Transparency as design publicity: explaining and justifying inscrutable algorithms

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

Design publicity does not require that accountable individuals provide fully persuasive and non-corrigible justifications; it is sufficient that they declare what they take to be the relevant elements, exposing themselves to public scrutiny.

“Notice that design publicity does not require that individuals that are accountable for algorithmic decisions provide fully persuasive and non-corrigible justifications. It is sufficient that they declare what they take to be the relevant elements, exposing themselves to public scrutiny, as the above-mentioned FR case exemplifies.”

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

The debate on algorithm inscrutability mostly depends on the meaning we attribute to the expression "explaining the model" and accordingly "understanding the model."

“The debate on algorithm inscrutability mostly depends on the meaning we attribute to the expression “explaining the model” and accordingly “understanding the model.””

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

The goal of an algorithm is usually a practical objective, such as profit or efficient allocation of scarce resources, but can include moral values such as equity, beneficence, trustworthiness, and socially accepted rules for the domain.

“The goal of an algorithm is usually a practical objective, such as profit or efficient allocation of scarce resources, but can include moral values such as equity, beneficence, trustworthiness, and the rules that are socially accepted as pertinent for the domain in which the model is employed.”

Classified as probably not a substantial claim.

In both cases, the goal introduces normativity in the model, as it represents something that there are good reasons to pursue.

“In both cases, the goal introduces normativity in the model, as it represents something that there are good reasons to pursue.”

Classified as probably not a substantial claim.

The first step of the justification of decisions taken by an algorithm requires evaluating the goals and constraints that the algorithm respectively achieves and respects.

“The first step of the justification of decisions taken by an algorithm, thus, requires evaluating the goals and constraints that the algorithm, respectively, achieves and respects.”

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

Design publicity can be more easily linked with the justification of the use and design of the algorithm, and of each individual decision following from it.

“We argue that design publicity can be more easily linked with the justification of the use and of the design of the algorithm, and of each individual decision following from it.”

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

If the goals and constraints pursued by a model do not reflect values worth pursuing, the decisions following from the model are not justified.

“If the goals and constraints pursued by a model do not reflect values worth pursuing, the decisions following from the model are not justified.”

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

Counterfactual explanations represent a practical strategy to explanation in presence of few variables and scenario choices, where the assumption that the model makes decisions that are normatively appropriate is taken for granted.

“Summing up, counterfactual explanations do represent a practical strategy to explanation in presence of few variables and scenario choices, where the assumption that the model makes decisions that are normatively appropriate is taken for granted.”

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

Transparency of machine learning algorithms, just as explanation, can be defined at different levels of abstraction.

“In this paper we argue that transparency of machine learning algorithms, just as explanation, can be defined at different levels of abstraction.”

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

Design publicity provides information about the goal and moral constraints the algorithm is designed to pursue and respect, the translation of that goal into a machine-learning problem, the performance of the algorithm, and proof that decisions are taken by consistently applying the same algorithm.

“Design publicity provides information about (a) the goal the algorithm is designed to pursue and the moral constraints it is designed to respect (value transparency); (b) the way this goal is translated into a problem that can be solved by machine learning (translation transparency); (c) the performance of the algorithm in addressing problem (performance transparency); and (d) a proof of the fact”

Classified as probably not a substantial claim.

The explainee of counterfactual explanations must accept as a presupposition that the decision is taken based on certain features.

“The explainee of counterfactual explanations must accept as a presupposition that the decision is (reasonably?) taken based on certain features.”

May combine two claims.

Both design transparency and consistency transparency are necessary for design publicity, because they are necessary to assess if and how the decision taken by or with the help of an algorithmic system is justified.

“In what follows, we argue that both design transparency and consistency transparency, as defined in “Design explanation of algorithms,” are necessary for design publicity, because they are necessary to assess if and how the decision taken by (or with the help of) an algorithmic system is justified (when it is).”

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

There is no straightforward and unique way to translate a goal into a mathematic construct.

“There is no straightforward and unique way to translate a goal into a mathematic construct.”

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

Value transparency requires that normative goals and the reasons for considering them are clearly specified, so that the choice of such normative goal is not a mere arbitrary decision by data scientists.

“Value transparency requires that these normative goals and the reasons for considering them are clearly specified—i.e. the choice of such normative goal is not a mere arbitrary decision by the data”

Classified as probably not a substantial claim.

Design publicity is intended to empower the public to debate choices about algorithmic fairness, so as to enable their revision.

“Design publicity is intended to empower the public to debate also such choices, so as to enable their revision.”

Classified as probably not a substantial claim.

The division of design transparency into three components, value, translation, and performance, enables the analysis of the algorithm at three different levels of explanations, which constitute three levels that a design explanation has to address.

“The division of design transparency into three components (value, translation, performance) enables the analysis of the algorithm at three different levels of explanations, which, as we will see, constitute also three levels that a design explanation has to address.”

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

At a level of abstraction where explanations clarify the purpose of algorithms, these explanations would be understandable to the public, from end users with low expertise to policymakers, corporate executives, and computer scientists and engineers.

“On the other hand, one could consider a level of abstraction where explanations clarify the purpose of algorithms; these would be understandable to the public, from the end users with low expertise to policymakers in the need of justifying the use of algorithmic-assisted decision-making, to corporate executives adopting such models, to computer scientists and engineers that design them.”

May combine two claims.

The most normatively important properties of model-based decisions are properties of the kind of patterns, such as the distribution of errors or benefits between groups, that emerge when the law of large numbers applies, as well as model accuracy and ability to generate benefits.

“they are properties of the kind of patterns, e.g. the distribution of errors, or of benefits, between groups, or of groups defined by morally salient properties, that emerge when the law of large numbers applies, as well as the model accuracy and ability to generate benefits (e.g. profit, or other forms of utility).”

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

The explainee of counterfactual explanations is modelled as having exclusively self-centered, concrete, pragmatic interest in the specific features that are relevant for the decision about her or him.

“The explainee is modelled as having exclusively self-centered, concrete, pragmatic interest in the specific features that are relevant for the decision about her or him.”

Classified as probably not a substantial claim.

Transparency as design publicity serves the particular purpose of normative justification.

“Our transparency idea serves a particular purpose: that of normative justification.”

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

In a justified algorithm, the goals and constraints reflect those values and constraints a reasonable person may want to see promoted or respected in the context of a service.

“In a justified algorithm, they reflect those values and constraints a reasonable person may want to see promoted/respected in the context of a service.”

May combine two claims.

Indirect discrimination or disparate impact can be morally or legally relevant in certain contexts but cannot be determined by reference to counterfactual explanations, because a protected group may be discriminated via proxy.

“A case in point of a morally salient property is indirect discrimination or disparate impact, which can be considered morally or legally relevant in certain contexts, but cannot be determined by reference to counterfactual explanations, because a protected group may be discriminated via proxy (e.g. ZIP code can be a proxy of race), so the protected group information will not appear as a feature in”

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

If the value translation is not declared and no reason to accept it is given, the decisions of the model are not even prima facie justified.

“If the value translation is not declared and no reason to accept it is given, the decisions of the model are not even prima facie justified.”

May combine two claims.

Different definitions of fairness for predictive models exist and it is often impossible to satisfy all of them simultaneously.

“Different definitions of fairness for predictive models exist and it is often impossible to satisfy all of them simultaneously (Berk et al. 2018).”

• cites Berk et al. 2018 sourceNot checked yet

  Not checked in the free run.

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

In comparison to post-hoc explanations of individual algorithmic decisions, design publicity meets a different demand, the demand for impersonal justification, of the explainee.

“In comparison to post-hoc explanations of individual algorithmic decisions, design publicity meets a different demand (the demand for impersonal justification) of the explainee.”

• AssumptionThe authors state this as an assumption: Individuals are able to take, or at least aspire to take, an impersonal standpoint when judging algorithmic systems.

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

This publicity is achieved thanks to special procedures, e.g. the consistent fulfillment of legal obligations emerging from civil law can be tested by going to court.

“This publicity is achieved thanks to special procedures. E.g. the consistent fulfillment of the legal obligations emerging from civil law can be tested by going to court.”

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

The design explanation of the model shows that an algorithmic model gives a decisional outcome because the model pursues a certain goal, which is translated into mathematical constructs implemented in the algorithm, which in turn enables one to verify whether the model achieves the goal.

“In conclusion, the design explanation of the model shows that an algorithmic model gives a decisional outcome because the model pursues a certain goal (value transparency), which is translated into mathematical constructs implemented in the algorithm (translation transparency), which in turn enables one to verify whether the model achieves the goal (performance transparency).”

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

This challenge would violate consistency.

“This would violate consistency.”

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

The translation of a moral constraint into a quantifiable performance measure should be given a normative grounding, and not simply be assumed.

“With reference to the FR example above, the translation of a moral constraint (e.g. anti-discrimination) into a quantifiable performance measure (e.g. equality in the false positive and false negative rate across racial and sex groups) should be given a normative grounding, and not simply be assumed.”

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

The performance can only be assessed by translating the goals in question into measurable quantities.

“The performance can only be assessed by translating the goals in question into measurable quantities.”

May combine two claims.

Performance metrics provide an important indication of the extent to which every value has been achieved, which is especially important for the overall justification of the system when a value can only be achieved at the expense of another value.

“Performance metrics provide an important indication of the extent to which every value has been achieved, which is especially important for the overall justification of the system when a value can only be achieved at the expense of another value.”

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

In cases of high-stake decisions, any change due to retraining should be at least publicized and justified by pointing out a considerable improvement in performance, which overrides consistency concerns.

“In such cases, any change due to retraining should be at least publicized, and justified, by pointing out a considerable improvement in performance, which overrides consistency concerns.”

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

If the same exception were made for everyone who had an interest to demand it, the algorithm would not achieve its design goals, which justify its use.

“on the other hand, if the same exception were made for everyone who had an interest to demand it, the algorithm wouldn’t achieve its design goals, which justify its use.”

May combine two claims.

Consistency transparency is necessary but not sufficient to justify each decision as fair.

“Nonetheless, as consistency transparency shows that identical cases are treated identically, it represents the first step towards fairness; it is a sort of basic requirement of fairness that, as we shall show, is necessary but not sufficient to justify each decision as fair.”

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

The primary goal of the algorithm is essential to show that decisions are not morally arbitrary.

“The primary goal of the algorithm is essential to show that decisions are not morally arbitrary.”

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

An algorithm pursuing such goals will achieve them to a determined degree, which is expressed by performance transparency.

“An algorithm pursuing such goals will achieve them to a determined degree, which is expressed by “performance transparency.””

Classified as probably not a substantial claim.

The paper proposes a form of transparency that consists in publicizing the design of an artifact, including value, translation and performance, as well as its consistent application.

“We propose a form of transparency that consists in publicizing the design of an artifact (including value, translation and performance) as well as its consistent application.”

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

Consistency transparency requires that changes in a model be declared because this is relevant for their justification.

“Consistency transparency requires that changes in a model be declared because, as we shall maintain, this is relevant for their justification.”

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

To assess if inscrutable algorithms satisfy their principles of justice, we consider their performance.

“To assess if inscrutable algorithms satisfy their “principles of justice” we consider their performance.”

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

Different types of justification are possible, for example in terms of common or philosophical morality, of the law, or by virtue of political principles and values that may be universal or characteristic of the society in which the model operates.

“Different types of justification are possible, for example in terms of common or philosophical morality, of the law, of by virtue of political principles and values that may be universal or characteristic of the society in which the model operates.”

Classified as probably not a substantial claim.

An individual subjected to an unfavorable decision may accept, in principle, that the algorithm is justified as a whole, yet challenge the necessity of implementing the model when taking a decision about him.

“An individual subjected to an unfavorable decision may accept, in principle, that the algorithm is justified as a whole, yet challenge the necessity of implementing the model when taking a decision about him.”

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

In the algorithmic case, the consistent application of inscrutably complex rules appears to lack transparency.

“In the algorithmic case, the consistent application of inscrutably complex rules appears to lack transparency.”

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

Decisions based on imperfect predictions about stochastic events will typically be often wrong, but sufficiently often right to justify the use of the model in practice.

“This is because decisions based on imperfect predictions about stochastic events will typically be often wrong, but sufficiently often right to justify the use of the model in practice.”

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

This kind of transparency provides an explanation of the artifact by design, an intentional explanation of its deployment, a justification of its use, and, when used consistently, a procedural justification of the individual decisions it takes.

“We maintain that this kind of transparency provides (1) an explanation of the artifact, namely, an explanation “by design”; (2) an intentional explanation of its deployment; (3) a justification of its use; (4) when used consistently, a procedural justification of the individual decisions it takes.”

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

Primary goals matter to justification when they are valuable goals, e.g. there are good reasons to pursue such goals, which can be explained by reference to values commonly accepted in society.

“Primary goals matter to justification when they are valuable goals, e.g. there are good reasons to pursue such goals, which can be explained by reference to values commonly accepted in society, including moral, political or legal values, as well as the profit generated by free market exchanges, in capitalist societies that rely on profit as the motive stimulating socially efficient economic”

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

Other goals, the constraining goals, typically reflect value considerations, e.g. privacy or fairness.

“Other goals (the “constraining” goals) typically reflect value considerations, e.g. privacy or fairness.”

Classified as probably not a substantial claim.

Publicly recognized forms of social utility, such as security or profit, may fill the purpose of showing that decisions are not morally arbitrary.

“Publicly recognized forms of social utility, such as security (in the FR example) or profit may fill this purpose.”

Classified as probably not a substantial claim.

The algorithm's performance connects the explanation of the artifact with the application of the rule to particular cases.

“The algorithm’s performance connects the explanation of the artifact (i.e. an algorithm, or rule) with the application of the rule to particular cases.”

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

For algorithmic decision-making, the principles of justice correspond to its design goals.

“For algorithmic decision-making, the principles of justice correspond to its design goals.”

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

Consistency is a normative goal, and showing that it is achieved by the model contributes to explaining why an individual decision is made by showing that it is explained by a normative consideration.

“Consistency is a normative goal and showing that it is achieved by the model contributes to explaining why an individual decision is made–namely, by showing that it is explained by a normative consideration.”

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

When consistency is among the reasonable goals of model deployment, the explanation of the decisions by the model includes consistency transparency.

“When, as in most cases, consistency is among the reasonable goals of model deployment, the explanation of the decisions by the model includes consistency transparency.”