EU guidelines on ethics in artificial intelligence: Context and implementation

BRIEFING

EPRS | European Parliamentary Research Service

Author: Tambiama Madiega

Members' Research Service

PE 640.163 – September 2019

EU guidelines on ethics

in artificial intelligence:

Context and implementation

SUMMARY

The discussion around artificial intelligence (AI) technologies and their impact on society is

increasingly focused on the question of whether AI should be regulated. Following the call from the

European Parliament to update and complement the existing Union legal framework with guiding

ethical principles, the EU has carved out a 'human-centric' approach to AI that is respectful of

European values and principles. As part of this approach, the EU published its guidelines on ethics

in AI in April 2019, and European Commission President-elect, Ursula von der Leyen, has announced

that the Commission will soon put forward further legislative proposals for a coordinated European

approach to the human and ethical implications of AI.

Against this background, this paper aims to shed some light on the ethical rules that are now

recommended when designing, developing, deploying, implementing or using AI products and

services in the EU. Moreover, it identifies some implementation challenges and presents possible

further EU action ranging from soft law guidance to standardisation to legislation in the field of

ethics and AI. There are calls for clarifying the EU guidelines, fostering the adoption of ethical

standards and adopting legally binding instruments to, inter alia, set common rules on transparency

and common requirements for fundamental rights impact assessments, and to provide an adequate

legal framework for face recognition technology. Finally, the paper gives an overview of the main

ethical frameworks for AI under development outside the EU (e.g. in the United States and China).

In this Briefing

• EU human-

centric approach to artificial

intelligence

• Key ethical requirements

• Implementation challenges

• Possible further EU action

• International context

• Outlook

EPRS | European Parliamentary Research Service

EU human-centric approach to artificial intelligence

Background

Artificial intelligence (AI) commonly refers to a combination of: machine learning techniques used

for searching and analysing large volumes of data; robotics dealing with the conception, design,

manufacture and operation of programmable machines; and algorithms and automated decision-

making systems (ADMS) able to predict human and machine behaviour and to make autonomous

decisions.

AI technologies can be extremely beneficial from an economic and social point of

view and are already being used in areas such as healthcare (for instance, to find effective

treatments for cancer) and transport (for instance, to predict traffic conditions

and guide

autonomous vehicles), or to efficiently manage energy and water consumption. AI increasingly

affects our daily lives, and its potential range of application is so broad that it is sometimes referred

to as the fourth industrial revolution.

However, while most studies concur that AI brings many

benefits, they also highlight a number of ethical, legal and

economic concerns, relating primarily to the risks facing

human rights and fundamental freedoms

. For instance, AI poses

risks to the right to personal data protection and privacy, and

equally so a risk of discrimination when algorithms are used for

purposes such as to profile people or to resolve situations in

criminal justice.

There are also some concerns about the impact

of AI technologies and robotics on the labour market (e.g. jobs

being destroyed by automation). Furthermore, there are calls to

assess the impact of algorithms and automated decision-

making systems (ADMS) in the context of defective products

(

safety and liability), digital currency (blockchain),

disinformation-spreading (fake news) and the potential military

application of algorithms (autonomous weapons systems and

cybersecurity). Finally, the question of how to develop ethical

principles in algorithms and AI design has also been raised.

EU approach

Policy-makers across the world are looking at ways to tackle the risks associated with the

development of AI. That said, the EU can be considered a front-runner with regard to establishing a

framework on ethical rules for AI.

Leading the EU-level debate, the European Parliament called on the European Commission to

assess the impact of AI, and made wide-ranging recommendations on civil law rules on robotics

January 2017. The Parliament drew up a code of ethics for robotics engineers and asked the

Commission to consider the creation of a European agency for robotics and AI, tasked with

providing the technical, ethical and regulatory expertise needed in an AI-driven environment.

Against this background, in 2018 the Commission adopted a communication

to promote the

development of AI in Europe, and in 2019 it published a coordinated plan on AI – endorsed by the

Council of the European Union – to coordinate the EU Member States' national AI strategies.

Building on this groundwork, in April 2019 the Commission published a set of non-binding

Ethics

guidelines for trustworthy AI. Prepared by the Commission's High-Level Expert Group on AI,

composed of 52 independent experts, this document aims to offer guidance on how to foster and

secure the development of ethical AI systems in the EU.

A recent report by

Algorithmwatch – a not-for-profit

organisation promoting more

transp

arency in the use of

algorithms – lists examples of

ADMS already in use in the EU. AI

applications are wide-ranging.

For instance, the Slovenian

Ministry of F

inance uses a

machine-

learning system to

detect tax evasion and tax fraud.

In Belgium, the police are using a

predictive algorithm to predict

car robberies. In Poland, this

technology is used to profile

unemployed people and decide

upon the type of assistance

appropriate for them.

EU guidelines on ethics in artificial intelligence: Context and implementation

Notion of human-centric AI

The core principle of the EU guidelines is that the EU must develop a 'human-centric' approach

to AI that is respectful of European values and principles.

The human-centric approach to AI strives to ensure that human values are central to the way

in which AI systems are developed, deployed, used and monitored, by ensuring respect for

fundamental rights, including those set out in the Treaties of the European Union and Charter

of Fundamental Rights of the European Union, all of which are united by reference to a

common foundation rooted in respect for human dignity, in which the human being enjoys

a unique and inalienable moral status. This also entails consideration of the natural

environment and of other living beings that are part of the human ecosystem, as well as a

sustainable approach enabling the flourishing of future generations to come.

While this approach will unfold in the context of the global race on AI, EU policy-makers have

adopted a frame of analysis to differentiate the EU strategy on AI from the US strategy (developed

mostly through private-sector initiatives and self-regulation) and the Chinese strategy (essentially

government-led and characterised by strong coordination of private and public investment into AI

technologies).

In its approach, the EU seeks to remain faithful to its cultural preferences and its

higher standard of protection against the social risks posed by AI – in particular those affecting

privacy, data protection and discrimination rules – unlike other more lax jurisdictions.

To that end, the EU ethics guidelines promote a trustworthy AI system that is lawful (complying

with all applicable laws and regulations), ethical (ensuring adherence to ethical principles and

values) and robust (both from a technical and social perspective) in order to avoid causing

unintentional harm. Furthermore, the guidelines highlight that AI software and hardware systems

need to be human-centric, i.e. developed, deployed and used in adherence to the key ethical

requirements outlined below.

Key ethical requirements

The guidelines are addressed to all AI stakeholders designing, developing, deploying,

implementing, using or being affected by AI in the EU, including companies, researchers, public

services, government agencies, institutions, civil society organisations, individuals, workers and

consumers. Stakeholders can voluntarily opt to use these guidelines and follow the seven key

requirements (see box on the right) when they are developing, deploying or using AI systems in

the EU.

Human agency and oversight

Respect for human autonomy and fundamental rights is at the

heart of the seven EU ethical rules. The EU guidelines prescribe

three measures to ensure this requirement is reflected in practice:

• to make sure that an AI system does not hamper EU

fundamental rights, a fundamental rights impact

assessment should be undertaken prior to its

development. Mechanisms should be put in place

afterwards to allow for external feedback on any potential

infringement of fundamental rights;

• human agency should be ensured, i.e. users should be

able to understand and interact with AI systems to a

satisfactory degree. The right of end users not to be

subject to a decision based solely on automated

processing (when this produces a legal effect on users or

significantly affects them) should be enforced in the EU;

The key EU requirements for

achieving trustworthy AI

• human agency and

oversight

• robustness and safety

• privacy and data

governance

• transparency

• diversity, non-

discrimination and fairness

• societal and environmental

well-being

• accountability

EPRS | European Parliamentary Research Service

• a machine cannot be in full control. Therefore, there should always be human

oversight. Humans should always have the possibility ultimately to over-ride a

decision made by a system. When designing an AI product or service, AI developers

should consider the type of technical measures that should be implemented to

ensure human oversight. For instance, they should provide a stop button or a

procedure to abort an operation to ensure human control.

Different types of fundamental rights impact assessments are already being used in the EU. The European

Commission adopted a set of guidelines on fundamental rights in impact assessments and uses this

checklist to identify which fundamental rights could be affected by a proposal and to assess systematically

the impact of each envisaged policy option on these rights. The General Data Protection Regulation (GDPR)

provides for a regulatory framework that obliges data controllers to apply a Data Protection Impact

Assessment (DPIA). The Government of Canada has also developed an Algorithmic Impact Assessment that

assesses the potential impact of an algorithm on citizens, with a digital questionnaire evaluating the potential

risk of a public-facing automated decision system. This tool will be mandatory in Canada as of 2020.

Technical robustness and safety

Another essential requirement is to have secure and reliable systems and software. Trustworthy

AI requires algorithms to be secure, reliable and robust enough to deal with errors or inconsistencies

during all life-cycle phases of an AI system. This requirement is about ensuring cybersecurity. In

practice, all vulnerabilities should be taken into account when building algorithms. This requires

testing AI systems to understand and mitigate the risks of cyber-attacks and hacking. AI developers

should put in place processes capable of assessing the safety risks involved, in case someone uses

the AI system they are building for harmful purposes. For instance, if the system is compromised, it

should be possible for human control to take over and abort the system. To tackle this important

question, the EU applies a twofold approach: first, fostering cooperation between the AI

community and the security community, and second, reflecting

on how to modify the legal

framework governing liabilities in the EU, and to go from a human-conduct-based liability regime

to a more machine-based liability regime.

Privacy and data protection

In the EU, there is a lot of consideration for data protection and privacy, and all AI stakeholders must

comply with the General Data Protection Regulation (GDPR

) as a matter of principle. Furthermore,

the EU guidelines on AI advise the AI community to ensure privacy and personal data are protected,

both when building and when running an AI system. Citizens should have full control over their

own data, and their data should not be used to harm or discriminate against them. In practice, this

means that AI systems should be designed to guarantee privacy and data protection. To this end, AI

developers should apply design techniques such as data encryption and data anonymisation.

Moreover, they should ensure the quality of the data, i.e. avoid socially constructed biased,

inaccuracies, errors and mistakes. To that end, data collection should not be biased and AI

developers should put in place oversight mechanisms to control the quality of data sets.

Transparency

Transparency is paramount to ensuring that AI is not biased. The AI guidelines introduce a number

of measures to ensure transparency in the AI industry. For instance, the data sets and processes that

are used in building AI systems should be documented and traceable

. Also, AI systems should be

identifiable as such, and humans need to be aware that they are interacting with an AI system.

Furthermore, AI systems and related human decisions are subject to the principle of explainability,

according to which it should be possible for them to be understood and traced by humans.

EU guidelines on ethics in artificial intelligence: Context and implementation

Diversity, non-discrimination and fairness

The guidelines focus strongly on avoiding unfair bias

when AI products and services are designed. In

practice, AI developers should make sure that the

design of their algorithms is not biased (e.g. by the

use of an inadequate data set). Stakeholders that may

be directly or indirectly affected by AI systems should

be consulted and involved in their development and

implementation. AI systems should be conceived with

consideration for the whole range of human abilities,

skills and requirements, and ensure accessibility to

persons with disabilities.

Societal and environmental well-

being

AI systems should be used to enhance positive social

change and encourage sustainability and

environmental responsibility of AI systems. In other

words, measures securing the environmental

friendliness of AI systems should be encouraged (e.g.

opting for a less harmful energy consumption

method) and the social impacts of these systems (i.e.

on people's physical and mental wellbeing) must be

monitored and considered. Moreover, the effects of

AI systems on society and democracy (including

regarding the electoral context) should be assessed.

Accountability

Mechanisms should be put in place to ensure responsibility and accountability for AI systems and

their outcomes. Internal and external independent audits should be put in place, especially for AI

systems whose use affects fundamental rights. Reporting of the AI systems' negative impacts

should be available (including for whistle-blowers), and impact assessment tools should be used

to that end. In situations where the implementation of the key ethical requirements creates conflicts

between them, decisions on the trade-off (i.e. the decision to choose to fulfil one ethical

requirement over another) should be evaluated continuously. Accessible redress mechanisms

should be implemented.

Implementation challenges

While the implementing phase of the guidelines has started, academics and stakeholders have

warned about a number of implementation challenges.

Need for clarification

The lack of clarity in the wording of the guidelines has been criticised in many respects. Thomas

Metzinger, professor of theoretical philosophy at the University of Mainz and a member of the

Commission's expert group on AI, warns

that the guidelines are short-sighted, deliberately vague

and do not take long-term risks into consideration. Furthermore, he regrets that the 'red-lines'

(i.e. non-negotiable ethical principles) in the draft guidelines were deleted or watered down in the

final text. Two of these 'red lines' had been that AI should never be used to build autonomous lethal

weapons or social scoring systems. However, after protracted negotiations, the final version of the

Explainability – part I

The wide-ranging concept of explainability is

about

making explanations on an

algorithmic decision-making system

available. The requirement for

explainable AI

addresses the fact that complex machines and

algorithms often cannot provide insights into

their behaviour and processes. This sometimes

results in a black box effect, i.e. a situation

where AI systems are capable of producing

results, but the process by which the results are

produced and the reasons why the algorithm

makes specific decisions are not fully

understandable by humans.

Explainability is therefore particularly

important to ensure fairness in the use of

algorithms and to identify potential bias in the

training data. This far-reaching requirement

means that an explanation should be available

on how AI systems influence and shape the

decision-making process, on how they are

designed, and on what is the rationale for

deploying them. Explainability must address

both the technical processes of an AI system

and the related human decisions taken in

accordance with the EU guidelines.

EPRS | European Parliamentary Research Service

text instead referred to these issues as 'critical concerns' and did not include a clearly formulated

prohibition.

Another expert group member, Andrea Renda, together with the AI task force of the Centre for

European Policy Studies (CEPS), also published a report

highlighting some shortcomings of the draft

ethics guidelines. The report warns in particular about the lack of a hierarchy of principles that

would otherwise have allowed EU institutions to tailor their policy approach.

Lack of regulatory oversight

The EU ethics guidelines are non-binding. However, concerns have been raised regarding the lack

of regulatory oversight to support their implementation. Non-profit research and advocacy

organisation AlgorithmWatch stresses

that most of the recommendations and guidelines on AI

issued so far do not provide any oversight mechanisms to ensure and enforce compliance with

voluntary commitments. Without such mechanisms, however, there is little incentive to adhere to

these ethical principles. Others also

warn about the risk of the technology industry financing and

shaping the ethical debate about algorithms and automated decision-making systems. The lack of

regulatory oversight raises the issue of the empowerment of public bodies or authorities to monitor

the enforcement of the EU ethical guidelines.

The AI Now Institute, argues for expanding the powers of regulators to oversee, audit, and monitor AI

technologies by domain. The institute favours a sector-speciﬁc approach that focuses on the application of AI

technologies within individual domains (e.g. health, education, transport).

Need for coordination of actions at EU and national levels

Several EU Member States have started work on establishing their own national frameworks on

ethics and AI in parallel to the EU initiatives. Below is an outline of these moves, by country.

France

Dating from March 2018, the French AI strategy sets out as one of its core principles the requirement

that AI technologies must be explainable to be socially acceptable. To that end, the government is

required to: put in place several policies in order to develop algorithm transparency and audits;

include ethics in training for AI engineers and researchers; carry out a discrimination impact

assessment (to encourage AI designers to consider the social implications of the algorithms they

produce); and ensure that the principle of human responsibility is applied (e.g. by setting

boundaries for the use of predictive algorithms in the law enforcement context). Furthermore, it is

proposed that a consultative ethics committee for digital technologies and AI be set up for the

purpose of organising a public debate in this field.

Germany

Initially, the ethics-related debate was essentially driven by sector-specific industry interests, and

resulted in the adoption in June 2017 of a set of

ethical rules for automated and connected vehicular

traffic by the Transport Ministry's Ethics Commission. In November 2018, the national AI strategy

was launched, setting out a range of measures on ethics. For instance, the document advocates

using an 'ethics by, in and for design' approach for all development stages and uses of AI. It

pledges to promote research into novel ways for pseudonymising and anonymising data and for

differential privacy. Furthermore, the federal government will review whether the German AI-

related legal framework covers all aspects related to algorithm-based and AI-based decisions,

services and products and, if necessary, adapt it in order to make it possible to verify whether there

is any undue discrimination or bias. The legislation governing the use of personal and non-

personal data for AI-based applications will be reviewed, and the possibility to establish and/or

expand government agencies or private-sector auditing institutions to verify algorithmic decision-

making processes will be examined.

EU guidelines on ethics in artificial intelligence: Context and implementation

Finland

In August 2018, the Ministry of Economic Affairs issued a report recommending to set up a

parliamentary monitoring group to promote the ethical value base of AI more extensively in

society, and to monitor and evaluate pilots and technology developments associated with the

ethical aspects of artificial intelligence. This group would be tasked with the creation of rules and

the assessment of practices in the context of defining responsibilities in situations where a machine

is taking decisions autonomously.

United Kingdom

The UK Committee on Standards in Public Life announced in March 2019 it was launching an inquiry

into the use of AI in public services, with the aim of examining whether the rules were sufficient to

ensure that high standards of conduct are upheld as technologically assisted decision-making is

adopted more widely across the public sector. In the UK, there is particular focus on the risks of

biometrics in ongoing discussions. A 2019 report from the UK Biometrics and Forensics Ethics

Group outlines some of the ethical issues raised by the use of live (real-time)

face recognition

technology (FRT) based on machine-learning techniques and recommends the development of an

adequate legal framework. Against this background, the House of Commons Science and

Technology Committee has urged the UK government to issue a moratorium on the current use of

FRT and to prohibit further FRT trials until a proper legislative framework has been introduced, and

guidance on trial protocols and an oversight and evaluation system have been established.

Risk of fragmentation. EU Member States are likely to enact some diverging national ethical rules on AI that

could fragment the EU landscape in this domain. Such fragmentation may hamper the emergence of pan-

European Al services. Therefore, coordinated actions at EU and national levels will be key to ensuring coherent

harmonisation of the EU ethical guidelines and avoiding any discrepancies within the EU.

Possible further EU action

Following the publication of the EU guidelines on ethics in AI, the Commission launched a pilot

phase in June 2019 and invited all stakeholders to provide feedback on the practical

implementation of the key requirements by the end of 2019. To this end, companies participating

in the pilot will report on their experience in implementing the guidelines. Based on the feedback

received, the High-Level Expert Group on AI will propose a revised version of the compliance

assessment list to the Commission in early 2020.

Something crucial in this context is to reflect on the following question: to what extent will voluntary

ethical rules, driven by industry’s pace and strategies, be sufficient to address the ethical issues

raised by AI development. There are calls for stronger intervention on the part of public authorities

to influence the development and enforcement of these rules. The European Commission

President-elect, Ursula von der Leyen, has announced

that she will put forward legislative

proposals for a coordinated European approach on the human and ethical implications of AI within

her first 100 days in office. Policy-makers, academics and stakeholders have called for further action

to implement and complement the ethics guidelines, and equally to ensure a harmonised approach

and avoid fragmentation. Possible further action focusing on ethical issues ranges from soft law

guidance, hard law legislation and standardisation.

Clarification of the guidelines

One of the main recommendations of the CEPS task force report on AI is to adopt some guidance

allowing to identify which applications or business models are potentially problematic and which

should be prohibited because they are incompatible with EU core values and legislation. The report

further stresses that extensive explanations should be provided to establish effective fairness

standards and that it is necessary to focus more extensively on setting up appropriate redress

EPRS | European Parliamentary Research Service

mechanisms for individuals.

Another great challenge is to clarify how to implement the

requirement of explainability in a context where the complexity of AI algorithms can make it

difficult to provide a clear explanation and justification for a decision made by a machine (i.e. black

box effect). Ensuring a harmonised application of the guidelines throughout the EU would require

spelling out this concept in more detail.

Explainability – part II

While AI systems can be made explainable, this may result in a trade-off between cost and interpretability.

In order to apply the guidelines consistently and efficiently, stakeholders would need additional

recommendations on key questions such as: i) do they need to ensure explainability by design; ii) can they

differentiate the level of transparency required when they face cases where AI supports decision-making

by humans which may raise fewer explainability issues than fully automated decision-making systems;

and

iii) to what extent should intellectual property rights and trade secret protection be limited by the

implementation of the explainability requirement. In this regard, the AI Now Institute argues that AI

companies should waive trade secrecy and other legal claims that inhibit full auditing and understanding of

their software, because such trade secrecy contributes to the black box effect and makes it hard to assess bias,

contest decisions or remedy errors.

Standardisation

Standardisation is expected to play an essential role in driving AI market adoption. Standards can

influence the development and deployment of particular AI systems through product certification,

and serve to disseminate best practices in AI as is the case in cybersecurity or environmental

sustainability.

A number of standardisation organisations are working on AI technical standards;

in parallel, ethical AI standards are also being developed.

For instance, the joint technical

committee of the

International Organization for Standardization (ISO) and the

International Electrotechnical Commission (IEC) is working on

developing standards to ensure trustworthiness in AI

technology from the outset. Expert working groups are

considering how to technically achieve AI systems'

robustness, resiliency, reliability, accuracy, safety, security and

privacy. Another leading standardisation organisation, the

Institute of Electrical and Electronics Engineers (IEEE)

published in 2019 an ethical framework setting out more than

100 ethical issues and recommendations to serve as a

reference for policy-makers, engineers, developers and

companies deploying, selling and using AI systems. In

practice, the IEEE seeks to develop specific industry

standards and processes – related to transparency,

accountability, and algorithmic bias – for the certification of

AI systems.

Against this background, the Commission’s 2019 Rolling plan for ICT standardisation

identifies three

main actions in relation to standard-setting in AI, namely: i) fostering coordination of

standardisation efforts on AI in Europe; ii) ensuring coordination between standardisation efforts on

AI in Europe and other international standardisation efforts; and iii) integrating the outcomes of the

High-Level Expert Group on Artificial Intelligence within the standardisation roadmaps.

However, launching a standardisation process raises many questions. Similar to technical standards,

ethical standards are voluntary measures. Standards can be made mandatory or become a

condition for awarding procurement contracts

so as to ensure that industry players implement

them. Some researchers stress, however, that there are not sufficient grounds for the adoption of

public certification or mandatory standards on AI in Europe, as the self-certification framework is

AI ethical standards

IEEE standard establishes a

process model by which engineers

and technologists can address

ethical consideration throughout

the various stages of system

initiation, analysis and design of

new IT products and systems.

Other standards address the

manner in which personal privacy

terms are offered and how they can

be read and agreed to by machines,

or describe specific methodologies

(e.g. selection of data sets) to

address and eliminate bias when

algorithms are created.

EU guidelines on ethics in artificial intelligence: Context and implementation

evolving and it is too early to anticipate with enough certainty how the AI market will develop over

time.

The fact that standards are vague and certification enforcement and oversight are

unclear (i.e. who performs the ethical certifications?) has been criticised too. One concern is that the

standardisation and certification bodies are focused on enabling AI to become 'ethically' marketable

and that existing market logic will control AI development.

The risk of a race to the bottom in

regulatory oversight because AI development organisations may choose to locate in jurisdictions

that impose more lax rules for implementing ethical standards has also been pointed out.

EU regulatory framework on AI

A number of proposals on AI legislation have been discussed,

including several described below.

Legislation on transparency of decision-making systems

Transparency is paramount to ensuring that AI is not biased and AI systems are explainable. There

are calls to legislate and make the transparency requirement mandatory. For instance, the

Finnish national AI strategy

paper recommends assessing how ethical obligations could be imposed

on platforms as is done in the GDPR. The paper stresses in particular that certain parts of an

algorithm developed and used by the platforms could be prohibited if it distorts or restricts

competition without justification.

The EU could build on existing legislative initiatives and research

on transparency conducted in recent years. In July 2019, the EU adopted the new

Regulation (EU) 2019/1150

requiring providers of online intermediation services and online search

engines to implement a set of measures to ensure transparency and fairness in the contractual

relations they have with online businesses (e.g. online retailers, hotels and restaurants businesses,

app stores) that use such online platforms to sell and provide their services to customers in the EU.

The Commission is also carrying out an in-depth analysis on

algorithmic transparency.

Against this background, a 2019 Parliament study recommends the creation of a regulatory body

for algorithmic decision-making tasked with defining i) criteria that can be used to differentiate

acceptable algorithmic decision-making systems (that should be subject to an algorithmic

impact assessment) and systems that should be prohibited; and ii) the obligations falling on

algorithmic decision-making system providers (such as the obligation to make their systems

auditable). New EU legislation could also address the responsibility for informing the persons

affected by such systems, while also clarifying the explainability requirements and setting

specific liability and certifications regimes.

Sector-specific legislation in the health sector

It is arguably more important to ensure rigorous implementation of the ethical rules in specific

sectors, such as healthcare, where human control over algorithms and decision-making systems is

paramount. Against this background, the Finnish national AI strategy

proposes to formulate AI

ethics rules specific to the healthcare ecosystem. A 2018 study by the University of Oxford stresses

the need to analyse the implementation of the GDPR in the field of health research, and where

needed, amend laws or create more clarity through interpretation and guidance.

Legislation on face recognition technology

The use of face recognition technology (FRT) is becoming widespread across Europe and is giving

rise to growing concerns. FRT is considered as processing 'biometric data'

under the GDPR, and is in

principle subject to strict terms and conditions of use. However, technology experts disagree on

whether the GDPR framework is robust enough to address all issues created by the growing use of

AI-based FRT, or whether additional legislation will be necessary to ensure EU fundamental rights

are protected.

Already, the adoption of national FRT legislation is being discussed in some

Member States (see in particular the UK debate mentioned above).

EPRS | European Parliamentary Research Service

A number of legally binding instruments could be adopted to translate ethical rules into hard law and make

them mandatory for the most influential AI industry players in the EU.

International context

While the EU is clearly a front-runner in the debate on the ethical and social implications of AI, other

government entities in the world are also looking at these issues.

United States

In the US, while a range of industry players have already developed some codes of conduct on ethics

and AI, there are calls for more government-led regulation. Collaborative industry groups, such

as the Partnership on AI

(including Microsoft, Amazon, Facebook and Apple), have pledged to

develop and share best practices, including on ethics. The Association for Computing Machinery

(ACM) also published in 2018 a

Code of Ethics and Professional Conduct to guide the ethical conduct

of computing professionals. Furthermore, companies are developing their own ethical guidelines.

For instance, Microsoft has its own

AI advisory board and Google has disclosed its AI principles, an

ethics charter to guide the responsible development and use of AI in research and products.

However, there is growing concern

that self-regulation will not be enough to tackle the ethical

challenges posed by the development of AI. In 2018 the AI Now Institute issued a report concluding

that internal governance structures in most technology companies are failing to ensure

accountability for AI systems. It argues therefore that government agencies need greater power to

oversee, audit and monitor AI technologies, especially those involving face recognition.

China

In China, there is growing interest in setting up an ethical framework for the development of AI. In

2017, China released its Next Generation Artificial Intelligence Development Plan

setting out long-

term strategic goals for AI development in the country by 2030. One objective is to establish

regulatory and ethical frameworks to ensure the healthy development of AI in China. China would

promote self-discipline of the AI industry and enterprises, and increase punishments for data abuse,

violations of personal privacy and unethical activities in this regard. The Artificial Intelligence

Industry Alliance, which brings together Chinese tech firms and universities, released

draft

guidelines for self-regulation in the field of AI in May 2019.

These call for implementing principles

of 'human-oriented', 'secure/safe and controllable' and 'transparent and explainable' AI similar to

the ones enshrined in the EU AI ethical guidelines. Furthermore, the New Generation AI Governance

Expert Committee, established by the Ministry of Science and Technology, released in June 2019 a

document outlining eight non-binding principles to guide AI development in China. These

principles largely mirror the EU rules on AI.

For instance, AI development should conform to 'human

values, ethics, and morality'; 'should be based on the premise of safeguarding societal security and

respecting human rights'; should 'eliminate bias and discrimination in the process of data

acquisition, algorithm design, technology development, product R&D, and application'; and should

'respect and protect personal privacy'.

Other countries and organisations in the world

Canada has already adopted a number of guiding principles governing the use of AI in the

administration and public services. Public institutions are required to incorporate some ethical

principles (including privacy and transparency concerns) in their application of AI.

The 2018

Directive on Automated Decision-Making for Federal Institutions

outlines the responsibilities of

federal institutions and provides rules to help them assess and mitigate the risks associated with

deploying an automated decision system. Australia is also well advanced. The Office of the

Australian Information Commissioner published

a Guide to data analytics and the Australian privacy

principles in 2018 and is working on a national ethics framework to address standards and codes of

EU guidelines on ethics in artificial intelligence: Context and implementation

conduct in the field of AI. Preparatory work for establishing an AI ethical framework is also ongoing

in India, New Zealand, Singapore, South Korea and Japan.

Some international organisations

are engaging in setting international rules in the field of ethics and AI. In May 2019, the OECD and

associated nations adopted a non-binding list of guidelines

for the development and use of AI.

The AlgorithmWatch AI Ethics Guidelines Global Inventory lists all of the ethical frameworks and principles

being developed across the globe. Recent years have seen a flurry of initiatives from companies, governments,

NGOs and research bodies to propose ethical rules on AI. The ethical principles laid down in other jurisdictions

seem relatively similar to those of the EU (though less detailed) and are essentially of a self-regulatory nature,

even though there is growing demand for more government oversight.

Outlook

Policy-makers all over the globe are looking at how to tackle the risks associated with the

development of AI. In April 2019, the EU published its guidelines on ethics in AI, becoming a front-

runner in the setting up of a framework for AI. Ethical rules on AI, where such exist, are so far

essentially of a self-regulatory nature, and there is growing demand for more government

oversight. In the EU, there are strong calls for clarifying the EU guidelines, fostering the adoption of

ethical standards and adopting legally biding instruments in order to, inter alia, set common rules

on transparency, set common requirements for fundamental rights impact assessments and

provide an adequate legal framework for face recognition technology.

MAIN REFERENCES

Boucher P., How artificial intelligence works, briefing, EPRS, March 2019.

The Age of Artificial Intelligence, European Political Strategy Centre, European Commission, March 2018.

Ethics Guidelines for Trustworthy AI, Independent High-Level Expert Group on Artificial Intelligence,

European Commission, April 2019.

Artificial Intelligence: A European Perspective, JRC, European Commission, 2018.

Kritikos M., Artificial Intelligence ante portas: Legal & ethical reflections, briefing, EPRS, March 2019.

A governance framework for algorithmic accountability and transparency, study, Scientific Foresight Unit

(STOA), EPRS, April 2019.

Understanding algorithmic decision-making: Opportunities and challenges, study, STOA, EPRS,

March 2019.

ENDNOTES

See definition provided by the High-Level Expert Group on Artificial Intelligence (glossary section of the Ethics

Guidelines for Trustworthy AI). There is no commonly agreed definition for AI. For an overview of the notion of AI and

the difficulty in defining it, see Philip Boucher's 2019 EPRS briefings on How artificial intelligence works and on Why

artificial intelligence matters.

See EPRS briefing on Economic impacts of artificial intelligence by Marcin Szczepański, July 2019.

See Data quality and artificial intelligence – mitigating bias and error to protect fundamental rights, FRA Focus, EU

Agency for Fundamental Rights, June 2019.

See two EPRS publications by Mihalis Kritikos: What if algorithms could abide by ethical principles?, ‘at a glance’ note,

November 2018; and Artificial Intelligence ante portas: Legal & ethical reflections, briefing, March 2019.

Furthermore, the Parliament adopted an own-initiative-resolution on a Comprehensive European industrial policy on

artificial intelligence and robotics in February 2019. The European Economic and Social Committee issued an own-

initiative opinion on AI in May 2017 also calling for a code of ethics for the development, application and use of AI.

Furthermore, on 10 April 2018, 25 European countries signed a Declaration of cooperation on Artificial Intelligence.

See definition provided by the High-Level Expert Group on Artificial Intelligence (glossary section of the Ethics

Guidelines for Trustworthy AI).

See The Age of Artificial Intelligence, European Political Strategy Centre, European Commission, March 2018.

Ibid. See as well Artificial Intelligence: A European Perspective, JRC, European Commission, 2018.

EU action could also be taken to clarify the issue of liability of damages and measures to encourage data-sharing.

EPRS | European Parliamentary Research Service

See Andrea Renda, Artificial Intelligence: ethics, governance and policy challenges, report of a CEPS Task Force, 2019,

p. 117.

See PricewaterhouseCoopers, 2018 AI predictions: 8 insights to shape business strategy, 2018. The report argues that

most AI can be made explainable but at a cost, and explains that, if every step must be documented and explained,

the process becomes slower and may be more expensive.

See OECD, above at p. 93.

See Peter Cihon, Standards for AI Governance: International Standards to Enable Global Coordination in AI Research &

Development, University of Oxford, April 2019.

See Jens Popper et al., Artificial intelligence across industries, International Electrotechnical Commission (IEC)

whitepaper, October 2018, pp. 71-78.

See Marc Böhlen cited in Industry standards won't give artificial intelligence a conscience, Matthew Linares,

openDemocracy media platform, February 2019.

See Alan F. T. Winfield, Ethical standards in robotics and AI, University of the West of England, Bristol, February 2019.

See A. Renda, op.cit., p. 118.

See Marc Böhlen cited in M. Linares, op.cit.

See P. Cihon, op.cit., p. 17.

See a draft European Commission priorities working document published by Politico in August 2019.

See Work in the age of artificial intelligence – Four perspectives on the economy, employment, skills and ethics,

Ministry of Economic Affairs and Employment of Finland, 2018 p. 54.

See the following EPRS studies: A governance framework for algorithmic accountability and transparency, April 2019;

Understanding algorithmic decision-making: Opportunities and challenges, March 2019; Cost of non-Europe in

robotics and artificial intelligence - Liability, insurance and risk management, June 2019.

On 21 August 2019, the Swedish Data Protection Authority (DPA) imposed its first fine since the EU GDPR came into

effect in May 2018. The fine was imposed on a school for creating a facial recognition program in violation of the GDPR.

More generally, how the GDPR affects machine learning techniques remains unclear in many respects. See also Mapping

regulatory proposals for artificial Intelligence in Europe, Access Now non-profit advocacy group, November 2018.

Work has started being undertaken on the application of FRT in Schengen border control. See Biometrics and the

Schengen Information System – Fostering identification capabilities, JRC blog, European Commission, July 2019.

Google also established an Advanced Technology External Advisory Council (ATEAC) featuring prominent academics

and designed to monitor its use of artificial intelligence, but decided to dissolved

it in April 2019 amid a controversy

about the appointment and independence of some of the Council members.

See Joint Pledge on Artificial Intelligence Industry Self-Discipline (Draft for Comment), June 17, 2019.

See Chinese Government Governance Principles for a New Generation of Artificial Intelligence: Develop Responsible

Artificial Intelligence, 17 June 2019. The eight principles with regard to AI are: harmony and friendliness; fairness and

justice; inclusivity and sharing; respect for privacy; secure/safe and controllable; shared responsibility; open

collaboration; and agile governance.

See Responsible Artificial Intelligence in the Government of Canada (Digital Disruption White Paper Series), Treasury

Board of Canada Secretariat, 10 April 2019.

For an overview, see Regulation of Artificial Intelligence in Selected Jurisdictions, the Law Library of Congress, United

States, January 2019.

DISCLAIMER AND COPYR

IGHT

This document is prepared for, and addressed to, the Members and staff of the European Parliament as

background material to assist them in their parliamentary work. The content of the document is the sole

responsibility of its author(s) and any opinions expressed herein should not be taken to represent an official

position of the Parliament.

Reproduction

and translation for non-

commercial purposes are authorised, provided the source is

acknowledged and the European Parliament is given prior notice and sent a copy.

Mopic / Fotolia.

[email protected]ropa.eu (contact)

www.eprs.ep.parl.union.eu

(intranet)

www.europarl.europa.eu/thinktank

(internet)

http://epthinktank.eu

(blog)

EU guidelines on ethics in artificial intelligence: Context and implementation