Algorithmic Impact Assessments under the GDPR: Producing Multi-layered Explanations
Year: 2019
Type: article
Abstract: Policymakers, scholars, and commentators are increasingly concerned with the risks of using algorithms for profiling and automated decision-making. This article addresses how a Data Protection Impact Assessment (DPIA), applied as an algorithmic impact assessment (AIA), links the two faces of the General Data Protection Regulation (GDPR) approach to algorithmic accountability: individual rights and systemic governance. We propose that AIAs simultaneously provide systemic governance of algorithmic decision-making and serve as an important ‘suitable safeguard’ (Article 22) of individual rights. As a nexus between the GDPR’s two approaches to algorithmic accountability, DPIAs have a heretofore unexplored link to individual transparency rights. Our examination of DPIAs suggests that the current focus on the right to explanation is far too narrow. We call, instead, for data controllers to consciously use the mandatory DPIA process to produce what we call ‘multi-layered explanations’ of algorithmic systems. This concept of multi-layered explanations not only more accurately describes what the GDPR is attempting to do, but also normatively fills potential gaps between the GDPR’s two approaches to algorithmic accountability. To date, the discussion of the GDPR’s regulation of algorithmic accountability has largely focused on whether there is an individual right to explanation of an algorithmic decision. Only more recently have legal scholars begun to focus on the GDPR’s systemic accountability tools.1 Impact assessments have received particular attention, on both sides of the Atlantic, as a tool for algorithmic accountability. This article aims to address how a Data Protection Impact Assessment (DPIA) (Article 35) links the GDPR’s two approaches to algorithmic accountability—individual rights and systemic governance—and potentially leads to more accountable and explainable algorithms. Examining the GDPR’s approach to impact assessments suggests that the scholarship has been getting explanations wrong. Algorithmic explanations should not be understood as static statements but as a circular and multi-layered process. The literature has largely to date focused on what information goes to whom and when; we argue that the impact assessment process plays a crucial role in connecting internal company heuristics and risk mitigation to outward-facing rights and in forming the substance of several different kinds of explanations. We begin by introducing the algorithmic accountability tools in the GDPR (section ‘Algorithmic accountability in the GDPR’). In section ‘Individual rights in the GDPR and the multi-layered explanation’, we explore the individual rights of data subjects as regards algorithmic decisions. In section ‘Collaborative governance in the GDPR’, we explain the GDPR’s collaborative governance of algorithms. In section ‘The DPIA as an algorithmic impact assessment’, we explain the requirements of the DPIA under the GDPR. In section ‘A model algorithmic impact assessment’, we discuss the broader literature on Impact Assessments and how our interpretation of the GDPR’s impact assessment requirement in fact leads to a better, more complex understanding of the GDPR’s explanations of algorithmic decision-making than a focus on Article 22 alone. We close by calling for what we call a multi-layered approach to explanations, stemming from the impact assessment process. As a methodological disclaimer: article a of to the and of the Article the Data Protection an of Data Protection We are that are not and be to be the only interpretation of the GDPR. the GDPR Article that the is to and in to the of the GDPR. the of the is not only for but also for the of Data Protection we and to be an of the GDPR’s to in scholarship as in between and To and as is to the role in the of the GDPR as a The GDPR has for algorithmic decision-making. the legal focused on whether the GDPR an individual right to an explanation of an individual algorithmic legal to focus on accountability in the of the in in by the Article the tools the of Data Protection and the requirement of DPIAs (Article 35) under As of has the GDPR a approach to algorithmic accountability. a of individual rights with a systemic governance by more the and the and two and individual right is also a to have and the in the algorithmic governance a data has a right to an individual algorithmic (Article to of automated decision-making (Article and to to information the (Article to of the GDPR in on data controllers using whether that decision-making a The GDPR an of systemic accountability the of (Article and DPIAs (Article from the of the in and in the on and on in and in by the Article the is also crucial to the the GDPR The GDPR largely in the of with the substance of data and in the of an approach in the legal literature as the use of This of has been to as to the and than by the the with both and with to with the substance and collaborative governance is not to the has an role to the GDPR governance to accountability on in the GDPR is not individual rights. is that process of with and from the from and from both and in the two of article more on both the individual rights and systemic governance of the GDPR’s approach to algorithmic accountability, to the role of the DPIA in the two The GDPR several important rights with to algorithmic decision-making. The GDPR both data rights as and the right to and rights to profiling as the right to also to algorithmic of the GDPR rights to algorithmic a right to be of a right of both and to information the a right to be of the of and of and a right not to be to automated (Article with and for the in is but are not a right to a to of and to (Article We not to the legal rights in but an be As discussion of individual rights has largely focused on whether be to As and is to that there is right to an explanation in the the GDPR’s to explain information the in in to and information in has been the of whether is a explanation to individual and what of explanation is The has focused on whether not Article 22 an right to explanation of an individual by an automated Our by of is that with be to in the that be to the decision-making process to be to rights under the the right to a of the Article of the GDPR have the Article 22 explanation in of the Article 22 right to largely the more complex approach to algorithmic transparency by the GDPR as a As we discuss in section a model our is that the GDPR’s transparency rights are as a the is understood as a of multi-layered have a right to both a but of the of an and more on individual We discuss of explanations in have been in the legal literature both in and in the the of to both rights and algorithmic from to to individual and call for a the that individual rights with systemic of for individual accountability in of and The individual rights the and of rights to and and to by profiling decision-making explanations also serve to the of a decision-making by and to and assessment as to whether are we and call a of algorithmic individual as we discuss also the between the GDPR’s two rights a crucial role in the GDPR’s systemic collaborative governance. the GDPR’s approach to algorithmic accountability has the potential to important in the literature the in of individual rights in algorithmic The of algorithmic governance in the GDPR is systemic governance This to address to and on and an is is largely collaborative a approach to We two of how in the GDPR. Article on is of in The GDPR’s not what using to individual rights (Article of but to both and to what are The a right to individual explanation more of be in several the of the Article but are not to algorithmic data and using and tools are systemic accountability and in a and the that what as be an between and and potentially to with of of (Article The GDPR to with both the substance and the of regulation in The GDPR’s approach to and in algorithmic decision-making is of collaborative governance in with on on the of in profiling and algorithmic decision-making. The GDPR not how to the that data for the and of that and use and to on the of data as on a the GDPR not what to the of what and with with what of algorithmic accountability—individual rights by but collaborative governance of DPIA plays a We that as applied to algorithmic the DPIA is understood as a nexus between the GDPR’s two approaches to algorithmic accountability. in to what is be and to the potential as in the GDPR. The on the GDPR to DPIAs for with Article a DPIA for and of to is on automated on are that produce legal the the As and have that a DPIA is not in a We that algorithmic impact assessments for under Article of the In we address the DPIA as an We what the of the DPIA is in the and what the in algorithmic governance both what the should be and to several in the current of The GDPR’s DPIA in the as an to be an for the using impact assessments as a tool to algorithmic accountability. In we discuss the GDPR’s requirements for a in the to requirements to algorithmic decision-making. The GDPR DPIAs only under The GDPR describes a DPIA as assessment of the impact of the on the of (Article the of the a of the the and the of the an assessment of the of in to the an assessment of the risks to individual rights and and the a company use to address risks and GDPR (Article The DPIA a company a a company a and propose data (Article the GDPR also the risk by a a company whether is with Impact Assessment (Article should also under and the DPIA The DPIA an more of that DPIAs should as a of be the and that should be a DPIA is a not a the DPIA This process and that information the risk The DPIA process by The GDPR also requirements for the from of the impact assessments in the literature and in section for DPIAs not a of that are to AIAs have an internal but in (Article The GDPR with a data has In of the GDPR with (Article This in for from than the The DPIA that for in the of by and to This with a of of as in the literature on impact assessments in section for The DPIA explain that not have an to In and have to As for of the but not from from and to data The GDPR not DPIAs to be by a Data Protection a risk assessment that in risk in the of by the to the a company with the (Article a company whether should be to as of the impact assessment process. In the from the impact assessment DPIAs are not to be to the As the a DPIA is not a legal requirement of the GDPR data controllers should of The that is a to of the are not the the DPIA of a of the As scholars have there are in of the assessment be by the of the data as in the of in and in The GDPR and the DPIA an but what a company in a DPIA in the of the impact assessments in the legal the GDPR not the DPIA both as This has to the GDPR’s DPIAs as with the and process As we discuss is not the role of the DPIA in the of the requirements for a we to the of the DPIA as an impact assessment’, both by our understanding of the GDPR’s approach to algorithmic and by the on far commentators have the GDPR on to the DPIA The DPIA as an plays a role in the GDPR’s approach to algorithmic decision-making. We that in the of the GDPR’s algorithmic governance the DPIA should be understood as a nexus between the two approaches to algorithmic decision-making. The DPIA links the GDPR’s individual rights to systemic governance of algorithms. the DPIA in both potential and leads to how the DPIA as an be and The DPIA is not a As a tool in the GDPR’s algorithmic governance has more potential than the The DPIA has two as a tool in the GDPR’s systemic governance and as an of the GDPR’s of individual rights. the DPIA in a between the two how is to as an to the of also leads to in the section a model the of algorithmic explanations and to be the GDPR. understood as of the GDPR’s collaborative governance of the DPIA be as a of the DPIA with and with with internal and under a of but has the DPIA as to both a decision-making and decision-making governance is concerned with and a The applied in the of algorithmic with risks of and and with with of This aims to heuristics by the GDPR’s the and the what a company in and algorithmic decision-making. The process of the from with an with a and both internal and and to internal company in a be and is not The DPIA also be understood in as a to as a and that be and by under the GDPR’s The DPIA has an unexplored role in the GDPR’s of individual rights. the DPIA serve as a of for the to algorithmic the individual and rights. data subjects have a right to the as as the and the of DPIA as of the and the of the (Article a as of the DPIA internal be to serve as the for in to to the in the of a DPIA an assessment of risks to the rights and of and have a right in the of to be of the and of decision-making as a company a should how the information in that process also the individual rights requirements under the GDPR. the DPIA as by the on to for individual rights as of the process. of DPIAs as to in individual the DPIA is an of to individual the on a of individual process. The GDPR using under the to on to to individual rights (Article Data controllers should use DPIAs to what to address risks The on that not the use of of systemic accountability, but also a of individual the the and of algorithmic a to a and a to of The on to individual rights in the GDPR’s Article 22 as a of risk the DPIA process. individual rights as of a for algorithmic decision-making that is not but that more a In we the GDPR on the DPIA as as a of to individual algorithmic process rights. individual rights as the on both provide a as to what be in a and with the process of a individual rights in the DPIA as a of company company decision-making heuristics to an assessment of individual process rights. simultaneously as a collaborative governance to in the in of individual process rights. the DPIA has a role in the of collaborative governance to individual rights the of systemic accountability as the DPIA only and not with In the of algorithmic and is more and be understood as an for algorithmic The goes as in the of algorithmic the use of and to in that in in data are and the risk of is that and that requirement the GDPR’s ‘suitable requirement from the of individual in the GDPR’s to a far broader of systemic accountability (Article The on of for algorithmic decision-making that use both internal and and This that in the of algorithmic a company the DPIA process and the risk mitigation the the fact that is not a requirement for DPIAs in the of are and as a of ‘suitable safeguard’ ‘suitable to individual rights in the of the in the on links individual rights with collaborative governance are with with to and to individual and is how to address is also as a crucial of individual rights in the in for to that are not to an to is that in the DPIA process the as with of with algorithmic and that and rights are by a also As the is the DPIA the two approaches to algorithmic governance in the GDPR. The of the DPIA is that not a for mandatory to the as in section of a model is understood by to be an of impact assessments as a both in the of to with and in the of the to in that company to the DPIA to both of are of a collaborative governance This be The GDPR a of in the of and in the of As the GDPR with with the substance of how individual rights be and how to address and algorithms. In the of how we be that of individual rights and collaborative governance is the is to use the GDPR’s have been in to the with also of the of and to the of impact the GDPR to in governance as not be for The DPIA also potentially to the and with using only and rights of the accountability for the GDPR’s collaborative governance. This is more link DPIA to what to the systemic of a DPIA to as the in a decision-making is that to be to provide the and company an individual has been the information has received a to a in the and the from the by This a more of getting the as and risks the DPIA process from individual rights. the GDPR’s DPIA to the DPIA as an we to a more We close article with a call for more on a model that serve as a for what we call multi-layered explanations of algorithmic decision-making. This to what and accountability be and and to how to and also a of how to link the the DPIA process to the individual under the GDPR. We are not the to focus on impact in We the to discuss AIAs not in but as a of the GDPR’s approach to algorithmic accountability. This the of the of impact assessments in from accountability an governance Our have for for AIAs in legal suggests that impact assessments serve a role in with accountability as of suggests that impact assessments a role both as a of and between the multi-layered individual explanations we are in the GDPR. We begin with an of the discussion that has recently AIAs have received a of on both sides of the as tools to address of algorithmic and in We discuss several important to the Impact Impact Assessments Impact Assessments Impact Assessments and Impact Assessments is important to from the of the impact assessment are are not by in legal In we discuss how different from the under have to more for The for impact assessment is the in the in by the impact requirement a to a the of the As a and whether the as a an Assessment a that and for whether to an impact the a of Impact goes on to an a the be and a on the impact of a be and The is to by the and and an is the of commentators have the as a model for impact assessments in for the as an to and a model for Impact to the is a model for data to in the of and the and call for Algorithmic Impact on the that an impact assessment is an regulation that to and with the describes the model not as an to but as a to that transparency a by information the of that in the to two scholars impact statements as both for the particular and for in a the model to far The process is static in only to the of a is than not from the process transparency and not for for impact assessments on as of in also to the for on the model of the The process by the is a and process on a by data and and and to as a model for to are assessments that legal to the of and with a of and of the the for the GDPR’s of the is the As in the the with in the as in the have been as are more the as a of the but goes on to a of important In several for the have as of the of under data a of of data to In to a from a a company to whether in with data This from the has and not serve as the for a the of an describes the assessment as on a than an in than in with to and of the of of the broader than legal not process and from and In and the concept of a Impact in than a but of a of in to the on and of a in to to the the that in DPIAs in the have to focus on data and data broader and legal impact to the We to for to on We both and in the We also a in literature that our on the GDPR to the use of an the with to to and use in be to using an to explore and by explain the the using the and the impact for the and a have to to be to the of be to a the have to mitigation in the To address the the of and and with not but of of a Impact in Algorithmic as a a than to algorithmic accountability and that should also potentially and the of on and whether the have an impact on a particular the as a by than by The a a that to on The of the call for a the to the use of an the is to the be mandatory than goes the the to an in a of an the be to a an the model and a an the model is as two an not a for a of is to to be by the both for and are and individual process for by the we to the of the GDPR. the of a and Impact Assessment in the between a and a the suggests that and in to the role of an impact assessment in the of data that the impact of on has is is and is circular in of a to by an of is but not is that is with the of the GDPR’, not the GDPR to a commentators have recently the DPIA and the role plays in the of algorithmic accountability more of the for impact assessments the of information to the This is both to how a is and to and for the use of a The of information to the be more in the of a an is not the a of with to the and information only to more call for and as a of the impact assessment use impact assessments to and a for a and Our has for for AIAs Our the GDPR’s of an suggests that the have largely several important AIAs are not understood as a In the of the are of a of Only the how impact assessments with tools in the the for and from In the of the both and and and to the role in algorithmic accountability, but not discuss to accountability tools in the Our suggests that impact assessments are tool in a and are not and are understood as with tools as individual rights. impact assessments serve as a between collaborative governance and individual The information a company the impact assessment process what to and to the The an impact assessment in serve not to and but also to a an This role is by the GDPR’s In the GDPR we that the GDPR’s of impact assessments is a of collaborative governance with the what as not how the DPIA to the broader of both collaborative governance tools and individual rights in the GDPR. as of a of there are unexplored between the GDPR’s DPIA and individual rights and is that of the GDPR’s individual rights and algorithmic decision-making are in an the GDPR’s of the has a for that a data should the risk of and The is the GDPR’s the not to but to both and the GDPR’s individual rights. there is a to company impact is a in by the links individual and systemic we the GDPR’s of the to be both the potential of and the between what we to as ‘multi-layered explanations’ in the GDPR. of the scholars, both and and the of and data The GDPR’s of individual rights by to the impact of in on and an of the an explanation for a and in have to that of to and by the the Impact Assessment with systemic approach to risk assessment and risk mitigation data controllers to how the not but We that systemic and explanations an and should be to and that a be that is under the GDPR. We close with a for a model under the GDPR. is in both the and to in we to that Our model is both by our understanding of the GDPR and by our of impact assessment we and as of an impact we our understanding of the GDPR to how the DPIA process and the GDPR’s of individual that we to as multi-layered explanations. our not only the discussion of impact but the individual algorithmic transparency rights in the GDPR. model process should the should the of as a of should and not but an is should to the of both of the and provide the for should not but in and to what we by model process should also the from algorithms as a in to algorithms as in the and use is a that of in the in the be for an is in The risks from the by and from the the and risks from how the using the are and not in use of model should be a process that but also assessment and for algorithms that and are in a model of the fact that is on a to and that a model be to not particular but that be to individual model an assessment of on a and and of to This not the of the of on in the of an in the GDPR’s current algorithmic accountability and As to different data controllers have different Article of the GDPR that the and risks of data the and to with the GDPR. algorithmic decision-making risks for data subjects should more As data controllers to a algorithmic accountability are in the GDPR both Article and the rights to to have a to a in the and to but are not and the on In the of more and the data should and a right to explanation of an individual decision. To the Impact Assessment process Data Protection be to and of the GDPR not in use the GDPR’s to particular and for This not to and the of the but to with algorithmic decision-making. use what have to more the of of algorithmic as in the on have in algorithmic decision-making on particular that is as with a impact assessment process. This approach of impact assessments to to internal company risk mitigation to the The more but from our is the DPIA process and should the substance of the GDPR’s individual algorithmic accountability rights. we a close the substance a company is to produce a on to the requirements of the GDPR. Article GDPR that a DPIA should a of the and the of the an assessment of the and of the in to the an assessment of the risks to the rights and of data subjects and the to address the and to the of data and to with Regulation the rights and of data subjects and This of a DPIA largely the GDPR’s individual algorithmic transparency rights. Article a DPIA to a of the be as the for the GDPR’s requirement that be of information the in algorithmic The on DPIAs the to the and of the of and and a of the and the on data of of a DPIA what is to information the we the GDPR’s transparency requirements for with and requirements for the are between DPIA and GDPR algorithmic accountability under the GDPR The on only more The on the requirements of a DPIA between DPIA and the on GDPR algorithmic accountability of the and the of the the by the The of data in the automated decision-making process is in the assessment of the and of the in to the the of data are is to the automated decision-making assessment of the risks to the rights and of data subjects to in kinds of legal under Article The to address the and to the of data and to with Regulation the rights and of data subjects and there are between information that a data is to produce a DPIA and what a data is to to as of the GDPR’s algorithmic transparency The data to the in a DPIA is to the algorithmic transparency to the of data in and how algorithmic profiling is The to the and of the in a DPIA is to the algorithmic transparency to explain the of data and the of the The to the data risks and on is to the transparency to explain the impact of the profiling use in the to of individual rights in the of is to the to and to address the risks in In in the of the of the DPIA process in the GDPR to to algorithmic transparency in the GDPR by Article the Our on the DPIA as systemic governance to
Source: International Data Privacy Law
DOI: 10.1093/idpl/ipaa020
Authors: Margot E Kaminski, Gianclaudio Malgieri
Language: English
Subfield: Law
Field: Social Sciences
Domain: Social Sciences
Keywords: algorithmic impact assessment, data protection impact assessment, General Data Protection Regulation, algorithmic accountability, right to explanation, automated decision-making, profiling, algorithmic transparency, collaborative governance, data controllers, explainable algorithms, individual rights, systemic governance, EU data protection law, AI regulation
Open Access status: hybrid
Estimated APC paid: $4,850
OpenAlex ID: W3126102167
