ERU, Alm.del - 2019-20 - Bilag 327: Det danske høringssvar til Kommissionens meddelelse Om kunstig intelligens: En europæisk tilgang til ekspertise tillid (COM (2020) 65 final), samt overliggernotat herfor, fra erhvervsministeren

Erhvervsudvalget 2019-20
ERU Alm.del Bilag 327
Offentligt

NOTAT

Annex: Specific comments from the Danish Government on the white

paper on artificial intelligence

Specific comments

An ecosystem of trust: Regulatory framework for AI

The overall aim must be to create a regulatory framework where trustwor-

thy, ethical, safe and secure AI goes hand in hand with the ability to provide

innovative solutions. This regulatory exercise must also include stocktak-

ing of existing legislation in order to make sure that existing legislation is

up to date and is able to address specific issues related to AI. A stocktaking

process should also ensure that potential new legislation does not overlap

with existing requirements.

The General Data Protection Regulation (GDPR) has illustrated how the

EU has been able to set global standards. However, the subsequent demand

for guidance on how to adhere to the new standards showed the importance

of legal clarity and user centric guidance. Therefore, clear and operable

regulation is a precondition and must be factored in from the outset when

establishing a regulatory framework for AI.

The scope for a new EU regulatory framework for AI

Where certain situations related to serious risks to individuals or to society

stemming from the use of AI are not best tackled by existing legislation,

the Danish Government finds it appropriate to address such risks in a new

risk-based regulatory framework at the European level, while taking into

consideration Member States’ competences in specific sectors.

This regu-

latory framework should address serious risks in relation to transgressions

of fundamental rights such as discriminating decision-making as well as

risks of infliction of injuries, especially where these may be irreversible.

Certain AI applications can be applied in critical sectors, which may in-

volve serious risks of societal or individual significance, thereby catego-

rising such application as high-risk. At the same time, it is important to

consider that AI is also being developed and applied in critical sectors for

the benefit of health, welfare and public services. Therefore, it is essential

to find the right balance in the risk-based regulatory framework between

minimizing risks and facilitating the development and uptake of new solu-

tions for societal challenges.

Defining high-risk AI

It is crucial that the definition of high-risk AI is clearly limited to applica-

tions, which can actually cause serious risks, as it is imperative that un-

problematic applications of the technology are not unnecessarily limited to

ERU, Alm.del - 2019-20 - Bilag 327: Det danske høringssvar til Kommissionens meddelelse Om kunstig intelligens: En europæisk tilgang til ekspertise tillid (COM (2020) 65 final), samt overliggernotat herfor, fra erhvervsministeren

the detriment of innovation, the development of new business models as

well as the creation of future jobs. Therefore, essential building blocks for

this new framework have to be a clear and targeted definition of high-risk

AI and provide businesses and public authorities with proper guidance on

how high-risk AI is defined in practice.

Firstly, such a definition should entail a clear definition of AI itself. The

definition of AI will need to be sufficiently flexible in order to accommo-

date technical developments, while being precise enough to provide the

necessary legal certainty. However, there is a need to elaborate further on

the definition set out in the white paper, as it is still uncertain which kinds

of AI is covered. The properties of AI as currently defined ranges from

statistical models to neural networks, whereas in our view, the former

should not be covered by the definition AI.

Secondly, the aim must be to identify the AI applications, which constitute

serious risks and determine how serious risks are defined. Such an identi-

fication should result in making the category of high-risk AI the exception

rather than the rule. With this in mind, the two cumulative criteria set out

in the white paper constitutes a good starting point, as an exhaustive list of

sectors combined with the application itself could provide the necessary

predictability and delimitation. However, this will furthermore depend on

the elaboration of the two criteria. The exhaustive list of sectors must cover

a limited set of sectors and the definition of application should be clarified.

This list should also be accompanied by clear guidance, as the term "appli-

cation" could cover many different aspects such as the process as well as

the end purpose. In order to bring clarity and cover the actual high-risk AI,

we would give greater weight to the end purpose of the application, as this

would characterize the potential outcome and thereby the risk more pre-

cisely.

In continuation hereof, applying the two cumulative criteria could never-

theless bring about different grey zones as well as overlook nuances with

respect to serious risk. Therefore, greater emphasis should be put on a-

built-in risk assessment, which for example should assess both individual

and societal risks as well as assessing the probability of the actual risk.

However, such a risk assessment should not in itself create legal uncer-

tainty or bureaucracy.

Furthermore, the role of human oversight already incorporated in the ap-

plication should be taken into account in the risk assessment. There should

be a difference in terms of adhering to a stricter regulatory framework,

when an application takes autonomous decisions compared to the situation

where the application constitutes a support tool for humans in the decision-

making process.

Lastly, we are wary of the exceptional instances, where the use of an AI

application for certain purposes is to be considered as high-risk irrespective

of the sector, as such an approach could undermine the legal certainty.

There could be certain applications with a horizontal nature where such an

exceptional category would make sense, but it must be composed of an

exhaustive and limited list of specific applications.

Requirements for high-risk AI

The key features set out in the white paper are a good basis for identifying

the relevant requirements for high-risk AI. An essential element of the fu-

ture discussion is the operationalization of them. In this context, it is im-

portant to build on the existing work concerning the guidelines for trust-

worthy AI and take into account the experiences stemming from their feed-

back process. As stated by the Commission, one of the key results of the

feedback process is that a number of the requirements are already reflected

in existing legal or regulatory regimes. As the requirements should com-

plement, but not overlap with existing legislation, such results must be

taken into account, especially as it is likely that high-risk AI applications

are also covered by sector-specific legislation.

In order to not unnecessarily hamper innovation and future solutions within

the area of high-risk AI, any new requirements imposed must be propor-

tionate, technology neutral and be able to address the specific risks associ-

ated with the AI application.

Furthermore, the Commission should incorporate, to the utmost extent pos-

sible, technical standards on AI as well as data technology in general, since

standards have proven to be a flexible way to regulate a field including

rapid technological developments. ISO/IEC and CEN/CENELEC as well

as the IEEE are currently working on different aspects of standardization

of AI and ethics/trust.



Training data

As AI is only as useful, as the data, which it is trained upon, it is essential

to set tangible data requirements for the development and use of high-risk

AI. These should be easy to interpret, and therefore it is necessary to elab-

orate as to what constitutes “sufficiently representative” training data with

regards to dimensions of gender, ethnicity and other possible grounds for

bias or discrimination as well as those of privacy and safety. Such require-

ments should not inhibit the innovation of AI, which does not touch upon

these dimensions.



Data and record-keeping

An essential part of AI oversight is the ability to review the decisions taken.

A complete transparency requirement might not be technically possible

and could in some cases be a de facto hindrance for innovation. Instead, it

should be required that certain information about the AI model such as

records regarding the data used for training as well as the techniques used

for developing the system is stored. These obtainable inputs will enable

explanation of the output performed by the model and thereby enhance the

transparency.

In terms of storing significant characteristics of datasets, it is vital that

these characteristics rely on existing standards prevalent in the market in

order to avoid imposing potential transition costs on authorities and busi-

nesses who are already using common metadata standards when describing

data.

As to not create unnecessary burdens, it is important to establish strict re-

taining limits as well as what specific information is necessary in order to

explain the AI’s output.

At the same time, this requirement should not lead

to a situation, where developers and deployers of high-risk AI are required

to publish data as well as algorithms. The objective should be to give the

competent authorities, if needed, the adequate insight in order to verify

whether applicable rules and requirements are complied with.

Furthermore, it is essential to take into account the requirements already

set out in the GDPR as well as to specify how developers and deployers of

AI are able to collect, store and use data for AI and still be GDPR-compli-

ant. In this aspect, it should be considered, if GDPR is a sufficient legal

framework in order to avoid creating new legislative burdens.



Information to be provided

Proactive information

related to the AI system’s capabilities and limita-

tions is essential, as such information could raise the understanding of the

system as well as how it should be deployed. It could furthermore help in

cases concerning liability, for example if a system is utilised in a way,

which contradicts its capabilities and limitations, which the deployer was

made aware of in advance. However, as to not lay unnecessary burdens

upon developers, it is important to establish what specific information is

needed

in order to explain the system’s capabilities as well as limitations.

When it comes to a requirement about informing users about the interaction

with an AI system, it is unclear whether such a requirement in practice will

empower citizens. The GDPR already sets out requirements, when per-

sonal data are obtained, concerning the existence of automated decision-

making. Furthermore, despite this requirement, the deployer or developer

of the AI

–

including public authorities - will still be responsible for the

output and citizens will still be able to challenge the output. Lastly, another

potential requirement for high-risk AI involves human oversight, which

seems to overlap with this particular requirement. Transparency should be

our point of departure in order to promote trust, but the aim should be to

enhance transparency by giving citizens a meaningful choice and/or giving

the possibility for the output to be put to the test.



Robustness and accuracy

For AI to be trustworthy, it requires both robustness and accuracy of the

application. However, in our view, both aspects require further elaboration,

also in terms of the specific definitions of robustness and accuracy. One of

the most important questions in terms of establishing such new require-

ments is whether AI must be more precise than humans, for example

should autonomous vehicles be more precise than humans driving tradi-

tional ones? In order to minimise risks, but without hindering innovation,

it is essential to find the proper balance in this question.

The issue of security is partly covered by the requirements on robustness

and accuracy. Given the importance of cybersecurity in relation to AI and

other data-based technologies, the item should be elaborated. Denmark

supports a horizontal approach to cybersecurity, where existing legislation

regarding cybersecurity would be updated, if necessary, in order to address

potential risks in relation to AI

–

instead of establishing a sectorial ap-

proach for AI in relation to cybersecurity.



Human oversight

When categorised as high-risk AI, there must be a requirement to have an

appropriate involvement of human oversight in the specific AI application.

One example could be when applying AI to diagnose patients or to choose

medical treatment. In such cases, human oversight must be required. For

example, a doctor must have the final say in the beforementioned matters,

thereby making the AI application a support tool for the doctor in order to

qualify the decision-making process.

As the AI applications within the category of high-risk AI may be a diverse

group, there will not be one size fits all-model in terms of human oversight,

meaning that human oversight could manifest itself in different forms. The

appropriate involvement of human oversight - type as well as degree

–

should be proportionate and take into account the specificities of the AI

application in question. In such determination, it would be useful to bring

into play the results from the build-in-risk assessment used for establishing

whether the AI application in question was high-risk or not.

Specific requirements for certain particular AI applications used

for remote biometric identification

Usage of technology for remote biometric identification calls for a very

cautious approach, as it must neither result in general surveillance of our



citizens nor erode fundamental rights as well as existing EU and national

legislation. At the same time, we acknowledge that certain, but strictly lim-

ited cases may justify its deployment for example in terms of the preven-

tion, investigation and prosecution of crimes. However, it is important that

we strike the right balance between conflicting considerations. We need to

find the right balance between fundamental rights and the prevention, in-

vestigation and prosecution of crimes. Biometric remote identification

must always have a clear purpose, be proportionate, allowed for a strictly

limited period as well as subject to adequate safeguards. As to whether

further guidelines and regulations are needed in this respect, we will re-

serve the right to form our position, once such a potential proposal has been

presented by the Commission.

Certification and control of high-risk AI

An efficient governance structure must strike the balance between the need

for a uniform application of the requirements and the need for avoiding

unreasonable burdens for developers and deployers of AI. An overarching

European governance structure should play a key role in facilitating the

implementation of the regulatory framework, as this is important to ensure

a consistent and harmonized application of the requirements related to AI.

Especially the responsibility of guidance as well as capacity to assist de-

velopers and deployers covered by the regulatory framework should be

clearly defined from the beginning.

The specific conformity assessment should be entrusted to notified bodies

designated by the Member States. This would increase assessment capac-

ity, as developers and deployers would have access to at least 27 different

notified bodies in order to obtain the conformity assessment. Furthermore,

sufficient capacity to certify AI products and services needs to be ensured

before entry into force of the regulatory framework.

Important considerations for the set-up of a conformity assessment should

be to avoid lengthy, burdensome processes as well as duplication with

other conformity assessments and tests in other fields such as medical de-

vices. Additionally, some specificities of the requirements are not relevant

for all types of AI products and services. The conformity assessment

should be able to take into consideration the different aspects of different

AI applications, focusing on those, which are relevant for the application

in question. Also, it could be considered if a conformity assessment could

be divided into different steps or similarly where an AI application in its

testing stage would not be obliged to go through a conformity assessment,

thereby giving room to innovate as well as minimizing the pressure on the

notified bodies.

It should be the responsibility of the developer or deployer of AI to assess

whether their applications are required to be certified. As grey areas are

unavoidable due to the continuous development of the technology, it is es-

sential to develop practical guidance tools in order to reduce legal uncer-

tainty. A practical tool could for example involve an algorithm assessment

tool, where developers and deployers could get immediate guidance on

whether they are subject to the requirements or not.

As AI is a fast-evolving technology with an algorithm, which is not always

static, recertification must be a requirement, if the AI application has been

significantly modified. In order to avoid a situation, where developers and

deployers find themselves in continuous conformity assessments, it should

be explored how to stipulate benchmarks for when recertification is needed

and whether the developer or deployer, which has already been assessed

could be able to have specific aspects assessed based on their prior assess-

ment.

Concerning SMEs, a conformity assessment could be a burdensome and

lengthy process. However, once a conformity assessment is in place and a

SME’s AI is categorized

as high-risk, the SME would for the most part be

requested to deliver a fully-fledged certification from buyers and users who

would otherwise turn to other certified providers. Therefore, an outright

exception for SMEs would not solve the issue. Instead, SMEs should be

given priority at the notified bodies and when establishing the Digital In-

novation Hubs, SMEs should be given assistance with their potential con-

formity assessment.

In order to ensure that the different requirements are complied with over

time and ensure consistent application of the requirements, compliance

must be monitored and controlled by the existing competent national au-

thorities, part of a market surveillance scheme. In our view, such control

does not call for a new enforcement setup, as AI is just one technology out

of many and would at the same time be controlled in connection with other

areas such as competition, product safety etc. However, there could be a

potential competency gap of the existing national authorities, which must

be addressed in order to be fully equipped to enforce the new requirements

related to high-risk AI.

Voluntary labelling for AI applications outside the high-risk category

As the general objective should be to enhance trustworthy AI - not only for

the AI which poses serious risks

–

the Danish Government strongly sup-

ports a European voluntary labelling scheme. In this regard, the Danish

Government will bring concrete experiences from the recent work with de-

veloping a Danish prototype of a data ethics seal as well as from the pri-

vately established Danish labelling scheme for IT security and data ethics

which is planned to launch by the end of 2020.

A European labelling scheme should promote human-centric and ethical

AI and data use in Europe by making it visible for consumers which com-

panies, products and services to trust and thus empower consumers to make

the ethical choice. A Danish study from 2019 showed that over 80 percent

of citizens would avoid shopping at places, if they suspected that their data

was not being processed responsible. At the same time, 65 percent of citi-

zens stated that their choice would be influenced by a labeling scheme. A

labelling scheme would therefore be a practical way of enhancing trust-

worthy AI for citizens as well as creating a push for making this a compet-

itive advantage for companies.

Scope of a European labelling scheme

The Danish Government supports the idea for a European labelling scheme

to target AI specifically. At the same time, we would underline the need

for the labelling scheme to cover broader issues in the digital economy

which citizens are concerned about in order to secure the relevance of the

labelling scheme, which should be demand-driven. Such issues would en-

compass for example the usage and storage of data, tackling cybersecurity

as well as ensuring unbiased decision-making processes, all of which are

essential parts of AI. The Danish study from 2019 also showed that citizens

do not differentiate between IT security and data ethics when using online

services. Based on these insights, we would advocate for a European label-

ling scheme, which both incorporates IT security as well as data ethics into

the scheme's criteria, in order to make the scheme relevant for citizens and

to avoid confusion.

Furthermore, it must be ensured that the scheme is not inconsistence with

corresponding regulation and requirements, for example the GDPR. How-

ever, the scheme should not be limited to the processing of personal data

or guaranteeing GDPR compliance and the scheme should therefore be an

addition to these.

Criteria for the scheme

The European labelling scheme should be rooted in a set of common crite-

ria based on the ethical guidelines for trustworthy AI. As a successful la-

belling scheme is based on a broad market uptake, there should be a differ-

ence between the very strict requirements for high-risk AI and the require-

ments in the labelling scheme, which would cover all other AI, even though

the requirements should be based on the same foundation. Such differenti-

ation should aim at not requiring all developers and deployers of AI to ad-

here de facto to the regulatory framework of high-risk AI.

The criteria could require that data ethics are anchored at the managerial

level of an organization, for example by requiring that businesses’ internal

policies for data ethics are included in their annual non-financial reporting.

The criteria should also address and be anchored at the technical level such

as fulfilling certain cybersecurity standards and ensuring that the use of

algorithms is explainable, transparent and unbiased.

As the category of AI outside the high-risk category would cover a diverse

group of developers and deployers of AI, it could be worth to consider a

differentiation of criteria, depending on a risk profiling. In the Danish pro-

totype, companies are categorized into four groups based on their organi-

zational complexity and data complexity. In order to obtain the data ethics

seal, each group is required to adhere to a set of different criteria, which

also vary in strictness. We would recommend for the European labelling

scheme to follow a similar approach.

Furthermore, it will be challenging for many developers and deployers in

the target group such as SMEs to comply with all the criteria. For this rea-

son, it could be considered to use a progressive disclosure model when

rolling out the labelling scheme, meaning that the applicants will not be

asked to comply with all the criteria at once. This strategy has multiple

purposes such as reducing complexity for the applicants, making it easier

to get started and thus increases the chance of buy-in, and giving the feeling

of progression by providing the applicants with the most relevant criteria

for its risk profile first and then increase the complexity-level progres-

sively.

Furthermore, the model and criteria should be flexible to reflect future

technological development and be revised continuously. In order to secure

trust in the labelling scheme, it should be voluntary to apply for the label,

but when first awarded, the criteria should be legally binding.

Governance of the scheme

An efficient governance structure for the scheme could be based on the set-

up for the EU cybersecurity certification framework. This entails giving an

EU agency the responsibility to provide the common criteria and certifica-

tion scheme for the label. An advisory group, which could be the high-level

expert group on AI could provide recommendations for drafting the com-

mon criteria and the Member States, should be responsible for appointing

national conformity assessment bodies. An alternative model could be to

anchor the governance of the labelling scheme in an advisory board and a

secretariat constructed by industry bodies.

Adjustments to existing EU legislation related to AI

The Danish Government supports a horizontal approach towards AI and

thus supports that the Commission also focuses on adjusting existing leg-

islation to address specific issues in relation to AI.

In this respect, the Danish Government welcomes a revision of the General

Product Safety Directive (GPSD) in order to ensure a set of minimum

standards for product safety in regard to AI embedded into consumer prod-

ucts. These should be complemented by updates to the existing sector spe-

cific legislation to address the specific needs pertaining to different cate-

gories of products. However, it is important that the sector specific regula-

tion is aligned with the general requirements in the GPSD in order to avoid

conflicting demands.

The Danish Government also supports the Commission’s proposed provi-

sions in relation to product safety in order to support responsible innova-

tion in AI technology, including a focus on the life cycle of products, qual-

ity of data, transparency and cooperation among economic operators and

public authorities. Additionally, the regulatory framework for product

safety should also address possible changes in consumer behaviour, when

interacting with products embedded with AI technology and ensure that

existing safety requirements - such as physical or mechanical means to de-

activate a product - is not compromised. Finally, updates to the regulation

must also address resilience towards external threats and be aligned with

existing legislation concerning cybersecurity.

Furthermore, it is essential to analyse whether and to what extent the cur-

rent legal framework on liability is still fitting in order to protect users in

the area of AI. The Danish Government supports that individuals having

suffered harm caused with the involvement of AI systems need to enjoy

the same level of protection as individuals having suffered harm caused by

other technologies, whilst technological innovation should be allowed to

continue to develop. In this respect, the Danish Government welcomes that

all options to ensure this objective should be carefully assessed, including

possible amendments to the Product Liability Directive and possible fur-

ther specific targeted rules in the specific area of AI.

Overall, it is important that the adjustments of existing legislation are

drafted as clear as possible

–

both regarding the scope, definitions and al-

location of responsibilities, but also that they take other Union or national

legislation into consideration. A common and clear European regulatory

framework will increase consumer protection and benefit European com-

panies.

An ecosystem of excellence

The Danish Government supports the Commission’s ambitious focus on

research concerning AI and recognizes the importance of international co-

operation among research centres with expertise in AI. For research initia-

tives to accelerate the innovation of AI, these must be coupled by adjusting

our framework conditions at the European level. Therefore, the Danish

Government further supports initiatives to make data, investment and test-

ing facilities available across the EU with the aim of making the technol-

ogy accessible for both the private as well as the public sector. In this re-

spect, effective coordination across fields, programmes, the Commission

and the Member States is essential in order to achieve the common objec-

tives in terms of AI.

Revising the Coordinated Plan

As AI is undergoing rapid development

–

both in terms of the technology

itself, but also due to initiatives set out in the Digital package as well as the

context with COVID-19 - the Danish Government looks positively at re-

visiting the Coordinated Plan with an aim to review the impact of its initi-

atives and based on such a review, adapt where necessary. The overall ob-

jective of the Coordinated Plan must still be to foster closer cooperation

and coordinating common priorities and initiatives within AI.

If Europe truly wants to be a front-runner in the global digital economy,

we need to be ambitious in our priorities as well as maximizing efforts

through closer cooperation which the Coordinated Plan caters for. At the

same time, the Coordinated Plan should utilize and build upon existing

well-functioning structures, institutions and clusters of expertise and ca-

pacities in order to improve the impact of the plan. This must be taken into

account when revisiting the plan.

With regards to the energy sector, the Danish Government strongly sup-

ports the focus on how digitalization, access to data and emerging technol-

ogies can support our efforts to meet the EU’s climate neutrality targets by

2050. In our view, the Coordinated Plan could to a greater extent become

an enabler of accelerating the twin transition towards a green and digital

economy

which must be the backbone of the EU’s recovery plan for a more

sustainable and resilient Europe. For instance, by establishing testing and

experimentation facilities that focus on developing and using solutions that

paves the way towards a carbon-neutral and circular economy.

Furthermore, the Danish Government suggests that a review of the Coor-

dinated Plan builds on the experiences obtained from the existing plan with

respect to the security implications related to AI. The review should also

include considerations on the security implications of AI in other critical

functions in society beyond transport, security and energy.

Focusing efforts on the research and innovation community

The proposed creation of reference testing centres for AI under the Digital

Europe programme should focus on providing a competitive advantage for

Europe by ensuring that the testing centres are relevant for a variety of

public and private sectors, applications and initiatives. It is important that

the selection of testing centres follows an open, transparent and inclusive

process and that the test centres commit to involve both SMEs as well as

start-ups in order to foster the entire ecosystem.

Testing centres should have a flexible set-up and not apply a one-size fits

all model, as AI can be used in many different settings. It is important that

the testing centres build upon and strengthen existing capabilities in Mem-

ber States and have a strong connection to existing AI ecosystems in order

to avoid duplication. Furthermore, the testing centres should be obliged to

develop and support responsible and trustworthy AI as well as having a

strong focus on promoting carbon-neutral and circular solutions across sec-

tors. We further believe that Denmark has several well-suited candidates

that could participate in the set-up of testing centres for the benefit of all

of Europe.

Furthermore, research and innovation actions under Horizon Europe will

be an important element in Europe’s approach to AI.

Denmark fully sup-

ports the proposed AI and Robotics intervention area of cluster 4 (Digital

and Industry) and notes that it is of particular importance to ensure effi-

cient, cross-cutting coordination with other clusters in Horizon Europe, as

well as with efforts from other relevant EU programmes.

Skills

The Danish Government recognizes that there is a great need for a larger

digitally qualified

workforce, while emphasizing Member States’ compe-

tences in this area. Fully harnessing the potential of AI requires investing

in people's digital competences and skills. The Danish Government recog-

nizes the importance of a continuous focus on the intake of students in

higher education programmes in information technology, including AI.

Here, the Digital Europe Programme will be an important element, as it

aims to develop world-leading masters programmes in AI. As the focus in

the Digital Europe Programme is based on the broad uptake of digital tech-

nologies in the economy, the activities should have a sufficiently broad

scope in order to meet the demand from businesses and public sectors.

Furthermore, there is a need for safe and secure use and handling of data

on all levels with a specific focus on cyber and information security skills

as part of the overall effort to empower individuals and investing in digital

skills for citizens and SMEs.

Focus on SMEs

It is important that the European Digital Innovation Hubs under the Digital

Europe Programme ensure broad uptake of digital technologies by busi-

nesses and especially SMEs. The Danish Government is currently taking

steps to integrate the Digital Innovation Hubs in the existing Danish struc-

tures for decentralized business promotion. Modalities and framework for

the hubs must be drafted in such a way that the hubs can contribute effec-

tively to accelerate the digital transformation and have a real added value

for European businesses and public authorities. Therefore, no hindrance

should be introduced for the utilization of existing structures which already

meet the needs of businesses and public authorities and which can carry

out the tasks.

As AI is a technology of horizontal nature with importance for a wide range

of sectors, the Danish Government supports the initiative that at least one

European Digital Innovation Hub under the Digital Europe Programme has

a degree of specialization in AI. Such a specialization requirement will still

cater for a flexible framework, which enables Member States to build upon

existing national structure and institutions, as the hubs should build on lo-

cal strengths available as well as the future needs of the local economy.

Partnership with the private sector

The Danish Government supports the Commission’s ambition to establish

a public-private partnership in AI, data and robotics and acknowledges the

importance of building strong ties between public and private institutions

in the research and development of AI solutions.

To this end and based on national experiences within data ethics and AI,

the Danish Government has established a national cluster organization for

digital technologies in accordance with the national smart specialization

strategy 2020-2023. The cluster organization aims at promoting innovation

in companies developing AI solutions by strengthening collaboration be-

tween the private and public sector.

Promoting the adoption of AI by the public sector

The Danish Government supports the proposal to initiate sector dialogues,

giving priority to healthcare, rural administrations and public service oper-

ators as set out in the white paper, which seems to resonate with the areas

incorporated in the existing Coordinated Plan, thereby strengthening exist-

ing efforts. However, we suggest that AI systems intended for green, cli-

mate-positive purposes are also added to the priority list.

It is important that a new action plan on AI aimed at the public sector al-

lows for both a top-down approach to identify critical functions in society

that must be given priority to ensure the strategic autonomy of EU member

states, as well as a bottom-up approach, for example an explorative ap-

proach that presents local authorities, such as municipalities and regions

with a high degree of self-determination to experiment with AI solutions

within the realm of the law.

Regarding the “Adopt AI programme”, Denmark proposes that the pro-

gramme not only focuses on the aspect of the public sector procuring AI

systems, but also includes the aspect of the public sector developing AI

systems.

A programme focusing on public development and testing of AI solutions

could support the “Adopt AI programme” by incentivising local authorities

to develop AI solutions for application in prioritised areas holding a poten-

tial to heighten the quality and capacity of the public sector through up-

scaling of the technology, but where experience is needed before a wider

uptake is possible. Such a programme could draw inspiration from the Dan-

ish “signature projects”, a joint public sector effort to test AI within the

prioritised areas of healthcare, public administration and employment.

Securing access to data and computing infrastructures

Ensuring access to computing infrastructure as well as high-quality data

across different sectors should be highly prioritized, as access to data is a

precondition for the development and wider uptake of AI. Therefore, it is

of utmost importance that the initiatives set out in the Commission’s Data

Strategy complement the initiatives set out in the white paper on AI. This

especially revolves around the development of data spaces, where stand-

ardisation is a key driver in establishing useful and well-functioning data

spaces,

cf. the Danish Government’s response to the Data Strategy.

International aspects

Alongside the EU’s priorities to develop and deploy trustworthy AI, the

Danish

Government supports the Commission’s continuous efforts to pro-

mote the ethical use of AI through international cooperation in order to

make this the international norm.