For EUTA members, artificial intelligence (AI)is critical to innovate and compete globally.Across the wide variety of B2B and B2Cservices we provide, we use AI to enhanceefficiencies and productivity while constantlyimproving the range of digital servicesaccessible to our customers. We stronglyencourage the development of a human-centric AI as a way to secure publicconfidence and public / private investments. We are keen to highlight the many day-to-day applications of AI used by EUTAmembers, and how these applications bringtangible benefits to EU businesses andconsumers alike. Simultaneously, and aswith any new transformative technology, it isuseful to identify and address anychallenges or risks that may arise, such aslegal uncertainty and ethical questions

Forewordaround AI. We aim to be a constructive voicein this regard, and act as a sounding boardwith both business acumen and hands-ontechnical AI expertise. With these objectives in mind, we urge the EUinstitutions to conduct a thoroughassessment of the existing legislation beforeintroducing any new proposals specificallytargeting AI-driven technologies andapplications. Any new rules should also seekto address loopholes in the existing dataprotection, data security andliability regime. With this in mind, we haveset out a series of high-level principles on AIto provide an insight into the key challengesand opportunities facing EU businesses withAI capabilities, and to also ensure thedevelopment of a fair and sound AIframework in Europe.

Gianpiero Lotito Magdelena PiechPresident of the EUTA Chair of the EUTA

Gianpiero Lotito Magdalena Piech

European Tech AllianceHigh-Level Principles on Artificial IntelligenceFebruary 2020

1

2

Understanding AI AI and machine learning are commonly usedby EUTA members to make small, everydaybusiness activities more efficient. First andforemost, AI enables us to make decisionsfaster while improving the customerexperience. Some key definitions are set outbelow: What is an algorithm? An algorithm is acalculus method (or a finite sequence ofwell-defined instructions) to resolve a classof problems. It is commonly used for onlinedata processing, computer operations andmathematical calculations. An algorithm canalso be used to leverage data in variousways, i.e. inserting a new data item in aspecific data collection structure, searchingfor a particular item or sorting a collectionof items. Not all cases of using algorithmsamount to use of AI What is AI? The definition of AI has changedover time. Today AI can be defined as theability of a machine, software or system toact intelligently, in the sense that it is notprogrammed to perform a single, repetitivemotion, but can learn and adapt itspredictions and behaviour depending on theenvironment and on the data it ingests. The“learning” element is what differentiates AIfrom algorithms, which are static and donot adapt to feedback. AI is therefore a

dynamic process that delivers increasinglyprecise results. At the same time, AI is onlydata-driven, and lacks contextualawareness, holistic understanding, andcannot be compared to a far more complexhuman intelligence. Today the generalconsensus is that we have achieved ‘narrowAI’ capabilities - where a technology canoutperform humans in a very narrowlydefined task - but we are still a long wayfrom ‘general AI’ capabilities at scale, wherea machine can apply knowledge and skills indifferent contexts, similar to humans. What is machine learning? Machine learningis a subcategory of AI, whereby a machinelearns without explicit programming, byidentifying patterns or trends in data inputsand adapting accordingly. Machine learningrelies on data to automatically re-programa system. What is deep learning? Deep learning is asubset of machine learning where artificialneural networks or computing systems aretrained to solve complex issues and learnfrom large, diverse, unstructured and inter-connected amounts of data. Deep learningcan process a wider range of data resourceswhile requiring less data preprocessing byhumans, and can often produce moreaccurate results than traditional machinelearning approaches.

& Key DefinitionsIntroduction to AI

3

for European Growth

Deliver personalised music recomm-endations based on users’ preferencesand listening habits which enable themto discover and enjoy more of thecontent they love; Provide customers with safe holidayaccommodation through fake propertydetection, thanks to data labelledimages and constant evaluation of themodel; Provide customers with the mostappropriate choice of clothing, shoes orbooks based on their personalpreferences (style, interest or price)amid millions of products; Highlight key information from customerreviews to ensure the information isuser-friendly and more easily-accessible; Combine Internet searches with marketstatistics to propose bespoke productrecommendations in a timely manner(for example, online adverts); Increase the customer serviceexperience, by predicting customers’intent & providing them with quick,easily understood advice throughchatbots;

Improve logistics by bundling e-commerce parcels together andcalculating the fastest way to the end-user; Prevent fraud, inappropriate content ordishonest behaviour on marketplaces ordating apps, e.g. detection of fakeprofiles, inappropriate images orcounterfeited goods; Boost operational efficiency by bulkinghuge amounts of data / automatedclassification of thousands of productsinto relevant categories on amarketplace; Personalise email communications (withproduct recommendations andvouchers) based on each client’sshopping habits; Offer instant translation of webpagesaround the world and around the clockto facilitate good customer experience. Automated playtesting using deeplearning and the cloud to getsignificantly faster feedback for leveldesigners allowing them to focus ontheir core creative roles and gettingplayers better content more quickly.

AI is already enabling breakthroughs in various sectors, from greater operational efficiency tomajor innovations. AI is employed for a wide variety of purposes ranging from customerservice improvement, internal processes automation, fraud detection, safety fault preventionor in radiotherapy, where deep learning is essential to optimise cancer treatment efficiency. For European tech companies, AI is increasingly vital to respond to customer demand forhigh-quality, personalised goods and services. Inter alia, AI allows EUTA members to performvarious tasks.

AI as an engine

Examples:

4

1.1.Promote European leadership andcompetitiveness by avoiding over-regulation of low-risk AI applications

EUTA High Level Principles onArtificial Intelligence

The aim of the EUTA’s high-level principles on AI is twofold: First, we wish toprovide insights on the key challenges and opportunities facing EUTA memberswith AI capabilities; and second, we would like to lay down principles that shouldguide the EU’s on-going assessment of a potential AI regulatory framework.

EUTA members strongly believe that their AI-driven applications and technologies arelow-risk, as these applications do not poseany risks to the physical well-being orfundamental rights of the end user. Similarly,most of our AI uses do not pose systemicthreats, for example, to public health andsafety, law or financial stability. We firmlybelieve risk is the right metric to define thescope of any new rules in relation to AI, andwe are supportive of the EU’s commitment toavoid over-regulation that would freezeinnovation and investment in AI.

At the same time, we acknowledge thatcertain risks do exist and a robust riskassessment framework is needed to helpmitigate these risks. In this context, it will beimportant to consider a variety of factors,including the rationale behind AI use, anypotential negative impact on specificindividuals, the category and quality of datainputs, existing technical and proceduralsafeguards to amend oversights or biases,model robustness and the degree of humanintervention. We stand ready to workalongside policymakers to help develop afuture-proof, risk-based framework that isflexible enough to reflect AI’s generalpurpose while addressing specific challenges.

5

3.3.Advocate for human-centric AIdevelopment In line with the European ParliamentResolution on AI from February 2019, EUTAmembers advocate for a human-centric AIdevelopment that would avoid possiblemisuses of AI technologies to the detrimentof fundamental rights. EUTA members arealso closely monitoring the work of the High-Level Expert Group on Artificial Intelligenceand release of Ethics Guidelines in 2019.Building on the Ethics Guidelines forTrustworthy Artificial Intelligence, EUTAmembers would support self-regulatoryprinciples for accountability and

transparency to help businesses treadmarket opportunities and the possibility forunfair bias and discrimination. The mostresource efficient approach in this regardcould be to focus on inputs, outputs andfeedback loops, while also considering thespecific characteristics of high-riskapplications of AI. As EUTA represents someof the largest and most innovativetechnology companies across Europe, webelieve that AI should be developed andused in line with ethical standards and havealready made significant investments in thisregard, including in bias-aware productdevelopment.

2.2.Ensure regulatory consistency on AI toavoid hampering R&D and innovation

While existing legal schemes and doctrinescan, in most cases, be readily applied to AIdevelopment and R&D, specific areas of AIdevelopment would benefit fromharmonisation through a single set ofEuropean rules to avoid a potentialpatchwork of national legislation. For EUTA members, regulatory consistencywith GDPR will be paramount to continueproviding best-in-class products andservices. Indeed, GDPR already placesrestrictions on automated decision-makingand provide individuals the rights to be

informed of the logic behind such decisionsand challenge the decisions, i.e. articles 14and following under GDPR. Any new EU rulesmust be carefully evaluated to avoidduplication or contradiction with the existingdata protection legal framework, but alsoall other adjacent frameworks such ascontractual law. EUTA also believes the e-Privacy Regulation as currently proposed willhamper Europe’s leadership in AI. We call onthe EU to make a fresh start with thisproposal, and withdraw the text until theGDPR’s implementation and effectivenesshave been thoroughly assessed.

6

4.4.All Internet users tend to be naturally biasedor express bias in their online activities. Datathat are fed into an algorithm canexacerbate an initial bias or personalpreference expressed by a specific user (e.g. ifa user only orders sports books from anonline marketplace, the algorithm will tendto exacerbate the initial bias expressed bythe user by suggesting books matching theuser’s interests). As a result, there is adifference between an unbalanced result anda biased result. An unbalanced result canseem biased, but is actually objectivebecause the data inputs are genuinelyrepresentative of the reality. A biased result,however, can be the result ofunrepresentative data inputs or even biasedsystem design (e.g. when a system isdesigned by white men only). Within this context, providing fulltransparency on algorithm bias or explainingautomated decisions in a meaningful way toall Internet users is complex, creatingunnecessary red tape and burdensomerequirements for European businesses relyingon low risk AI applications. The mostappropriate technical and organisationalmeasures for mitigating AI risks from theoutset will be situation-dependent. For low-risk applications, EU businesses should havethe flexibility to choose measures that willdeliver the best outcomes. EUTA membersalso believe that EU rules could helpencourage self-regulation while alsoproviding a useful transparency toolbox.

- Transparency could be fed into DataProtection Impact Assessments, a keycompliance tool for certain forms of AIprocessing under GDPR [2]. Controllers usingAI systems to process data or make decisionsabout individuals are required to take stepsto prevent discriminatory behaviours in suchsystems to avoid breaching existing EUlegislation; - On technical measures, if the bias ordiscrimination is caused by an unbalanceddata set, ‘re-balancing’ the data, withaddition or removal of new data, can betaken as a solution; - Other solutions, albeit even more complexthan ‘re-balancing’ data, can consist intraining the system differently, or modifyingthe system post-training to achievealgorithmic fairness; - Striving to have a diverse workforce canalso be a powerful tool to manage bias in AIsystems, as recently suggested by the UKICO[3].

Set EU rules to encourage self-regulation to mitigate negative bias

[2]: The GDPR Recital 71 states that data controllers should take measures to prevent ‘discriminatory effects on natural persons on the basis of racial or ethnicorigin, political opinion, religion or beliefs, trade union membership, genetic or health status or sexual orientation, or processing that results in measures havingsuch an effect’.[3]: Accessible online at: https://ai-auditingframework.blogspot.com/2019/06/human-bias-and-discrimination-in-ai.html

Some examples are:

7

5.5.Build up Pan-European Data Commonsto tackle AI research siloes andovercome restrictions on access topublic datasetsPublic data sources are often disparate andcomplex, leading to a general fragmentationin terms of access and quality. Siloedworking, inconsistent collection methods anda lack of rules around data formatinteroperability inhibit research andinnovation that could be derived fromdatasets. EUTA members are supportive ofbuilding up a voluntary AI platform servingas a central point to gather and provideaccess to all publicly funded AI relatedknowledge and algorithms. For publiclyfunded datasets to be accessed by a widercommunity of users, the French Etalab is agood initiative that could be replicatedacross EU Member States and potentiallymerged to optimise common resources.

6.6.Without contemplating any new specificprovision for AI in the near future, theEuropean Parliament initially recommendedto “regularly re-evaluate current legislationto ensure that it is fit for purpose withrespect to AI while also respecting EUfundamental values and to seek to amend orsubstitute new proposals where this is shownnot to be the case”, while also monitoringthe relevance and effectiveness ofintellectual property rules. In line with theEuropean Parliament, EUTA members wouldwelcome a wider use of ‘regulatorysandboxes’ which consist of offering to AIoperators an opportunity to test, under realconditions, the safety and efficacy of theirtechnologies by temporarily releasing themfrom regulatory constraints.

Ensure that AI regulations are future-proof with regulatory sandboxes toavoid preempting AI innovation

8

8.8.Attract the best talents in EuropeTo fuel AI research and development, EUTAmembers need to attract and retain the besttalents in Europe. Member States should beencouraged to provide special residencepermits for AI specialists coming from thirdcountries to work for European companies. Inthe same vein, promoting relevant andattractive university programmes, with aspecial emphasis on bringing girls or womento these fields of expertise, should also be apriority of European institutions andgovernments. Ensuring a regulatory framework which istechnology neutral and provides for theopportunity to develop the most advancedforms of AI is also key today to retainingtalents trained in Europe.

9

7.EUTA members are key contributors toprivate research in AI. EUTA members arecommitted to closing the skills gap in AI andare supportive of the DG CONNECT’sinitiatives in this regard. Yet despite massiveinvestments and the EU academic excellence,Europe is still lagging behind in the globalskills race. It is therefore urgent for the EU toboost the emergence of AI training modulesacross the EU, while also facilitatingresearcher mobility between the public andprivate sectors. Public research institutionsare the EU key asset in the field of AI andmore synergies between the public andprivate sectors should be urgently explored.

7.Create more synergies betweenpublic and private research in AI

10

9.9.Cybersecurity is a key component for AIdevelopment in Europe. In its Resolution onAI published in February 2019, the EuropeanParliament devoted an entire section tocybersecurity & called for preventingloopholes, cyberattacks and misuse of AI, byimplementing “product safety controls bymarket surveillance authorities andconsumer protection rules.” EUTA memberssupport the European Commission’s proposalto boost EU owned cloud capabilities,through an EU investment plan in new cloudcapabilities and new requirements for cloudservices marketed in the EU (aroundcybersecurity, data protection, energyefficiency & portability/ interoperability).

Promote EU capacity-building andindependence in the supply ofcomponents and computing systems,data infrastructure & web platforms

However, any initiative seeking to addressthe EU dependency on the foreign supply ofcomponents and computing systems, datainfrastructure and web platforms, mustconsider that EU-based companies arecurrently relying on components, systems &infrastructure coming from third countries –mostly China and the US.  As a result, it willtake years before any EU initiative can closethe EU IT capability gap, due to the legacy ofmassive investments already made in otherregions. Overall, EU companies should retaincommercial freedom to choose the bestpossible offer suited to their business,irrespective of the country of origin.

11

The European Tech Alliance (EUTA) brings together and give a voiceto the major European digital champions, scaleups and leadingstartups. We believe that Europe is good at tech and our sector isdriving jobs and growth across the continent. With an overarchinggoal of fostering innovation in Europe, EUTA members are keen toprovide expert insights to the EU institutions and promote the EUcompetitiveness in the global AI race. This paper has been developed at a preliminary stage in the policydiscussions in order to share our members’ expertise and inform thedebate. It is not directly attributable to any individual member andwe invite you to contact our members, should you like to betterunderstand their specific situation.

12