1.1. What are mixed-functions IoT objects? A methodological explanation

It is important to explain the meaning of the expression mixed-functions IoT objects. As a preliminary remark, the IoT technology is extremely adaptable as it consists of creating objects connected with sensors to a cloud system where a reply is elaborated in response to an input received from the sensors.Footnote ⁴ Let us start with some examples of IoT as medical devices and then some examples of mixed-functions IoT devices to better clarify what legal regime is applicable to each of the abovementioned groups. IoT objects can be medical devices such as real-time monitoring vests. In this case, an IoT is considered from a regulatory point of view as a medical device because its functions fall within the ones described by Article 2 MDR and because the IoT manufacturer decides to follow the MDR rules concerning the safety certification of medical devices to market it into the EU. The MDR is a special safety regime compared to the general objects’ safety rules applicable in the EU (infra 2.2). Instead, the general objects safety rules apply to consumer objects. It goes almost without saying that IoT devices can also be marketed as consumer objects. Think about a home cleaning robot as an example of a consumer IoT application. The objective of the application of consumer safety rules is to obtain an EU conformity certification (CE) as well as in the case of medical devices. Nevertheless, the rules to follow in this last case are less streamlined and strict than in the case of medical devices. This article wants to investigate the private law liability issues concerning an emerging third category of IoT objects which formally are consumer ones, hence they follow the so-called, consumer law acquis (including the general safety regulations), but they do have a non-certified/certifiable medical function which in this article will be called e-health function.

Let me make an example to make things clearer. A smartwatch can be defined as an Internet of Things (IoT)-powered object. Usually, devices like smart-watches are called “wearables,” which are “miniaturised electronic devices that can be easily donned on and off the body or incorporated into clothing or other body-worn accessories” Footnote ⁵ but this does not mean that all IoT objects are wearables. Following with the same example, a smartwatch not only tells you the time, but also becomes part of your daily routine. Some models can track down your vitals (e.g. your heart rate or the steps you take daily through the help of sophisticated sensors) and provide services such as displaying push-up notifications for emails received and calendar appointments. Sometimes, they can inform you of an altered and potentially dangerous health risk or call the emergency number on your behalf.Footnote ⁶

The difference between two objects such as the smartwatch and a monitoring IoT vest might be non-existent for an average consumer/patient, but it is an important one in terms of certifications and regulatory frameworks. In the first case, the object is considered a consumer object despite its e-health function. This category of devices is subject to CE certification, which can be obtained by complying with the rules on the safety of products. Tracking down your vitals and signalling a potential health risk is an e-health function because the object is marketed as a consumer object. I cannot say that those functions are medical ones as they are not performed by an object certified as a medical device. However, they might have de facto the same function. In the IoT vest case, the object is a proper medical device and follows different certification paths for its safety of use.

In this article, as a methodological choice, I will deal mainly with the EU policy proposals and legal acts concerning IoT objects with both e-health and consumer functions. The smartwatch case, to be clearer.

Investigating the regulatory and liability schemes of this emerging kind of connected objects is important because two relevant and consequent issues could be developing soon. The first one is the creation of a regulatory gap: if mixed-functions IoTs are treated as simple consumer objects even when they do have an e-health or lato sensu medical function (such as vitals monitoring) the risk is that the application of EU consumer and general safety rules will create a lower standard of protection for life and health than the one that is provided by the MDR. The second issue is a consequence of the previously cited regulatory gap: if mixed-functions IoTs cause damage to consumers, then the application of product liability or AI liability rules (applicable or soon to be applied) must at least be easier for consumers (see infra 3.2 and 3.3).

1.2. The importance of mixed-function IoT devices: a multi-layered explanation

It is important to focus on this new branch of IoT devices because both the regulatory and liability issues mentioned before are caused by the interconnection of several factors. The first factor concerns the ties between regulation and interoperability which might give rise to incentives to design IoT objects that are the most interoperable possible, regardless of their functions. Scholars have already spotted a pattern in the EU digital policy where interoperability seems to be one of the most sought-after results by mentioning and requiring harmonized standards, common specifications, and self-certification schemes.Footnote ⁷ Soon, the substantial common core of these definitions could play a role in eroding the distinction between consumer IoT and medical devices as the ways of functioning and operating between the two might become de facto the same. As a secondary effect, it could also play a role in making the distinction between “traditional/ex-post” liability and “ex-ante” risk management and compliance more nuanced (see infra in 1.3).

The second factor to keep in mind while explaining the future rise of mixed-functions IoT objects concerns technology development boosted by the pandemic. As an example of this, let us think about how the most acute phase of the COVID-19 pandemic accelerated innovation and technologies that had already been present for yearsFootnote ⁸ : telemedicine and health monitoring became more widespread in just some months’ time.Footnote ⁹ Also, there was an increase in making homes smarter, by relying on integrated voice assistants which could be the interface with the security of the home (e.g. smart cameras and alarms) but also energy (e.g. thermostats, shutters, and fridge controls).Footnote ¹⁰ These first two markets are both boomingFootnote ¹¹ and there might be a space for a new market which connects them.

In fact, this new generation of medical devices and the smart-home IoTs often need a link (another IoT device most of the time) to work together. The solution is building mixed-functions IoT objects which could play mostly an intermediary role between smart home appliances (e.g. voice assistants for the home) and the new generation of medical devices (e.g. medical IoT such as monitoring vests). However, there could also exist mixed-functions IoTs which could be an interface with a doctor or a medical structure, such as the case of the smartwatch I mentioned before. This technological evolution is consistent with Mark Weiser’s idea of “Ubiquitous Computing,” which is a state in which computing power and applications are so close to us that are almost invisible.Footnote ¹² This idea was also applied to the home environment in the early 2000s.Footnote ¹³ The phenomenon of technological convergence, which means the combination of different technologies,Footnote ¹⁴ has been blurring the boundaries among the latest years’ cutting-edge technologies: Artificial Intelligence (AI) systems work in the cloud to analyse and react to the inputs collected at a device level from an IoT object. An example is how we interact with an integrated voice assistant such as Google Home, Alexa or Bixbi: we give a voice command; the speaker (sensor level) collects personal dataFootnote ¹⁵ and sends them to a proprietary cloud where complex algorithms translate the input and elaborate a reply.

The third factor to consider is partially a consequence of the previous two: allowing and incentivising interoperability between different IoT objects by regulation, increasing development of smart-home and smart-rehabilitation objects and programs, and, most importantly, increasing the importance of the interface part of mixed-functions IoT objects, will make the traditional distinction between medical and consumer IoT devices more and more unclear, if not useless. In a few years, medical device developers might be tempted to make their e-health solutions interoperable with the IoT objects for the home. Moreover, even creators of IoT applications for the home or mixed-functions IoT manufacturers might be inclined to embed more developed e-health set of functions in their connected objects, which could be used in conjunction with a proper and technologically advanced medical device. Theoretically, this is the European Health Data Space’s (EHDS) field of application.Footnote ¹⁶ In this way, even the patient/consumer might be facilitated in making their home also an effective rehabilitation centre. This process might already be in its early days: in 2022, Amazon bought a chain of private medical clinics and it is also one of the most famous manufacturers of IoT products for the home.Footnote ¹⁷ It is realistic to expect that Amazon’s Alexa might become interoperable with private telemedicine applications or IoT medical devices developed by or for the chain of medical clinics, thus making mixed-functions IoT devices more popular and more widespread. In the EU, the same thing might happen, but things might different because of the current strict division between medical devices and consumer objects rules.

1.3. Mixed-functions IoT and private law liability rules

In the previous subsections, it was clarified that mixed-functions IoT devices are always consumers’ objects according to today’s EU regulatory approach. Hence, in terms of EU tort liability, if manufacturers do not comply, the rules on strict liability (the product liability directiveFootnote ¹⁸ and, prospectively, its update, the PLDU) and the future AI liability directive proposal (AILDP) will apply. It is still unclear whether these two regimes are fit for mixed-functions IoT and their fast-paced evolution (more infra in 3.1 and 3.2).

However, if we understand liability with a larger meaning to encompass accountability as well, it will be impossible not to mention, even briefly, the multi-layered structure of compliance duties which mixed-functions IoT manufacturers will need to face ex-ante and for which they will be held accountable and, eventually, liable in case of damages to consumers (infra 2).Footnote ¹⁹

The ensuing structure of the article will start from the two main assumptions stated in this first section: if it is true that interoperability and technological convergence are making distinctions between IoT as medical devices and consumer objects less clear, then it might be useful to know what the main regulatory acts applicable ex-ante to mixed functions IoT objects are. I am convinced that only by mapping down the legislative acts that a manufacturer needs to be compliant with ex-ante (sect. 2), it will be possible to clarify what consequences there will be in terms of ex-post private liability rules (sect. 3). From the combination of these concise regulatory and legislative maps, I hope that the EU legislators might take notice of the gaps that will be likely to be apparent in the upcoming years and that will concern common-use objects with a hybrid nature (sect. 4).

2.1 AI regulation

On 13 March 2024, the EU institutions adopted an agreement concerning a horizontal and general regulation aiming at the different forms of AI that are being developed and the ones that have existed for quite some time. This proposal was known since the beginning as the AI Act (AIA).Footnote ²⁰ The main rationale underpinning this regulation is two-fold. On the one hand, there is a fundamental rights protection rationale from also the legitimate uses of AI systems.Footnote ²¹ On the other hand, as in the GDPR, the AI act follows a risk management rationale.Footnote ²² This means that the higher the risk for EU citizens is from the AI systems employment, the more the AI systems providers will need to comply with ex-ante duties such as the certification of high-risk AI systems with specialised audit and certification bodies, called Notified Bodies,Footnote ²³ and the need to train data by respecting the transparency principleFootnote ²⁴ or by ensuring a meaningful “human oversight.”Footnote ²⁵

The AIA relevance for the IoT objects, including the mixed-IoT is apparent from the provision referring also to “product manufacturers placing on the market or putting into service an AI system together with their product and under their own name or trademark.”Footnote ²⁶

It is indeed relevant to mention the new categories of AI systems as some of mixed-functions IoTs might already use one or more kinds of AI systems. The new AIA regulates AI systems (primarily intended as software)Footnote ²⁷ in three categories: prohibited AI systems,Footnote ²⁸ high-risk AI systems,Footnote ²⁹ and general-purpose AI (GPAIs).Footnote ³⁰ There is also a not-explicitly mentioned fourth category which comprehends all the AI systems that do not fit in the previously mentioned category and that are considered low-risk (more infra). GPAIs is a category that refers to foundation models which are mainly known as the technology behind chatbots such as Chat-GPT or Gemini which was the core of the AI act negotiations from end of 2022 until the AIA approval. It is relevant to mixed-functions IoTs as many integrated voice assistants are now using foundation models to train and function. The AIA then further divides the GPAIs systemic risk and non-systemic risks GPAIs with different compliance burdens depending on the ability to generate a systemic risk.Footnote ³¹ It is not yet clear however how we can tell if a GPAI can cause a systemic risk given their “general” and “unfinished” core which makes them able through the training and scaling technique to adapt to different tasksFootnote ³² . This means that there is no reasonable way to exactly know in advance whether a GPAI can cause a widespread risk to fundamental rights. Both the reasonable foreseeability and the violation of fundamental rights are the two conditions requested by the “systemic risk”Footnote ³³ definition.

As far as prohibited practices high-risk AI systems and GPAI, the AIA assigns precise duties to AI providers,Footnote ³⁴ the actors who develop AI systems, and deployers, the people who employ those AI systems,Footnote ³⁵ depending on the risk generated by the AI system or model. Moreover, the AIA builds a complex transnational and EU governance system to govern the most impactful of those AI systems.Footnote ³⁶ Some uncertainty is still present as far as the compliance rules to apply for AI systems not falling explicitly into the previously described systems. Specifically, I am referring especially to the category of “low-risk” AI systems which was present in the first 2021 AI proposal but was subsequently “forgotten” from the AIA negotiations and the introduction of GPAIFootnote ³⁷ really stole the negotiations spotlight.

However, Article 4 AIA on AI literacy is a general, horizontal norm that applies to all kinds of AI systems and models and states several obligations that AI providers must be accountable for by instructing their staff and the AI system’s affected people of the risks that their kind of AI might cause and which rights they do have if they sustain damage.Footnote ³⁸

2.2. Product safety laws

The second group of laws that manufacturers will need to deal with is the one concerning product safety. There will be three regulations to discuss, from the most general to the most specific.

The more general one is the recently approved General Product Safety Regulation (GPSR) which will become applicable from 13 December 2024.Footnote ³⁹ It is important for mixed-functions IoT manufacturers as it considers as part of its applications field IoT objects – in its “product” definitionFootnote ⁴⁰ – as well as online marketplaces that sell them.Footnote ⁴¹ This new regulation is inspired by the administrative system created by the previous General Product Safety Directive (GPSD)Footnote ⁴² but includes some innovations, most notably a new recall system and the provision of traceability requirements for objects.Footnote ⁴³ In sum, the manufacturers’ common goal is to get the EU CE conformity label both under the GPSD and the GPSR. Moreover, in the GPSR there are complex and extended obligations falling on all the economic operators involved in the mixed-functions IoT objects production chain. This way of including all the economic operators involved is not exclusive of the GPSR but also of the Machinery Regulation (infra) and it started in the MDR which is the oldest and most specialised of these regulations (infra). The rationale for having so many actors with duties is not to leave accountability gaps along the product and value chain and this is why the PLDU mirrors this scheme in Article 7 PLDU from a liability point of view (see 3.1). In fact, in the GPSR, manufacturersFootnote ⁴⁴ have specific obligations such as the obligation to carry out a risk analysis of the marketed product.Footnote ⁴⁵ Furthermore, the other economic operators have specific obligations such as authorized representatives,Footnote ⁴⁶ importers,Footnote ⁴⁷ and distributorsFootnote ⁴⁸ as well as online marketplaces.Footnote ⁴⁹ Some rules shift manufacturers’ obligations onto other economic operators in specific casesFootnote ⁵⁰ , in order not to leave accountability gaps. In addition, the application to IoTs in general, and to mixed-functions IoTs in particular, is highlighted in Article 6 GPSR which sets a non-exhaustive list concerning criteria on how to evaluate the safety of the product and explicitly mentions the effect that the product might have on other ones also because of their interconnection.Footnote ⁵¹

The recent approval of the AIA act demonstrated what it had only been presumed during its negotiations: that there is a complementarity relationship between the AIA and the GPSR for the residual categories of the once called “low-risk AI systems”Footnote ⁵² and that is apparent by reading Recital 166 AIA. In the AIA, the only non-binding but explicit obligation is to draw codes of conduct inspired by the ethical principles of the High-Level Expert Group’s “Ethical Guidelines for a trustworthy AI” principles for those kinds of algorithms.Footnote ⁵³ The recently approved AIA confirms this as it uses the term “safety net” referring to the GPSR concerning other than high-risk AI systems which do not have a specific regime to follow. What would that mean in practice? It seems that the principles guiding the development non-high-risk AI systems will still be in the AIAFootnote ⁵⁴ , mainly in Article 4 on AI literacy and corresponding AI recitals as well as Article 70 AIA which concerns AI national authorities. However, for the safety and market surveillance obligations, AI systems providers (and, most likely mixed- functions IoT manufacturers) will need to consider the GPSR. Hence low-risk AI-powered mixed-functions IoT manufacturers will need to keep in mind this two-pronged system (AIA and GPSR) as far as compliance and accountability are concerned.

A more specific regulation in this group is the recently approved Machinery Regulation (MR) which can applyFootnote ⁵⁵ to mixed functions IoT objects and their components if they are of the same kind as the ones of the list described in Article 2 and Annex II MR, including software as a safety component.Footnote ⁵⁶ In addition to that, the MR rules applicable to mixed-functions IoT devices might also need to coordinate with Article 6 and Annex I (1) of the AIA. This AIA Article concerns high-risk AI systems which are either safety components or are AI systems covered by Annex I of the EU harmonization legislation and are required to undergo a third-party conformity assessment. The MR is part of those harmonization regulations contained in Annex I AIA. Reciprocally, Annex 2 (19) MR, mentions “…Safety components with fully or partially self-evolving behaviour using machine learning approaches ensuring safety functions.” In some specific cases, meaning when software is a safety component then, probably, the MR will need to be applied in conjunction with the AIA compliance discipline of high-risk AI systems if the software as a safety component falls within the parameters of Article 6 (1) and Annex I of the AIA. As for low-risk AI and the GPSR, it might be difficult to coordinate with the rather specific and complex MR discipline involving sometimes the same actors in the value chain.

Another element of similarity among these three legislations, including the MR, concerns the attribution of duties to economic operators: manufacturer, authorized representative, importer, and distributor.Footnote ⁵⁷ Each of these subjects has its own duties, the most complex of which belong to the manufacturers who need to create the product which must respect the basic health and safety conformity requirements and send the necessary technical documentation which can undergo a third-party conformity assessment depending on the category and level of risk of the machinery.Footnote ⁵⁸

In 1.1, I explained that the manufacturer is the subject responsible for the certification strategy of their products, and they should do that by verifying the product safety rules that are applicable to their product. In case the manufacturer does not opt for the “general safety” laws as they believe the IoT object has medical functions and not only an e-health one, then the path to follow is the one set in the MDR. Even if the MDR field of application is specific and, regarding IoT objects, it only concerns medical IoT objects, which I will not analyse here, it also gives a little guidance concerning mixed-functions IoT objects. If one reads in conjunction Articles 2(2) and 2(11) MDR they come across the meaning of accessory of a medical device. More precisely, it can be a device or software (e.g. a smartwatch or a function for an integrated vocal assistant) which ‘… whilst not being itself a medical device, is intended by its manufacturer to be used together with one or several particular medical device(s) to specifically enable the medical device(s) to be used in accordance with its/their intended purpose(s) or to specifically and directly assist the medical functionality of the medical device(s) in terms of its/their intended purpose(s).’Footnote ⁵⁹ If we want to expand even more the picture, Article 2(11) MDR explores the notion of system which is a combination of products, either packaged together or not, which are intended to be interconnected or combined to achieve a specific medical purpose. Nevertheless, the rigid distinction between medical and mixed-functions IoT still holds: Article 22(c) MDR provides that the other objects used in conjunction with one or more medical devices must bear a CE certification that is obtained through the processes described in the GPSD. This means that the GPSR will soon fulfil the GPSD role. Despite the MDR is not directly applicable to the mixed-functions IoT, it is important to point out that the MDR mentions non-medical devices used in conjunction with medical devices, such as software as a medical device, as it reinforces the hypothesis that in the near future medical IoT devices and mixed-functions IoT devices might work together although having a different regulatory discipline.

3.1. The new Product Liability Directive Update (PLDU) and the mixed-functions IoT

Concerning the first criteria of evaluation, the still applicable Product Liability Directive (PLD)Footnote ⁶⁰ and PLDU (after its approval) might be the most relevant harmonized liability systems for the mixed-function IoT objects for several reasons. The first reason is that this kind of IoT devices are considered part of a larger consumer objects category under EU law, hence they will respond to this kind of liability even today when the PLD is still applicable. Moreover, as recalled in the explanatory memorandum for the PLDU, product liability is a gap filler for those EU legislative proposals and acts that might have a more decisive regulatory character the Machinery Regulation, and the General Product Safety Regulation.Footnote ⁶¹ In those acts, liability is not addressed. Hence, if damage occurs, the PLDU will be applicable. To sum up, the gap-filling functionFootnote ⁶² will be the ultimate scope of EU product liability.

The process of the PLD review for the REFITFootnote ⁶³ review of the EU legal acts in 2018 crossed paths with the scholarly debate on the fitness of the PLD for the digital age.Footnote ⁶⁴ In the end, the idea was to update the PLD with the consolidated Court of Justice of the EU (CJEU) jurisprudence and the changes brought by new digital technologies at the same time. That is why it is relevant to focus on the new PLD Update (PLDU) proposal which was made public last 28 September 2022 together with the AILP (infra 3.2).

The PLDU functioning is not completely different from the current PLD one as the complainant needs to prove the damage, the defectiveness, and the causal link between the first two elements.Footnote ⁶⁵ However, the application of the new PLDU will be especially interesting for the mixed-function IoTs for several reasons. Firstly, it will be a general regime that has the role of covering the mixed-functions IoT, even when they are part of a medical device in the form of ancillary products or as part of a system of medical devices according to the MDR. At this moment, both medical devices and consumer IoTs need the PLDU application whenever the kind of damage covered by this directive proposal takes place. Secondly, what is interesting to notice is that the new PLDU proposal is similar to the MDR under a terminology aspect. For instance, the “producer” has disappeared and has been substituted by the “manufacturer.”Footnote ⁶⁶ Moreover, other terms such as economic operators, or authorized market representatives are more in line with the vocabulary of the MDR and the product safety laws (see sect. 2.3).

It is interesting to point out that the concept of a product (which can become defective) extends explicitly beyond the tangible part of the object. In fact, in Article 4(1) PLDU the definition of product comprehends software but also digital manufacturing files and components, the latter ones must be with the manufacturer’s control, as well as the related services.Footnote ⁶⁷ The PLDU describes these last elements by stressing their function of integration and interconnection but also by pointing out that the lack of said services would prevent the product from performing one or more of its functions. Furthermore, for the first time, the definition of damage encompasses not only physical damage (including psychological damage) and physical property but also the loss or corruption of data.Footnote ⁶⁸ This is likely to result in one of the most common damages that could be caused by using mixed-functions IoTs. It will be interesting to analyse how the different national judges will find a way to effectively measure the damage created by data loss according to their national laws and the interaction of data loss damage with damage reparation rules set in Article 82 GDPR.Footnote ⁶⁹ The importance of data and software in general over the hardware part of products is also reflected by Article 10(2) PLDU on liability exemptions. Whenever software or its absence, or a related service caused the damage the manufacturer cannot state that the defect did not exist at the time of putting into the market the product or the service.

One of the most interesting elements of the PLDU concerns its relationship between standards and liability. As in the current PLD, and maybe symbolically, Article 6 PLDU deals with defectiveness. From the letter of the provision, it seems that the safety paradigm stays the same, but its point of reference changes: the “public at large” instead of what a person can reasonably expect. One could wonder what the public could exactly expect from a mixed-functions IoT device, which can also work in connection with one or more medical devices. The perceptions of risk and safety would likely change in all the possible applications of mixed-functions and medical IoT devices working together or alone. The defectiveness indicators list that follows in Article 6 PLDU is likely to be interpreted, as the one in current Article 6 PLD, as not a conclusive and final one.Footnote ⁷⁰ What is important to notice is that Article 6(f) and (g) PLDU consider as one of the elements to evaluate defectiveness the non-compliance with product safety requirements, including safety-relevant cybersecurity requirements. This is essential given the space that is left to standards and common specifications not just in the MDR but also in the AIA and the product safety regulations. It is a good sign that the PLDU does not mention any of the regulations for cybersecurity, data-sharing, and safety. In fact, the contrary will most likely make the PLDU grow old before its time.

Moreover, it is necessary to highlight the importance of Article 8 PLDU on evidence disclosure and of Article 9 (2–5) PLDU concerning the defectiveness and causal link (rebuttable) presumptions. As in Articles 3 and 4 AILP, Articles 8 and 9 are PLDU are entwined, and they are the result of EU experts report about the impact that AI and new technologies have on liability.Footnote ⁷¹ Article 8 allows claimants to ask judges to require manufacturers to show how their product works on the condition that enough facts support this request for compensation. At the same time, the intellectual property rights of the manufacturer must be somehow protected. Article 9(2) sets three non-cumulative conditions to prove the defectiveness of the product which are the following: (i) the defendant has failed to comply with Article 8 evidence disclosure obligation (ii) the claimant demonstrates that the product does not comply with mandatory safety requirements laid down in Union law or national law that are intended to protect against the risk of the damage that has occurred; (iii) or, the claimant demonstrates that the damage was caused by an obvious malfunction of the product during normal use or under ordinary circumstances. Article 9(3) introduces a rebuttable presumption by stating that if the damage is consistent with a certain cause a causal link can be presumed. Ultimately, the judge can freely decide, because of technical difficulties experienced by the complainant, to presume either/both the causal link or/and defectiveness provided that the complainant proves that the product contributed to the damage and it is likely that its defectiveness was likely to have caused the damage.Footnote ⁷² All these presumptions might be interpreted more in favour of consumers than one might expect, as judges might feel more sympathetic towards a consumer who does not know how a seemingly straightforward, but complex, mixed-function IoT device might work.

Despite these premises, the future enforceability of the PLDU could create problems for consumers. One of the underlying issues in Article 7 PLDU is the lack of flexibility in the liability regime applicable to the different economic operators involved in the production of technological products, such as the mixed-functions IoT objects. In theory, the directive’s rationale is pro-consumer as the consumer must always have an EU-based subject they could address.Footnote ⁷³ The problem is that the first subject the consumer needs to contact in case of damages is always the manufacturer which could also be not EU-based in case it is a technological object such as a mixed-functions IoT. That might often be the case for technological products. Besides, Article 7 PLDU provides for a strict subjects list (from the further one to the nearer one to the consumer) to ask for compensation. Such order is understandable under a regulatory perspective for accountability reasons but might discourage consumers from taking advantage of this kind of liability because of the length of this list, the time, the costs that it might entail and, finally, the risk of being time-barred.Footnote ⁷⁴

3.2. The proposed AI liability directive (AILDP)

In terms of liability schemes, mixed-function IoT devices could also be potentially subject to the proposed AI non-contractual liability directive. This is a fault-based liability directive although it does not harmonize the concept of fault directly. It does that by exemplifying breaches of duties of care when it comes to AI development and through legal presumptions concerning the causal link between the damage and the AI system’s functioning.Footnote ⁷⁵ It applies to mixed-functions IoT because the technologies concerned (the IoT and the AI) are already working – if not within the same physical device – at least in combination through sensors and cloud (mostly proprietary) networks. This way of operating is the same both for mixed-function IoT devices and for certified IoT medical devices.

The AILDP will respond to non-contractual kind of claims caused by AI. Given that the damage from software needs to be considered also as falling within the PLDU regimeFootnote ⁷⁶ , the AILDP will cover only the claims in which damage is caused by software that:

(i) could be defined as AI system by the new text of the AIA and

(ii) injures, in a material and non-material way depending on the tort national rules, a party not involved in the AI system creation and deployment which I will generally refer to here as third party.

For instance, I use a visor with augmented reality to do rehabilitation, but the object is an IoT which is not certified as a medical device. However, this visor works in collaboration with the hospital and in this way, the doctor knows that I am doing physical exercise at a number of times per day. A friend comes to my place and tries on the visor. The problem is that the virtual reality landscape they see is incorrect. Hence, they jump from the sofa, and they break their ankle. Leaving aside for the moment my responsibility in leaving that object unattended, the friend could sue the IoT manufacturer (which could also be the AI provider) by using the national implementation of the AILDP. In this case case, the national rules concerning general non-contractual liability will be harmonized with the presumptions contained in Article 4 AILDP and the new procedural rules concerning evidence disclosure in Article 3 AILDP.

Article 4 AILDP is particularly interesting as three conditions must be satisfied by the claimant for the national courts to presume a causal link. The first one is to prove the fault of the defendant, the second one is that it must considered likely that the defendant’s fault was the cause of the AI-induced damage, and the third one is to demonstrate that the output produced by the AI system or the failure of the AI system to produce an output gave rise to the damage.Footnote ⁷⁷ However, if the system is high-risk according to the AIA, the first condition could be presumed if the claimant manages to prove at least one of the breaches of duty of care set in Article 4(2). These duties concern, among others, duties of transparency and fairness of data training to make just two examples. However, this will not apply when the defendant proves that the state of the art makes it possible to prove causality without using Article 4(1) AILDP presumptions for high-risk AI systems.Footnote ⁷⁸

This analysis of the AILDP suffers from the interconnection of this legislation with the AIA first proposal and might be subject to changes quite soon. When the AILDP proposal was presented, it referred to a version of the AI act that is quite different from the one that was approved recently. Hence, this analysis might become outdated too in a short time. There is no mention of GPAIs in the AILDP and one can ask if they can be assimilated to high-risk AI systems in terms of AILDP presumptions especially if we consider GPAI with systemic risk or that have an opaque way of functioning and AI deployers/implementers do not know how to mitigate risks which might injure an undetermined but high number of people.

Even if we consider that the AILDP will not change (or change much) from the text that we are discussing today, I think that its application might be residual even for the national judges’ law practice, specifically for AI systems that correspond to the old “low-risk” AI systems and which fall in the general Article 4(1) AILDP scheme. These low-risk AI systems are relevant as they are most systems that power mixed-IoT objects today. One problem for the complainant’s lawyers could be the following: for most of them, it will be difficult to demonstrate that there is no other way to prove causation than to use Article 3 AILDP and gain access to how the algorithms work. Nevertheless, lawyers do not have the skills to understand how AI works, unless serious training is organized. But, even with training, it will be complex to find expert AI consultants in the short run. However, a more complainant friendly way to interpret Article 4(1) is through Article 4(5) which states that for low-risk AI systems judges can “only apply where the national court considers it excessively difficult for the claimant to prove the causal link mentioned in paragraph 1.” This cannot work if the complainant uses the mixed-functions IoT is for a personal or nonprofessional activity according to Article 4(6) AILDP.

Even if the complainant’s lawyer manages to prove through Article 3 AILDP the condition set at Article 4(1)(a), there are still other two conditions to prove together with the fault of the AI provider. In sum, the new harmonized general tort rules might not be as favourable to injured people as one might think they are. A liability system indeed needs to strike a balance between the different actors involved. However, in this case, it seems too tilted in favour of AI providers/manufacturers. They are the subjects that do know best how AI systems work.

What is going to be likely is that for mixed-IoT objects, if the AILDP needs to be applied, is that national judges will try to interpret the AILDP provisions more favourably towards the injured party, especially if the AI system is complex, independently from the level of risk and maybe by taking inspiration by Article 9 PLDU presumptions and/or by adding national implementation rules on the matter. This would follow the interpretation of Article 1(4) AILDP which allows Member States to adopt or maintain national rules that are more favourable towards injured people if they are compatible with EU law. It seems that only Recital 14 AILDP suggests that this directive should be interpreted as a minimum harmonization directive, but this indication is not present in the main text. This poses serious concerns because what would be the rationale of having a minimum harmonization directive for AI non-contractual liability if every member state can create EU law-compliant and yet more complainant-friendly AI tort rules when this proposal’s legal basis wants to avoid excessive market and legal fragmentation?

If there will be a change following the AIA approval in the AILDP structure, it will be a good occasion also to clarify this last point concerning a clearer division between the EU and national liability regimes, and making it explicit that it is a minimum harmonization directive, as well as whether or not to include presumptions for GPAI models and systems.