. 15. ↑
31. Ibid. ↑
32. Ibid.; H. Amberg, C. Büchler, M. Buess, Étude nationale sur la cyberadministration 2025. Rapport ciblé sur l’IA, Lucerne, Adlingenswil, Interface Politikstudien Forschung Beratung AG et DemoSCOPE AG, 2025, p. 9 ss. ↑
33. The notion of “trustworthy AI” plays a key role as a justification for the AI Act of the EU an act that isn’t applicable to Switzerland Regulation (EU) 2024,1689 of the European Parliament and of the Council of 13 June 2024 laying down harmonised rules on artificial intelligence, du 13 juin 2024; for a commentary on the notion see for example J. Eichenhofer, Trustworthy AI and Fundamental Rights, in Eur. Rev. Public Law, 2024. ↑
34. L. Alzahrani, W. Al-Karaghouli, V. Weerakkody, Analysing the critical factors influencing trust in e-government adoption from citizens’ perspective, vol. 12; M. Z. I. Lallmahomed, N. Lallmahomed, G. M. Lallmahomed, Factors influencing the adoption of e-Government services in Mauritius, in Telemat. Inform. 34 – 4, 2017, p. 67; R. Fakhoury, B. Aubert, Citizenship, trust, and behavioural intentions to use public e-services: The case of Lebanon, in Int. J. Inf. Manag. 35-3, 2015, p. 348. ↑
35. See for example S. W Palmer, The Role of Trust in Facilitating Human-Artificial Intelligence Relationships, Aarhus University, 2024. ↑
36. For such a study, see for example on the topic of “healh data spaces” P. Daniore, F. Zavattaro, F. Gilles, Public Trust in a Swiss Health Data Space, Zürich, University of Zürich – Digital Society Initiative, 2024, online on the website http://www.health.dsi.uzh.ch. ↑
37. In 2002, Technological developments have a significant influence on the drivers of eGovernment (free translation of 1st K. Schedler, L. Summermatter, Was treibt das eGovernment?, in D. Spahni (dir.), eGovernment 2: Perspektiven und Prognosen, P. Haupt, Bern. 2002 p. 120; The expectation from the users still support this orientation eGovernment MONITOR 2025, Initiative D21 e. V, Technical University of Munich, p. 14. ↑
38. As an example, the federal administration established a report in 1979 on the electronic data processing in the federal administration Bericht uber den Einsatz der elektronischen Datenverarbeitung in der Bundesverwaltung, Berne, Office federal de l’organisation, 1979. ↑
39. Digital first is at the center of the vision for the Swiss digital strategy 2026 Stratégie Suisse numérique 2026, du 12 décembre 2025, 39, p. 2. ↑
40. L. Stenberg, S. Nilsson, Factors influencing readiness of adopting AI : A qualitative study of how the TOE framework applies to AI adoption in governmental authorities, 2020, p. 55. ↑
41. The “vision” of the federal administration centers around optimisation, helping the employees and developing “new” solutions Stratégie Utilisation de systèmes d’IA dans l’administration fédérale, vol. 29, p. 3; N. Braun Binder, M. Spielkamp, C. Egli et al., Einsatz Künstlicher Intelligenz in der Verwaltung: rechtliche und ethische Fragen, 2021, p. 15. ↑
42. L. Stenberg, S. Nilsson, Factors influencing readiness of adopting AI, vol. 40, p. 42 s.; N. Braun Binde et al., vol. 41, p. 16. ↑
43. P. Ringeisen, A. Bertolosi-Lehr, L. Demaj, Automatisierung und Digitalisierung in der öffentlichen Verwaltung: digitale Verwaltungsassistenten als neue Schnittstelle zwischen Bevölkerung und Gemeinwesen, in Yearb. Swiss Adm. Sci. 9-1, 2018, p. 51 ss. ↑
44. For an independent overview of automation projects in Switzerland see the Atlas of Algorithm watch, vol 6 it is possible to select only government authorities in the searchbox; For an actual list of AI project on the federal level, see the AI Projektdatenbank of the CNAI 01.26, vol. 6. ↑
45. With organisational performance, these systems also search to improve the effectiveness of the State by delivering better public service O. Neumann, K. Guirguis, R. Steiner, Exploring artificial intelligence adoption in public organizations: a comparative case study, in Public Manag. Rev., 2022, p. 5; This point is even a domain of action in the last federal strategy Stratégie Utilisation de systèmes d’IA dans l’administration fédérale, vol. 29, p. 7. ↑
46. N. Braun Binder et al., vol. 41, p. 16 ss. ↑
47. For a summary of the expectations of public administration see Des algorithmes dans les administrations publiques 30 mars 2021, online on the website www.ictjournal.ch. ↑
48. Although fraud detection doesn’t need AI see A. Bănărescu, Detecting and Preventing Fraud with Data Analytics, in Procedia Econ. Finance, 32, 2015, pp. 1827‑1836; The developments of such AI systems for tax purposes in Switzerland in this domain seem to have only begun Automatisation et systèmes de gestion des risques au sein des administrations fiscales cantonales – swissprivacy.law, online on the website http://www.swissprivacy.law., M. Nufer, Künstliche Intelligenz in der Steuerveranlagung, in ASA, 2019, p. 88; For a french outlook see M. Wolf, S. Dale, Les technologies numériques au service de l’administration de l’impôt , in ERA Forum, 19-3, 2019, pp. 457‑464. ↑
49. In a field in constant evolution, it can be hard for the scholar to frame a subject at a given point, see for example W. G. de Sousa, E. R. P. de Melo, P. H. De Souza Bermejo et al., How and where is artificial intelligence in the public sector going? A literature review and research agenda, in Gov. Inf. Q., 36-4, 2019. ↑
50. On the federal level for example there are at least 10 chatbots projects underway as of 2026 AI Projektdatenbank 01.26, vol. 6, p. 7, 17, 48, 49, 69, 107, 114, 118, 119, 121. ↑
51. For example, of uses of chatbots worldwide, see for example Y. Wang, N. Zhang, X. Zhao, Understanding the Determinants in the Different Government AI Adoption Stages: Evidence of Local Government Chatbots in China, in Soc. Sci. Comput. Rev., 40-2 ,2022, p. 537. ↑
52. O. Neumann, K. Guirguis, R. Steiner, Exploring artificial intelligence adoption in public organizations, vol. 45, p. 18. ↑
53. Chatbots (general) in Public Administration, AlgorithmWatch CH, online on the website http://www.algorithmwatch.ch. ↑
54. P. Ringeisen, A. Bertolosi-Lehr, L. Demaj, Automatisierung und Digitalisierung in der öffentlichen Verwaltung, vol. 43, p. 54. ↑
55. H. Amberg, C. Büchler, M. Buess, vol. 32, p. 10. ↑
56. On this see for example V. Habib Lantyer, Hallucination by Design: The Hidden Incentives of AI, in SSRN Scholarly Paper, Rochester, NY, Social Science Research Network, 8 septembre 2025. ↑
57. S. Wachter, B. Mittelstadt, C. Russell, Do large language models have a legal duty to tell the truth? , (in R. Soc. Open Sci., 11-8, 2024. ↑
58. See T. Poledna, Staatliche Informationspflichten (Grundversorgung mit elektronischen Daten), in Tag. 2003 Für Inform. Recht Vom 29 August 2003 Bern Journ. 2003 Inform. Jurid. 29 Août 2003 À Berne. ↑
59. Federal Constitution (Cst.) of the 18 April 1999, RS 101. ↑
60. Of the 21 March 1997, RS 172.010. ↑
61. See for example for canton of Geneva Loi [genevoise] sur l’information du public, l’accès aux documents et la protection des données personnelles (LIPAD), du 5 octobre 2001, RS GE A 2 08. ↑
62. Of 17 December 2004, RS 152.3. ↑
63. For an analysis of these reglementations see V. Mabillard, M. Pasquier, Les lois d’accès à l’information en Suisse. Mise en perspective internationale, 2022, pp. 325‑344. ↑
64. T. Poledna, vol. 58, p. 88. ↑
65. The project started in 2022 and is offline as of 2026 RAV Baselland startet einen Chatbot-Pilot – Kanton Basel-Landschaft, online on the website http://www.baselland.ch. ↑
66. Loi fédérale sur l’assurance-chômage obligatoire et l’indemnité en cas d’insolvabilité (Loi sur l’assurance-chômage, LACI), of the 25 june1982, RS 837.0. The federal law charges the cantons of the execution of the proceedings in matter of unemployment. In particular, the registration of unemployed people (art. 17 al. 2 LACI and art. 85 LACI). ↑
67. Gesetz über die Arbeitsvermittlung und die Arbeitslosenversicherung (AVLG), of the 25.03.1999, RS-BL 837. ↑
68. IBM, What is a chatbot, in IBM explainers blog, online on the website www.ibm.com. ↑
69. Although the terminology “digital administrative assistant” is sometime used to describe conversational agents, see P. Ringeisen et al., Automatisierung und Digitalisierung in der öffentlichen Verwaltung, , vol. 43, p. 57 ss.; These types of hopes are already mentioned in descriptions of future project on the federal level, see for example the CFC- Copilot , Suisse numérique, online on the website www.digital.swiss. ↑
70. Such guided forms can be found on numerous e-government proceedings. For example in order for a person to announce their address changes to the authorities via http://www.eumzug.swiss, to create a new company on easygov visit the website www.easygov.swiss. Similar forms are available on the cantonal and communal level and are, to our knowledge always considered as eGovernment forms, more than chatbots. ↑
71. This is the case for Alva form canton Basel Stadt and the BarBot of the federal archive (that indicates using GPT-4o). ↑
72. AI Chatbot, Deeplink.ai (16 février 2026), online on the website www.deeplink.ai. ↑
73. Due to a lack of public informations, see the publicity from the enterprise that installed the solution “Voicebot from UMB”, online on the website http://www.umb.ch; their system uses the vocal solution from “Plateforme Conversationnelle Omnicanale Spitch”, Spitch, online on the website http://www.spitch.ai. ↑
74. Amt für Migration, Integration und Bürgerrecht – Kanton Basel-Landschaft, online on the website http://www.baselland.ch. ↑
75. Datenschutzerklärung Chatbot – KI-Bot Basel-Landschaft, du 25 avril 2025. ↑
76. Disclaimer, baselland.ch, online on the website http://www.baselland.ch. ↑
77. Conditions générales d’utilisation – Police Region Morges, Conditions générales d’utilisation – Police Region Morges, online on the websitewww.prm-vd.ch. ↑
78. The chatbot is available at the bottom of the webpages and on it’s dedicated page Alva, Kanton Basel-Stadt, online on the website, http://www.bs.ch. ↑
79. The link is called «datenschutzbestimmungen» and point toDatenschutzerklärung, Kanton Basel-Stadt (16 mai 2024), online on the website http://www.bs.ch. ↑
80. the chatbot of the Swiss army, online on the website h http://www.armee.ch.in the infobox on the right corner of the chat has one “disclaimer” (Clause de non responsabilité) and one document about the “Information about the use of the chatbot” that contains rules about the timeline of the suppression of the metadata. From a historical standpoint, the FragEsi chatbot [V.n. 10] had two separate documents one named «data protection», and one «terms and conditions». ↑
81. Archives fédérales suisses AFS, online on the website http://www.bar.admin.ch. ↑
82. Archives fédérales suisses AFS, Clause de non-responsabilité, déclaration de protection des données et conditions d’utilisation du chatbot, that can be downloaded online on the website http://www.bar.admin.ch. ↑
83. Idem, p. 5 which contains a link to the «Conditions of use and data protection policy», online on the website www.recherche.bar.admin.ch. ↑
84. «Terms and conditions», online on the website www.admin.ch. ↑
85. The chatbot is available only with a Swiss internet connection on the bottom right corner of the page, online on the website www.armee.ch. ↑
86. Free translation of «Clause de non responsabilité». ↑
87. Free translation of «Mentions légales», the popup contains indications about the authority in charge of the chatbot and means to contact them (address, email and telephone). ↑
88. The HofdererBot present with a button with the logo of the municipality at the bottom left corner Home – Gemeinde Hochdorf, 19 février 2026, online on the website http://www.hochdorf.ch. ↑
89. Section about the Integration of third-party services and content, Einbindung von Diensten und Inhalten Dritter, in Datenschutzerklärung, online on the website www.hochdorf.ch. ↑
90. The link to this text is present after the name of the company, but the document is hosted on the municipality’s website Datenschutzrichtlinie_für_den_Chatbot_der_CAUSE_Systems, online on the website www.hochdorf.ch. ↑
91. Voicebot UMB, vol. 73. ↑
92. Stadt Kreuzlingen, Einwohneramt, online on the website www.kreuzlingen.ch. ↑
93. Stadt Kreuzlingen, Datenschutzerklärung, online on the website https:,,www.kreuzlingen.ch,datenschutzerklaerung. ↑
94. Voicebot UMB, vol. 73. ↑
95. Free translation of the indication present on the chatbot before it’s “activation” by the user Wenn Sie mich aktivieren, unterstütze ich Sie bei der Suche nach den richtigen Informationen, Startseite, Kanton Basel-Stadt, online on the website https: www.bs.ch. ↑
96. Free translation of the first messages of the chatbot «je vais vous accompagner dans la recherche d’informations sur nos services» in Accueil – Police Region Morges, online on the website www.prm-vd.ch. ↑
97. Free translation of «Comment puis-je vous aider?». ↑
98. Free translation of «Es kann sein, dass einzelne Informationen nicht korrekt sind. Bitte überprüfen Sie wichtige Informationen», text of the chatbot present on the webpage of the federal archives Archives fédérales suisses AFS, vol. 84. ↑
99. Free translation of «le chatbot de l’armée suisse peut commettre des erreurs. Vérifiez les informations importantes». ↑
100. Free translation of «Meine Antworten werden von einer künstlichen Intelligenz generiert und sind deshalb nicht immer korrekt», vol. 98. ↑
101. Free translation of «je suis à mes débuts et j’apprends tous les jours». ↑
102. Free translation of «[a]ll information provided without guarantee» at the bottom of the chat window, vol. 91. ↑
103. Free translation of «[l]a version actuelle est un prototype et sert à effectuer d’autres tests utilisateurs». ↑
104. Kanton Basel Stadt migration office chatbot. ↑
105. Bar Bot of the federal archives. ↑
106. Alva. ↑
107. So the Swiss army and Morges municipal police chatbot. ↑
108. Data protection Basel Stadt, vol 79. ↑
109. Conditions générales chatbot archives, ,vol. 85, p. 2. ↑
110. Clause 5 indicates an assistance and contact point Conditions générales chatbot archives, p. 5; The Municipal police chatbot includes a contact email in the first messages, vol. 99; While Alva includes an email adress for « further information of questions about these terms and conditions» : «weitere Informationen oder Fragen zu diesen Nutzungsbedingungen» (n.8 of the KI ALVA terms) Data protection Basel Stadt, vol. 79 In contrast, the general disclaimer of the Kanton Basel Stadt gives no contact informations baselland.ch>disclaimer, vol. 76. ↑
111. See for example the numerous links to legal texts in the AFS data protection policy, vol. 86. ↑
112. Free translation of «[d]er Kanton Basel-Stadt übernimmt keine Gewähr für die Richtigkeit, Vollständigkeit oder Aktualität der von Alva zu Verfügung gestellten Informationen. Die Informationen dienen lediglich zu allgemeinen Informationszwecken und sind rechtlich nicht bindend. Wir bitten Sie, die von Alva zu Verfügung gestellten Informationen auf den offiziellen Websites des Kantons zu überprüfen» Data protection Basel Stadt, vol. 79. ↑
113. «Gewährleistungsausschluss», Idem. ↑
114. Free translation of «Der Kanton verwendet all seine Sorgfalt darauf, dass die Informationen auf den Websites zum Zeitpunkt der Veröffentlichung korrekt und aktuell sind und keine Rechte Dritter (insbesondere Rechte des Persönlichkeitsschutzes) verletzen. Dennoch können weder der Kanton noch einzelne Ämter oder Dienststellen eine Gewährleistung hinsichtlich der Genauigkeit, Zuverlässigkeit oder Vollständigkeit der Information machen», Idem. ↑
115. Free translation of point 2 of the conditions for Alva: «Alva basiert auf künstlicher Intelligenz und kann daher Fehler machen oder ungenaue Informationen bereitstellen. Insbesondere können Antworten korrekt erscheinen, inhaltlich jedoch falsch sein. Für die Folgen, die aus der Nutzung der von Alva bereitgestellten Informationen resultieren, wird vom Kanton Basel-Stadt keine Haftung übernommen», Idem. ↑
116. Free translation of the exclusion of liability clause «Der Kanton lehnt jegliche Haftung für Schäden oder Folgeschäden ab, die sich aus dem Zugriff auf seine Websites beziehungsweise auf einzelne Teile davon (wie z.B. heruntergeladene Dokumente) deren Benutzung (bzw. aus der Unmöglichkeit des Zugriffs oder der Benutzung) oder Links auf andere Websites ergeben» Idem. ↑
117. Free translation of «L’utilisation du chatbot et ses éventuelles conséquences relèvent de votre responsabilité. Les Archives fédérales ne garantissent en aucun cas la véracité, l’exactitude, l’exhaustivité et la fiabilité des réponses fournies par le robot. Des réponses fausses, inexactes ou inventées ne sont pas exclues. Vous devez toujours vérifier les informations et les sources indiquées. Les Archives fédérales ne sont en aucun cas responsables des actions entreprises par les utilisateurs ou de leur décision de ne pas agir sur la base des informations fournies par le chatbot, ni de leurs quelconques conséquences. Toute prétention en matière de responsabilité est expressément exclue», Terms and conditions of BarBot, colo. 85, p. 2. ↑
118. Free translation of «le chatbot est un interlocuteur central pour les militaires. Il traite sans délai les questions standard et transfère les requêtes plus complexes aux services compétents» «Sur armee.ch, un chatbot répond aux questions 24h, 24 et 7j,7», online on the website https: www.vtg.admin.ch. ↑
119. Free translation of «Le chatbot de l’armée suisse a été développé pour fournir des informations et des renseignements généraux sur l’Armée Suisse», the disclaimer of the Swiss army chatbot seeems directly inspired from the now defunct FragEsi chatbot that had a very similarly worded introcution: «Esi wurde entwickelt, um allgemeine rechtliche Informationen und Auskünfte rund um die ESA bzw. das Schweizer Stiftungsrecht bereitzustellen». ↑
120. Free translation of «L’utilisation du chatbot se fait à vos propres risques et périls. L’Armée suisse ne garantit pas l’exactitude, la précision, l’exhaustivité et la fiabilité des réponses fournies. Il faut s’attendre à des réponses fausses, erronées ou inventées de toutes pièces. Vérifiez chaque réponse, y compris les éventuelles sources indiquées. L’Armée suisse n’est pas responsable si l’utilisateur ou l’utilisatrice entreprend ou s’abstient d’entreprendre des actions en se fiant à l’exactitude et à l’exhaustivité des réponses et qu’il en résulte des inconvénients ou des dommages. Toute prétention en responsabilité causée par l’utilisation est expressément exclue» infobox «Clause de non-responsabilité», on the chatbot of the Swiss army website; this disclaimer uses the same exact wording of the now defunct Fragesi Chatbot (with the exception of the name of the authorities). ↑
121. Free translation of «Der Kanton unternimmt alle Anstrengungen, um einen sicheren Betrieb der Webseite sicherzustellen. Dennoch können Mängel wie z.B. Datenverlust oder -verfälschung, Virenbefall, Betriebsunterbruch u. a. nicht ausgeschlossen werden. Der Zugriff auf die Webseite erfolgt deshalb auf eigene Gefahr und in eigener Verantwortung». ↑
122. Free translation of «Die Publikationen auf den Webseiten des Kantons Basel-Landschaft weisen lediglich Informationscharakter auf. Der Kanton Basel-Landschaft behält sich ausdrücklich vor, Inhalte jederzeit ohne Ankündigung ganz oder teilweise zu ändern, zu löschen oder die Veröffentlichung zeitweise oder endgültig einzustellen. Haftungsansprüche gegen den Kanton Basel-Landschaft wegen Schäden materieller oder immaterieller Art, welche aus dem Zugriff auf die oder der Nutzung bzw. Nichtnutzung der veröffentlichten Informationen oder durch technische Störungen entstanden sind, werden abgelehnt», Idem. ↑
123. Free translation of «[d]e plus, [l’association] s’efforce de donner des informations exactes et actualisées sur son site. Néanmoins, elle ne peut garantir l’exactitude, la fiabilité et l’exhaustivité des informations qui y figurent ou auxquelles il est fait référence. En outre, des défaillances du système telles que des interruptions, des attaques par des virus ou des pertes de données ne peuvent pas être exclues» Conditions d’utilisation – police Morges, vol. 64. ↑
124. It is worth noting that these unreliabilities already made their way to the press À Morges, l’IA de la police locale affirme qu’on peut rouler à 58 km,h sans amende, in 24 heures, 3 septembre 2025, online on the website http://www.24heures.ch. ↑
125. On this point in US law, see for example L. Y. G. Tenzer, Defamation in the Age of Artificial Intelligence, in SSRN Scholarly Paper, Rochester, NY, Social Science Research Network, 18 août 2023. ↑
126. The authorities have a certain responsability to inform the public and to close off dangerous ways, on the matter see G. Seferovic, Die Haftung des Gemeinwesens für Schäden durch Naturgefahren auf Wanderwegen zwischen Werkeigentümer- und Staatshaftung, in Sicherheit&Recht , 1-2018, p. 54 ss. ↑
127. R. G. Briner, Haftung für Informationen auf Websites, in Sic!, 2002, p. 236 ss. ↑
128. For special types of terms and conditions on the internet, see M. Dill, Internet-Verträge, in Aktuelle Juristische Prax, 2000, p. 1513 ss; For online platforms see for example A.-C.. Fornage,Les plateformes en ligne et la protection du consommateur … vers un changement de paradigme, in Pichonnaz, Werro (dir.), La place du consommateur au quotidien, Schulthess, Genève – Zurich – Bâle, 2022, p. 165‑195, online on the website www.iris.unil.ch. ↑
129. Briner, Sic!, 2002 p. 231, vol. 130. ↑
130. P. Moor et al., Droit Administratif 1, vol. 16. p. 95. ↑
131. G. Malinverni et al., DCS I, vol. 16, p. 1812 ss; See also C. Grisel Rapin, La légalité, vol. 17. ↑
132. P. Moor, E. Poltier, Les actes administratifs et leur contrôle, Entièrement revue, mise à jour et augmentée, coll. Droit administratif Vol. 2, Stämpfli, Berne, 2011, p. 12. ↑
133. Ibid., p. 415 ss. ↑
134. G. Malinverni et al., DCS I, vol. 16, p. 1532 ss. ↑
135. For a general presentation see for example Ibid., p.1670 ss. ↑
136. Ibid., p. 1674. ↑
137. P. Moor et al., Droit Administratif 1, vol. 16, p. 420 ss. ↑
138. N. Braun Binder, L. Obrecht, G. Wittmer, Vertrauensschutz bei fehlerhaften Behördenauskünften durch Chatbots, in IusNet Digit, Recht Datenrecht, 2024. ↑
139. J. Klemperer, P. Hauenstein, Vertrauensschutz bei behördlichem Webauftritt, in Aktuelle Juristische Prax. 2023, pp 871‑880. ↑
140. On the federal level, the Loi fédérale du 14 mars 1958 sur la responsabilité de la Confédération, des membres de ses autorités et de ses fonctionnaires (Loi sur la responsabilité, LRCF), du 14 mars 1958, RS 170.32. ↑
141. According to F. Uhlmann, J. Kaspar, Schweizerisches Staatshaftungsrecht, Dike, Zürich, 2017, p.149 there should be an exam for each disclaimer if applicable laws allow for private agreement. ↑
142. J. Klemperer, P. Hauenstein, vol. 142. ↑
143. The fourth line of action reads : «In order to determine and establish liability in the event of damage, accident or breach of law, it is necessary to clearly establish responsibility when using AI» (free translation) in Intelligence artificielle – lignes directrices pour la Confédération, 2020, online on the website http://www.sbfi.admin.ch., p. 5. ↑
144. Since the law is both at the basis and the limit of the authorities’ actions (art. 5 al. 1 Cst.), referencing the applicable legal basis for a public activity is at least a “good practice” since it allows users to confirm the information they are given. The presence of such texts could also empower the users to exercise their rights. ↑
145. Because of the trust he has in the authorities to follow the core constitutional principles of legality and good faith (art. 5 al. 1 and al. 3 Cst.). ↑
146. Baselland.Ch>Disclaimer, Der Zugriff auf die Webseite erfolgt deshalb auf eigene Gefahr und in eigener Verantwortung, vol. 76. ↑
147. Haftungsausschluss armee.ch, vol.123. ↑
148. Free translation from the point 2 of the terms of uses of the KI Assistent Alva, Data protection Basel Stadt, vol. 79. ↑
149. For example only the the data protection statement includes a section «your rights» (vos droits) Terms and Conditions of BarBot, vol. 85, p. 4. ↑
150. We note that this point is part of the guiding principles of the federal strategy for the implementation of AI in the federal administration Stratégie Utilisation de systèmes d’IA dans l’administration fédérale, vol. 29, p. 3. ↑
151. Zur Erlangung des Vertrauens der Schweizer Bevölkerung in diese vom Bund bereitgestellte Lösung wird es keine zweite Chance geben, in Prüfung des Schlüsselprojektes, Eidgenössische Finanzkontrolle (EFZ), Bern, 2025, p. 25. ↑
152. For and example of such testing framework, see for example the way the eVoting systems are tested – for twenty years in Switzerland, with both ad hoc legislation and constant evaluations see the dossier about the trials online on the website www.bk.admin.ch. ↑
153. On the federal level, a framework for pilot projects is envisionned, which doesn’t include provisions in favor of the users, see art. 15 Loi fédérale sur l’utilisation des moyens électroniques pour l’exécution des tâches des autorités (LMETA), du 17 mars 2023, RS 172.019; For legistical consideration see for example A. Flückiger, (RE)FAIRE LA LOI: traité de légistique à l’ère du droit souple, Stämpfli Éditions, Genève, 2018, p. 645 ss., for evaluation of legislation, and 653 ss for experimental legislation in particular. For a specific approach on sandboxes see for example S. Volz, KI Sandboxen für die Schweiz?, in SZW RSDA 1,2022, p. 51. ↑
154. For similar recommendations, see Amberg,Büchler, Buess, vol. 32, p. 14. ↑
155. We see similar call for caution for example in Amberg, Büchler, Buess, vol. 32, p. 14. ↑
156. See for a study on the environmental impact, for example T. Le Goff, Recommandations pour une action publique en faveur d’une IA générative respectueuse de l’environnement, 2023, p. 4 ss. ↑
157. For a discussion on the different cost of the implementation of AI see D. Acemoglu, Harms of AI, in J.B. Bullock, Y. Chen, J. Himmelreich et al. (dir.), The Oxford Handbook of AI Governance, Oxford University Press, Oxford, 2024, p. 0. ↑