Beyond explainability: justifiability and contestability of algorithmic decision systems

Probably not a substantial claim: may not pass S2 (checkable). May combine two claims.

Explanations and justifications have different goals and should not be conflated.

“Even if they often support each other, explanations and justifications have different goals and should not be conflated: a user can understand the logic leading to a particular outcome without agreeing on the fact that this outcome is good; vice versa, he/she may want to contest an outcome (being convinced that it is bad) without knowing or understanding the logic behind the algorithm.”

Classified as probably not a substantial claim.

Accountability is a condition for input legitimacy.

“Kees Van Kersbergen and Frans Van Waarden refer to accountability as a condition for input legitimacy 9 [45] and Steven Bernstein highlights the importance of justifications and contestations to ensure legitimacy [9].”

• cites [45] HLEG-AI (2019) Ethics guidelines for trustworthy AI. European Commission High-Level Expert Group on Artificial Intelligence. https://digital-strategy.ec.europa.eu/en/library/ethics-guidelines-trustworthy-aiNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

The first norm in Fig. 1 is an example of rule while the two others are examples of objectives.

“The first norm in Fig. 1 is an example of rule while the two others are examples of objectives.”

Probably not a substantial claim: may not pass S2 (checkable).

Justifications are complete only if they establish a continuous link between the high-level objectives of the ADS (the applicable norms) and its implementation, with arguments about the fact that the implementation complies with the objectives.

“Justifications are complete only if they establish a continuous link between the high-level objectives of the ADS (the applicable norms, for example nondiscrimination, reduction of recidivism rate, or compliance with a given legal requirement) and its implementation, with arguments about the fact that the implementation complies with the objectives.”

May combine two claims.

Broad requirements such as transparency, fairness and accountability are generally ill defined, seldom required by law and difficult to implement.

“Broad requirements such as transparency, fairness and accountability are often presented as ways to limit these risks but they are generally ill defined, seldom required by law and difficult to implement[19, 42].”

• cites [19] Bovens M (2006) Analysing and assessing public accountability. a conceptual framework. CONNEX and EUROGOV networks. https://www.ihs.ac.at/publications/lib/ep7.pdfNot checked yet

  Not checked in the free run.

• cites [42] Henin C, Le Métayer D (2021b) A framework to contest and justify algorithmic decisions. Springer AI Ethics. https://doi.org/10.1007/s43681-021-00054-3 sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S2 (checkable), S4 (based). May combine two claims.

Justifications alone would lead to a unilateral process whereas contestations introduce a dialectic, conversational process.

“Justifications alone would lead to a unilateral process whereas contestations introduce a dialectic, conversational process.”

Probably not a substantial claim: may not pass S2 (checkable).

Justifications and contestations have the potential to bring a radical change to the integration of ADS within human environments.

“In addition to being essential components of accountability and legitimacy, justifications and contestations have the potential to bring a radical change to the integration of ADS within human environments.”

Probably not a substantial claim: may not pass S3 (load-bearing).

For ADS embedded in critical systems that make decisions automatically without interaction with any individual, the beneficiary of justifications and potential issuer of contestations could be an auditor or an expert working for a certification body.

“For this type of ADS, the beneficiary of the justifications and the potential issuer of contestations could be an auditor or an expert working for a certification body.”

• AssumptionRests on an assumption: Justifiability and contestability cannot be imposed on any ADS in any context.

Classified as probably not a substantial claim. May combine two claims.

Professionals and affected people are involved in the operational phase of the ADS and have to cope with particular situations (their own situation for affected people; the situation of the persons they have to serve for professionals).

“In contrast to experts, professionals and affected people are involved in the operational phase of the ADS and they have to cope with particular situations (their own situation for affected people; the situation of the persons they have to serve -patients, litigants, or customers - for professionals).”

May combine two claims.

The interactions presented in Fig. 2 and Fig. 3 show that Algocate is useful both for users contesting the decisions of the ADS and for users who want to justify them.

“The interactions presented in Fig. 2 and Fig. 3 show that Algocate is useful both for users contesting the decisions of the ADS (e.g. individuals affected by the decisions or professionals who are not sure about the suggestion of the ADS) and for users who want to justify them (usually the operator of the ADS).”

May combine two claims.

The need to provide ways to challenge an ADS is recognized by law but legal requirements are generally expressed as ill-defined rights to obtain explanations or human intervention.

“The need to provide ways to challenge an ADS is recognized by law but legal requirements are generally expressed as ill-defined rights to obtain explanations or human intervention.”

• cites Recital 71 of the GDPRNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

Different types of justifications and different modes of contestation are appropriate depending on the level of expertise of the stakeholders and the nature of their involvement in the ADS.

“Different types of justifications and different modes of contestation are appropriate depending on the level of expertise of the stakeholders and the nature of their involvement in the ADS.”

Probably not a substantial claim: may not pass S2 (checkable).

The shift from facts to values is precisely the shift from explanations, which reflect the logic of the ADS, to justifications, which refer to external, debatable norms.

“In a sense, this shift from facts to values is precisely the shift from explanations, which reflect the logic of the ADS, to justifications, which refer to external, debatable norms.”

May combine two claims.

Explainability is useful but not sufficient to ensure the legitimacy of algorithmic decision systems.

“In this paper, we point out that explainability is useful but not sufficient to ensure the legitimacy of algorithmic decision systems.”

Classified as probably not a substantial claim.

The benefit of the protocol in such cases is that a party relying on a norm is compelled to elicit it and to submit it to the approval of the other party (or the human decision maker as a last resort).

“The benefit of the protocol in such cases is that a party relying on a norm is compelled to elicit it and to submit it to the approval of the other party (or the human decision maker as a last resort).”

Probably not a substantial claim: may not pass S2 (checkable). May combine two claims.

Justifications and contestations are essential parts of accountability.

“3.4 Justifications and contestations are essential parts of accountability”

• cites [12] Binns R (2018) Algorithmic accountability and public reason. Philos Technol 31:543–556 sourceNot checked yet

  Not checked in the free run.

Classified as probably not a substantial claim. May combine two claims.

A user can understand the logic leading to a particular outcome without agreeing that the outcome is good; conversely, a user may want to contest an outcome without knowing or understanding the logic behind the algorithm.

“a user can understand the logic leading to a particular outcome without agreeing on the fact that this outcome is good; vice versa, he/she may want to contest an outcome (being convinced that it is bad) without knowing or understanding the logic behind the algorithm.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

The death limitation norm prioritizes people with high death risk, while the spread limitation norm prioritizes people with a high potential of spreading the virus.

“The first one prioritizes people with high death risk (for instance the elderly and people with co-morbidities) while the second one prioritizes people with a high potential of spreading the virus (e.g. medical staff, teachers and professionals with many contacts).”

May combine two claims.

Professionals and affected people have a different focus (more local than global) and require different modalities for justifications and contestations (more interactive and less demanding in terms of expertise).

“Therefore, they have a different focus (more 'local' than 'global') and they require different modalities for justifications and contestations (more interactive and less demanding in terms of expertise).”

May combine two claims.

These conditions are not met for most ADS relying on machine learning since their specifications can generally not be characterized by logical formulas and their performances are better assessed through statistical means.

“These conditions are not met for most ADS relying on machine learning since their specifications can generally not be characterized by logical formulas 22 and their performances are better assessed through statistical means.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

The three norms in the vaccine allocation example have different sources of legitimacy (based on fairness or utilitarian considerations) and can be in tension.

“These norms have different sources of legitimacy (based on fairness or utilitarian considerations) and can be in tension.”

Classified as probably not a substantial claim. May combine two claims.

Fairness can be defined in many ways and some of these definitions are incompatible with one another and may be in tension with the accuracy objective.

“fairness can be defined in many ways and some of these definitions are incompatible with one another and may be in tension with the accuracy objective.”

• cites [8] Berk R, Heidari H, Jabbari S, et al (2017) Fairness in criminal justice risk assessments: the state of the art. Sociol Methods Res. https://doi.org/10.1177/0049124118782533 sourceNot checked yet

  Not checked in the free run.

May combine two claims.

Explanations are not sufficient to justify or to contest decisions.

“However, as argued in the previous section, explanations are not sufficient to justify or to contest decisions.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Algocate includes three types of norms (rule, objective and reference).

“Technically speaking, Algocate includes three types of norms (rule, objective and reference) [38].”

• cites [38] Guidotti R, Monreale A, Ruggieri S et al (2018) A survey of methods for explaining black box models. ACM Comput Surv (CSUR) 51:93Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S2 (checkable).

The elicitation of norms justifying the outcomes of an ADS and their use in a particular context is a sound discipline to increase the chances that the ADS has a legitimate objective and is used in an appropriate way.

“The elicitation of norms justifying the outcomes of an ADS and their use in a particular context is not only useful to enable contestability: it is also a sound discipline to increase the chances that the ADS has a legitimate objective and is used in an appropriate way.”

• AssumptionRests on an assumption: Critical assumptions about the ultimate goal of an ADS and the reasons to believe that it is well-suited to address them are often left implicit or unclear.

Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

Contestations and justifications are dual.

“As discussed in the previous section, contestations and justifications are dual.”

Classified as probably not a substantial claim.

Algocate does not take for granted that the legitimacy of a norm is accepted by all parties.

“Algocate does not take for granted that the legitimacy of a norm is accepted by all parties.”

Classified as probably not a substantial claim. May combine two claims.

Enumerating all possible outputs given all possible inputs may be computationally or logistically infeasible, and we may be unable to flag all undesirable outputs.

“Enumerating all possible outputs given all possible inputs be computationally or logistically infeasible, and we may be unable to flag all undesirable outputs.'”

• cites Doshi-Velez and Kim [28]Not checked yet

  Not checked in the free run.

May combine two claims.

The GDPR conditions are too restrictive because they seem to exclude decisions taken on the basis of an ADS but not based only on it, and decisions that would not have a significant impact on individuals.

“They are too restrictive for at least two reasons: they seem to exclude decisions taken on the basis of an ADS (but not 'based only' on it) and decisions that would not have a significant impact on individuals.”

• AssumptionRests on an assumption: The first criterion to take into account for requiring justifiability and contestability is the potential impacts of the decisions on individuals and on society in general.

Probably not a substantial claim: may not pass S2 (checkable). May combine two claims.

For ADS used by professionals, it is critical that the professional gets appropriate justifications and has means to contest the outcomes of the ADS.

“In this context, it is critical that the professional gets appropriate justifications and has means to contest the outcomes of the ADS.”

• AssumptionRests on an assumption: Justifiability and contestability cannot be imposed on any ADS in any context.

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing). May combine two claims.

The human may want to guard against certain kinds of discrimination, and their notion of fairness may be too abstract to be completely encoded into the system.

“For example, as mentioned in [28], 'the human may want to guard against certain kinds of discrimination, and their notion of fairness may be too abstract to be completely encoded into the system'”

• cites [28] Danaher J (2016) The threat of algocracy: reality, resistance and accommodation. Philos Technol 29:245–268 sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

Contestations can take different forms depending on the status of the disputed norm.

“Contestations can take different forms depending on the status of the disputed norm (e.g. legal proceedings for a legal norm).”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Even if death and spread limitations seem to support each other in the long term, they result in different vaccination strategies.

“For example, even if death and spread limitations seem to support each other in the long term, they result in different vaccination strategies.”

Probably not a substantial claim: may not pass S4 (based).

Little work has been done so far to facilitate justifications and contestations for non experts.

“However, little work has been done so far to facilitate justifications and contestations for non experts.”

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing).

Such allocation systems give rise to various ethical considerations.

“Such allocation systems give rise to various ethical considerations [64].”

• cites [64] Monahan J, Skeem J (2016) Risk assessment in criminal sentencing. Annu Rev Clin Psychol 12:489–513 sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

Risk-assessment tools emphasize one major justification at the detriment of the others: incapacitation.

“Risk-assessment tools emphasize one major justification at the detriment of the others: incapacitation.”

• cites [23] Chetali B, Nguyen Q-H (2008) Industrial use of formal methods for a high-level security evaluation. In: Cuellar J, Maibaum T, Sere K (eds) FM 2008: formal methods. Springer, Berlin, pp 198–213 sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S2 (checkable). May combine two claims.

The means presented in Section 4 and the mere elicitation of applicable norms can provide useful inputs to the deliberation about the legitimacy of the ADS but they should not overshadow the debate, which must involve subjective (political, philosophical, ethical) considerations.

“The means presented in Section 4 (in particular, in Section 4.1) and the mere elicitation of applicable norms (as argued in Section 3.3) can provide useful inputs to the deliberation about the legitimacy of the ADS but they should obviously not overshadow the debate, which must involve subjective (political, philosophical, ethical) considerations.”

Probably not a substantial claim: may not pass S3 (load-bearing).

The need to conceive explanations as an interactive process has been argued by several authors [56, 58].

“The need to conceive explanations as an interactive process has been argued by several authors [56, 58].”

• cites [56] Loi M, Ferrario A, Vigano E (2020) Transparency as design publicity: explaining and justifying inscrutable algorithms. Ethics Inf Technol. https://doi.org/10.1007/s10676-020-09564-w sourceNot checked yet

  Not checked in the free run.

• cites [58] Malgieri G, Comandé G (2017) Why a right to legibility of automated decision-making exists in the general data protection regulation. Int Data Priv Law. https://doi.org/10.1093/idpl/ipx019 sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing).

The interest for more interactions with humans in the design and exploitation phases of machine learning systems takes different forms.

“The interest for more interactions with humans in the design and exploitation phases of machine learning systems takes different forms [1].”

• cites [1] Abdul A, Vermeulen J, Wang D et al (2018) Trends and trajectories for explainable, accountable and intelligible systems: an HCI research agenda. Proceedings of the 2018 CHI conference on human factors in computing systems-CHI ’18. ACM Press, London, pp 1–18Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

A series of works [13, 15, 16, 55, 61, 69] refer to justifications as ways of ensuring that a decision is good (in contrast to understanding a decision), which is in line with the definition proposed in this paper.

“However, a series of works [13, 15, 16, 55, 61, 69], refer to justifications as ways of ensuring that a decision is good (in contrast to understanding a decision), which is in line with the definition proposed in this paper.”

• cites [13] Biran O, Cotton C (2017) Explanation and justification in machine learning: A survey. In: IJCAI-17 Workshop on Explainable AI (XAI), p 8Not checked yet

  Not checked in the free run.

• cites [15] Biran O, McKeown K (2017a) Human-centric justification of machine learning predictions. In: Proceedings of the Twenty-Sixth International Joint Conference on Artificial Intelligence, IJCAI-17, pp 1461–1467 sourceNot checked yet

  Not checked in the free run.

• cites [16] Biran O, McKeown KR (2017b) Human-centric justification of machine learning predictions. In: IJCAI, pp 1461–1467 sourceNot checked yet

  Not checked in the free run.

• cites [55] Liao B, Anderson M, Anderson SL (2020) Representation, justification, and explanation in a value-driven agent: an argumentation-based approach. AI Ethics. https://doi.org/10.1007/s43681-020-00001-8 sourceNot checked yet

  Not checked in the free run.

• cites [61] Mitchell M, Wu S, Zaldivar A et al (2019) Model cards for model reporting. Proc Conf Fairness Account Transp. https://doi.org/10.1145/3287560.3287596 sourceNot checked yet

  Not checked in the free run.

• cites [69] Persad G, Wertheimer A, Emanuel EJ (2009) Principles for allocation of scarce medical interventions. Lancet 373:423–431. https://doi.org/10.1016/S0140-6736(09)60137-9 sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Another project [41] acknowledges the need to provide mechanisms for users to challenge model predictions and states that doing so may require users to marshal evidence and create counter narrative that argue precisely why they disagree with a conclusion drawn by an AI system.

“Another project going in the same direction is described in [41]: the paper acknowledges the need 'to provide mechanisms for users to challenge model predictions' and states that 'doing so may require users to marshal evidence and create counter narrative that argue precisely why they disagree with a conclusion drawn by an AI system.'”

• cites [41] Henin C, Le Métayer D (2021a) A multi-layered approach for tailored black-box explanations. Pattern recognition. ICPR international workshops and challenges. Springer Verlag, Cham, p 12663Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

The need to provide appropriate justifications of judicial decisions has been analyzed by several authors [26, 29, 73].

“The need to provide appropriate justifications of judicial decisions has also been analyzed by several authors [26, 29, 73] 24 .”

• cites [26] Cowls J, Floridi L (2018) Prolegomena to a white paper on an ethical framework for a good AI society. SSRN Electron J. https://doi.org/10.2139/ssrn.3198732 sourceNot checked yet

  Not checked in the free run.

• cites [29] de Licht KF, de Licht JF (2020) Artificial intelligence, transparency, and public decision-making. AI Soc 35:917–926 sourceNot checked yet

  Not checked in the free run.

• cites [73] Rouvroy A (2013) The end(s) of critique: data-behaviourism vs. due process. Privacy, due process and the computational turn. Philosophers of Law Meet Philosophers of Technology. Routledge, LondonNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

A good illustration of this approach is the Canadian Directive on Automated Decision-Making [35] which introduces four 'Impact Assessment Levels' associated with different 'Impact Level Requirements'.

“A good illustration of this approach is the Canadian Directive on Automated Decision-Making [35] which introduces four 'Impact Assessment Levels' associated with different 'Impact Level Requirements'.”

• cites [35] Fredriksson M, Tritter J (2017) Disentangling patient and public involvement in healthcare decisions: why the difference matters. Sociol Health Illn 39(1):95–111 sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

In this context, norms are made available to autonomous agents to make their decisions.

“In this context, norms are made available to autonomous agents to make their decisions.”

May combine two claims.

In this paper, we have also shown that justifiability and contestability are not illusory objectives: they can be operationalized in different ways for different types of recipients.

“In this paper, we have also shown that justifiability and contestability are not illusory objectives: they can be operationalized in different ways for different types of recipients.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

In the area of administration, [63] concludes that the process of giving a proper justification for an administrative decision involves more than just transparency for the sake of transparency, and that it is an argumentative process that provides targeted information.

“In the area of administration, [63] concludes that 'the process of giving a proper justification for an administrative decision involves more than just transparency for the sake of transparency. It is an argumentative process that provides targeted information.'”

• cites [63] Mohseni S, Zarei N, Ragan ED (2020) A multidisciplinary survey and framework for design and evaluation of explainable AI systems. ACM Trans Interact Intell Syst 1:1Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S2 (checkable).

The role of the technical means presented here is to support an overall combination of legal, political and organizational framework to enhance the legitimacy of ADS.

“The role of the technical means presented here is to support an overall combination of legal, political and organizational framework to enhance the legitimacy of ADS.”

Probably not a substantial claim: may not pass S3 (load-bearing).

The project described in [41] does not describe precise mechanisms to support this type of contestation.

“However, it does not describe precise mechanisms to support this type of contestation.”

• cites [41] Henin C, Le Métayer D (2021a) A multi-layered approach for tailored black-box explanations. Pattern recognition. ICPR international workshops and challenges. Springer Verlag, Cham, p 12663Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

The advantages of justifications over explanations to enhance trust in ADS has been analyzed by Karl de Fine Licht and Jenny de Fine Licht [32] who argue that a limited form of transparency that focuses on providing justifications for decisions has the potential to provide sufficient grounds for perceived legitimacy in AI decision-making.

“The advantages of justifications over explanations (in the sense used in this paper) to enhance trust in ADS has also been analyzed by Karl de Fine Licht and Jenny de Fine Licht [32] who argue that 'a limited form of transparency that focuses on providing justifications for decisions has the potential to provide sufficient grounds for perceived legitimacy in AI decision-making.'”

• cites [32] European Commission (2020) Proposal for a regulation of the European Parliament and of the Council on a Single Market For Digital Services (Digital Services Act) and amending Directive 2000/31/EC.Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing). May combine two claims.

AI and ADS can bring great benefits to society but there is strong risk of public backlash if ethical issues are not considered seriously.

“AI and ADS can bring great benefits to society but there is strong risk of public backlash if ethical issues are not considered seriously [10, 21, 32, 62] 25 .”

• cites [10] BEUC The European Consumer Organization (2019) Artificial Intelligence: what consumers say. Findings and policy recommendations of a multi-country survey on AI. https://www.beuc.eu/publications/beuc-x-2020-078_artificial_intelligence_what_consumers_say_report.pdfNot checked yet

  Not checked in the free run.

• cites [21] Castelluccia C, Le Métayer D (2020) Position paper: analyzing the impacts of facial recognition. In: Antunes L, Naldi M, Italiano GF et al (eds) Privacy technologies and policy. 8th Annual Privacy Forum, APF 2020. Springer International Publishing, Cham, pp 43–57Not checked yet

  Not checked in the free run.

• cites [32] European Commission (2020) Proposal for a regulation of the European Parliament and of the Council on a Single Market For Digital Services (Digital Services Act) and amending Directive 2000/31/EC.Not checked yet

  Not checked in the free run.

• cites [62] Mittelstadt B, Russell C, Wachter S (2018) Explaining explanations in AI. Proc Conf Fairness Account Transp. https://doi.org/10.1145/3287560.3287574 sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

The framework introduced in [51] relies on norms (called 'principles') inferred from ethical judgments using inductive logic programming.

“For example, the framework introduced in [51] relies on norms (called 'principles') inferred from ethical judgments using inductive logic programming.”

• cites [51] Langley P (2019) Explainable, normative, and justified agency. Proc AAAI Conf Artif Intell 33:9775–9779. https://doi.org/10.1609/aaai.v33i01.33019775 sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

The notion of 'design transparency' proposed by Michele Loi and his co-authors [52] is close to what the paper described as 'justifications for experts' in Section 4.1.

“As far as operationalization is concerned, the notion of 'design transparency' proposed by Michele Loi and his co-authors [52] is close to what we described as 'justifications for experts' in Section 4.1.”

• cites [52] Laugel T, Lesot M-J, Marsala C, et al (2019) The dangers of post-hoc interpretability: unjustified counterfactual explanations. In: Proceedings of the 28th International Joint Conference on Artificial Intelligence (IJCAI-19). https://www.ijcai.org/proceedings/2019/0388.pdf sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

The proposal for a regulation 'on a Single Market For Digital Services (Digital Services Act)' [30] published by the European Commission in December 2020 also emphasises the ability for users of online platforms to contest decisions.

“In this regard, it is worth noting that the proposal for a regulation 'on a Single Market For Digital Services (Digital Services Act)' [30] published by the European Commission in December 2020 also emphasises the ability for users of online platforms to contest decisions.”

• cites [30] Doshi-Velez F, Kim B (2017) Towards a rigorous science of interpretable machine learning. https://arxiv.org/abs/1702.08608Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

A lot of work has been done on argumentation and dialog games, but the focus in these areas is generally the logical structure of the framework to express and to relate arguments or the protocol to exchange arguments.

“A lot of work has also been done on argumentation and dialog games [6, 11, 51, 53] but the focus in these areas is generally the logical structure of the framework to express and to relate arguments or the protocol to exchange arguments.”

• cites [6] Atkinson K, Baroni P, Giacomin M et al (2017) Towards artificial argumentation. AI Mag 38:25–36. https://doi.org/10.1609/aimag.v38i3.2704 sourceNot checked yet

  Not checked in the free run.

• cites [11] Bex F, Walton D (2011) Combining explanation and argumentation in dialogue. Argum Comput 7:55–68 sourceNot checked yet

  Not checked in the free run.

• cites [51] Langley P (2019) Explainable, normative, and justified agency. Proc AAAI Conf Artif Intell 33:9775–9779. https://doi.org/10.1609/aaai.v33i01.33019775 sourceNot checked yet

  Not checked in the free run.

• cites [53] Lei T, Barzilay R, Jaakkola T (2016) Rationalizing neural predictions. In: Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing. Association for Computational Linguistics, pp 107–117 sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

The term 'justification' is sometimes used in the field of autonomous agents to refer to motivations to perform a specific action.

“The term 'justification' is sometimes used in the field of autonomous agents to refer to motivations to perform a specific action.”

Probably not a substantial claim: may not pass S3 (load-bearing).

The 'human-in-the-loop' approach leverages on human feedback during the training process to obtain more accurate classifiers.

“The 'human-in-the-loop' approach leverages on human feedback during the training process to obtain more accurate classifiers [46].”

• cites [46] Irving G, Christiano P, Amodei D (2018) AI safety via debate. https://arxiv.org/abs/1805.00899Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

The work in [43] relies on 'debates' between two competing algorithms exchanging arguments and counterarguments to convince a human user that their classification is correct.

“Closer to the notion of justification, [43] relies on 'debates' between two competing algorithms exchanging arguments and counterarguments to convince a human user that their classification is correct.”

• cites [43] Hildebrandt M (2019) Privacy as protection of the incomputable self: from agnostic to agonistic machine learning. Theor Inq Law 20:83–122 sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

The limitations of transparency (seen as 'looking inside the black box') for accountability has already been argued in [4].

“The limitations of transparency (seen as 'looking inside the black box') for accountability has already been argued in [4].”

• cites [4] Ananny M, Crawford K (2018) Seeing without knowing: limitations of the transparency ideal and its application to algorithmic accountability. New Media Soc 20:973–989. https://doi.org/10.1177/1461444816676645 sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

In the field of explainable AI, the distinction between explanations and justifications is sometimes blurred.

“In the field of explainable AI, the distinction between explanations and justifications is sometimes blurred.”

Probably not a substantial claim: may not pass S2 (checkable).

In this paper, we have argued that justifiability and contestability are key conditions to ensure the legitimacy of ADS that can have a significant impact on individuals' lives or on society in general.

“In this paper, we have argued that justifiability and contestability are key conditions to ensure the legitimacy of ADS that can have a significant impact on individuals' lives or on society in general.”

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing). May combine two claims.

Ethical regulation of the design and use of AI is a complex but necessary task; the alternative may lead to devaluation of individual rights and social values, rejection of AI-based innovation, and ultimately a missed opportunity to use AI to improve individual well-being and social welfare.

“As firmly asserted by Mariarosaria Taddeo and Luciano Floridi [70], 'Ethical regulation of the design and use of AI is a complex but necessary task. The alternative may lead to devaluation of individual rights and social values, rejection of AI-based innovation, and ultimately a missed opportunity to use AI to improve individual well-being and social welfare.'”

• cites [70] Peter F (2017) Political legitimacy. In: Zalta EN (ed) The stanford encyclopedia of philosophy. Springer, BerlinNot checked yet

  Not checked in the free run.

Classified as probably not a substantial claim.

Algocate is neutral, in the sense that it is designed to find the best arguments for both parties.

“Algocate is neutral, in the sense that it is designed to find the best arguments (i.e. statements with evidence supporting them) for both parties.”

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based).

It is appropriate to present justifications as a type of explanation in this setting since norms are internalised in agents.

“It is worth noting that it is appropriate to present justifications as a type of explanation in this setting since norms are in internalised in agents.”

Probably not a substantial claim: may not pass S2 (checkable).

This neutrality is of prime importance, given the usual imbalance of powers between individuals who are affected by the decisions and the designers or operators of the ADS.

“This neutrality is of prime importance, given the usual imbalance of powers between individuals who are affected by the decisions and the designers or operators of the ADS.”

• AssumptionRests on an assumption: There is a usual imbalance of powers between individuals who are affected by the decisions and the designers or operators of the ADS.

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

The goal of the work in [43] is to 'align an agent's actions with the values and preferences of humans' which is seen as a 'training-time problem'.

“However, the goal of this work is to 'align an agent's actions with the values and preferences of humans' which is seen as a 'training-time problem'.”

• cites [43] Hildebrandt M (2019) Privacy as protection of the incomputable self: from agnostic to agonistic machine learning. Theor Inq Law 20:83–122 sourceNot checked yet

  Not checked in the free run.

Classified as probably not a substantial claim.

Giving absolute primacy to either the decision-maker or the decision-subject would render algorithmic accountability too one-sided, allowing one party to hold the other to standards that they could not reasonably accept.

“As stated by Reuben Binns [12], 'Giving absolute primacy to either the decisionmaker or the decision-subject would render algorithmic accountability too onesided, allowing one party to hold the other to standards that they could not reasonably accept.'”

• cites Binns [12] Binns R (2018) Algorithmic accountability and public reason. Philos Technol 31:543–556 sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based).

A value-driven agent refers to an ethical preference ordering of the actions before making any decision.

“In a nutshell, a valuedriven agent refers to an ethical preference ordering of the actions before making any decision.”

Classified as probably not a substantial claim.

Even though the Canadian Directive applies only to administrative decisions of the federal government, it shows the way to the integration of algorithmic impact assessment into the development cycle of all critical ADS.

“Even though it applies only to administrative decisions of the federal government, it shows the way to the integration of algorithmic impact assessment into the development cycle of all critical ADS.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

The recent recommendations of the European Parliament [31] point in the same direction, considering that the determination of whether AI, robotics and related technologies should be considered high-risk, and thus subject to mandatory compliance with legal obligations and ethical principles, should always follow from an impartial, regulated and external ex-ante assessment based on concrete and defi

“In Europe, the recent recommendations of the European Parliament [31] points in the same direction, considering 'that the determination of whether artificial intelligence, robotics and related technologies should be considered high-risk, and thus subject to mandatory compliance with legal obligations and ethical principles as set out in the regulatory framework for AI, should always follow from an”

• cites [31] Doshi-Velez F, Kortz M, Budish R, et al (2019) Accountability of AI under the law: the role of explanation. https://arxiv.org/ftp/arxiv/papers/1711/1711.01134.pdfNot checked yet

  Not checked in the free run.