Cited by 132Ethics and Social Impacts of AIExplainable Artificial Intelligence (XAI)Digitalization, Law, and Regulation

The fallacy of inscrutability

Open access
Kroll, Joshua A.·2018·Philosophical Transactions of the Royal Society A: Mathematical, Physical and Engineering Sciences

Abstract

Contrary to the criticism that mysterious, unaccountable black-box software systems threaten to make the logic of critical decisions inscrutable, we argue that algorithms are fundamentally understandable pieces of technology. Software systems are designed to interact with the world in a controlled way and built or operated for a specific purpose, subject to choices and assumptions. Traditional power structures can and do turn systems into opaque black boxes, but technologies can always be understood at a higher level, intensionally in terms of their designs and operational goals and extensionally in terms of their inputs, outputs and outcomes. The mechanisms of a system's operation can always be examined and explained, but a focus on machinery obscures the key issue of power dynamics. While structural inscrutability frustrates users and oversight entities, system creators and operators always determine that the technologies they deploy are fit for certain uses, making no system wholly inscrutable. We investigate the contours of inscrutability and opacity, the way they arise from power dynamics surrounding software systems, and the value of proposed remedies from disparate disciplines, especially computer ethics and privacy by design. We conclude that policy should not accede to the idea that some systems are of necessity inscrutable. Effective governance of algorithms comes from demanding rigorous science and engineering in system design, operation and evaluation to make systems verifiably trustworthy. Rather than seeking explanations for each behaviour of a computer system, policies should formalize and make known the assumptions, choices, and adequacy determinations associated with a system. This article is part of the theme issue ‘Governing artificial intelligence: ethical, legal, and technical opportunities and challenges’.

Integrity signals

Registry

Crossref · 10.1098/rsta.2018.0084

Cites

46 of 46 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Claims the paper makes

0 of 36 cited sources checked · 0 supported · 0 flagged

  • Findingcitation

    The COMPAS audit reveals a fairness issue that was not previously observed: black arrestees are rated incorrectly as having a high risk of being re-arrested upon release nearly twice as often as white arrestees.

    “the audit nonetheless reveals a fairness issue that was not previously observed: black arrestees are rated incorrectly as having a high risk of being re-arrested upon release nearly twice as often as white arrestees.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [39]

      Not checked in the free run.

  • Findingcitation

    A review by ProPublica discovered a disparity in false positive rates for black arrestees in the COMPAS evaluation.

    “For example, a review by ProPublica famously discovered a disparity in false positive rates for black arrestees in the COMPAS evaluation, which is used to support bail decisions and other criminal justice professional judgement [39].”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [39]

      Not checked in the free run.

  • Interpretationreasoning

    The claimed inscrutability of computer systems is a category error because critics ascribe values to mechanical technologies and not to the humans who designed, built and fielded them.

    “That is, the claimed inscrutability of computer systems is a category error: by claiming that a system's actions cannot be understood, critics ascribe values to mechanical technologies and not to the humans who designed, built and fielded them.”
    Quote in the paperNot a substantial claim
  • Interpretationnone given

    Traditional power structures can and do turn systems into opaque black boxes.

    “Traditional power structures can and do turn systems into opaque black boxes, but technologies can always be understood at a higher level, intensionally in terms of their designs and operational goals and extensionally in terms of their inputs, outputs and outcomes.”
    Quote in the paperNot a substantial claim
  • Interpretationreasoning

    Because learning technologies are difficult to understand at a mechanistic level, applying them without considering phenomena or ends creates a barrier to understanding.

    “Because such learning technologies are, as a general matter, difficult to understand at a mechanistic level, such an approach serves to create a barrier to understanding.”
    Quote in the paperNot a substantial claim
  • Classified as probably not a substantial claim. 60
  • May combine two claims. 30
  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation

    Tests are often considered to be some of the best and most up-to-date documentation of a software system, to the point that some methodologies advocate writing tests in advance of writing the actual software, as tests constitute a de facto specification for a system.

    “For this reason, tests are often considered to be some of the best and most up-to-date documentation of a software system, to the point that some methodologies advocate writing tests in advance of writing the actual software, as tests constitute a de facto specification for a system [27].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [27]

      Not checked in the free run.

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

    Interpretationnone given

    A right to explanation fails to provide individual understanding of particular decisions.

    “Such a right sidesteps questions of aggregate effects like group fairness and nondiscrimination, while failing to provide individual understanding of particular decisions or to address power dynamics between individuals and computer systems or to facilitate governance of those”
    Quote in the paperNot a substantial claim
  • Probably not a substantial claim: may not pass S2 (checkable).

    Interpretationreasoning

    A focus on machinery obscures the key issue of power dynamics.

    “The mechanisms of a system's operation can always be examined and explained, but a focus on machinery obscures the key issue of power dynamics.”
    Quote in the paperNot a substantial claim
  • Probably not a substantial claim: may not pass S4 (based). May combine two claims.

    Premisenone given

    Software products have far and away more defects and bugs than other kinds of engineered artefacts, to the point that most people inside and outside the industry consider bugs and defects in software inevitable.

    “Further, software products have far and away more defects and bugs than other kinds of engineered artefacts, to the point that most people inside and outside the industry consider bugs and defects in software inevitable.”
    Quote in the paperNot a substantial claim
  • Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

    Interpretationreasoning

    The role of detailed, mechanistic explanations in creating understandable software systems is much smaller than the role of careful validation, experimentation, and a focus on capturing robust mechanisms, especially causal ones.

    “The role of detailed, mechanistic explanations in creating understandable software systems is much smaller than the role of careful validation, experimentation, and a focus on capturing robust mechanisms, especially causal ones.”
    Quote in the paperNot a substantial claim
  • May combine two claims.

    Interpretationreasoning

    Explanation is not an unalloyed good, because it is only useful when it properly engages the context of the tool being explained and because explanations at a technical level do not necessarily provide understanding or improve interpretability.

    “Explanation is not an unalloyed good, both because it is only useful when it properly engages the context of the tool being explained and because explanations, at a technical level, do not necessarily provide understanding or improve the interpretability of a particular technical tool.”
    Quote in the paperNot a substantial claim
  • Probably not a substantial claim: may not pass S4 (based). May combine two claims.

    Interpretationreasoning

    The priorities of the organization managing the software development are mediated into the product and tested for fit with the market rapidly, causing the benefits of sustained iteration to accrue not to overall social good, but to the software development organization alone.

    “Instead, the priorities of the organization managing the software development are mediated into the product and tested for fit with the market rapidly, causing the benefits of sustained iteration to accrue not to overall social good, but to the software development organization alone.”
    Quote in the paperNot a substantial claim
  • Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

    Interpretationnone given

    Systems can be understood in terms of their design goals and the mechanisms of their construction and operation.

    “Systems can be understood in terms of their design goals and the mechanisms of their construction and operation.”
    Quote in the paperNot a substantial claim
  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisenone given

    Iterative methodologies allow things learned during development, implementation and maintenance to be fed back into design objectives to create requirements for future iterations.

    “To improve these sequential engineering paradigms, the software industry invented iterative methodologies, where the phases form a true cycle and the things learned during development, implementation and maintenance can be fed back into design objectives to create requirements for future iterations of the software product.”
    Quote in the paperNot a substantial claim
  • Probably not a substantial claim: may not pass S4 (based).

    Interpretationnone given

    A right to explanation sidesteps questions of aggregate effects like group fairness and nondiscrimination.

    “Such a right sidesteps questions of aggregate effects like group fairness and nondiscrimination, while failing to provide individual understanding of particular decisions or to address power dynamics between individuals and computer systems or to facilitate governance of those”
    Quote in the paperNot a substantial claim
  • May combine two claims.

    Interpretationcitation

    Transparency of code and data alone does not provide understanding; there are fundamental technical limits on the capacity for software analysis.

    “Transparency of these instruments alone does not provide understanding, however: there are fundamental technical limits on the capacity for software analysis [15,16].”
    Quote in the paperSubstantial claim0 of 2 sources checked
    • cites [15]

      Not checked in the free run.

    • cites [16]

      Not checked in the free run.

  • May combine two claims.

    Interpretationreasoning

    A right to explanation fails to address power dynamics between individuals and computer systems or to facilitate governance of those systems.

    “Such a right sidesteps questions of aggregate effects like group fairness and nondiscrimination, while failing to provide individual understanding of particular decisions or to address power dynamics between individuals and computer systems or to facilitate governance of those”
    Quote in the paperSubstantial claim
  • Classified as probably not a substantial claim.

    Interpretationreasoning

    In iterative models, design goals are a living product of the process of software engineering rather than a static input to software development.

    “In these models, design goals are a living product of the process of software engineering rather than a static input to software development.”
    Quote in the paperSubstantial claim
  • Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

    Interpretationreasoning

    Understanding how computer systems are designed and built, including trade-offs between competing goals, dispels much of the inscrutability the systems may have when viewed as a whole or from the perspective of someone affected.

    “Understanding how computer systems are designed and built, including understanding when trade-offs between competing goals were made by the designers and why, dispels much of the inscrutability the systems may have when viewed as a whole or from the perspective of someone affected by such a system.”
    Quote in the paperNot a substantial claim
  • Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

    Interpretationreasoning

    It may in many cases be unnecessary to understand the precise mechanisms of an algorithmic system, so long as we describe the outlines of the system's interaction with the world.

    “It may in many cases be unnecessary to understand the precise mechanisms of an algorithmic system, just as we do not understand how humans make decisions, so long as we describe the outlines of the system's interaction with the world (e.g.doing good science, understanding causal mechanisms versus correlations).”
    Quote in the paperNot a substantial claim
  • Probably not a substantial claim: may not pass S4 (based).

    Interpretationreasoning

    Because both the software product and its specification are living outputs of the engineering process, approval, review and control suffer.

    “Further, because both the software product and its specification are living outputs of the engineering process, approval, review and control suffer.”
    Quote in the paperNot a substantial claim
  • Probably not a substantial claim: may not pass S4 (based). May combine two claims.

    Premisenone given

    Software systems are designed to interact with the world in a controlled way and are built or operated for a specific purpose, subject to choices and assumptions.

    “Software systems are designed to interact with the world in a controlled way and built or operated for a specific purpose, subject to choices and assumptions.”
    Quote in the paperNot a substantial claim
  • Probably not a substantial claim: may not pass S2 (checkable).

    Interpretationreasoning

    Inscrutability in these contexts is a choice, perhaps one born of laziness, on the part of the system's controllers.

    “But this barrier is illusory-inscrutability in these contexts is a choice, perhaps one born of laziness, on the part of the system's controllers.”
    Quote in the paperNot a substantial claim
  • Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

    Interpretationnone given

    Too often, systems are built irresponsibly by simply applying learning technologies to data without considering what phenomena the results should capture or to what ends the resulting system should function.

    “That is not to say that all systems are of necessity understandable-too often, systems are built (in whole or in part) irresponsibly by simply applying learning technologies to data without considering what phenomena the results should capture or to what ends the resulting system should function.”
    Quote in the paperNot a substantial claim
  • Probably not a substantial claim: may not pass S2 (checkable). May combine two claims.

    Interpretationreasoning

    Sometimes it is necessary to consider a system's context dynamically and to account for how the system changes and is changed by its context.

    “Sometimes, it is necessary to consider a system's context dynamically and to account for how the system changes and is changed by its context.”
    Quote in the paperNot a substantial claim
  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    It is common and a recognized best practice for software to be tested for consistency with its specification.

    “Regardless of the development methodology being used, it is common and a recognized best practice for software to be tested for consistency with its specification [26].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [26]

      Not checked in the free run.

  • May combine two claims.

    Interpretationreasoning

    Technologies can always be understood at a higher level, intensionally in terms of their designs and operational goals and extensionally in terms of their inputs, outputs and outcomes.

    “Traditional power structures can and do turn systems into opaque black boxes, but technologies can always be understood at a higher level, intensionally in terms of their designs and operational goals and extensionally in terms of their inputs, outputs and outcomes.”
    Quote in the paperNot a substantial claim
  • Probably not a substantial claim: may not pass S2 (checkable).

    Interpretationreasoning

    Because controllers have determined it is appropriate to use technology for particular ends, they have also provided an understanding of the outcomes created by their technology: the reasons they have for believing that the system is adequate to that purpose.

    “In turn, because they have determined that it is appropriate to use technology for particular ends, a system's controllers have also provided an understanding of the outcomes created by their technology: the reasons they have for believing that the system is adequate to that purpose.”
    Quote in the paperNot a substantial claim
  • May combine two claims.

    Premisecitation

    Large software development projects are more prone to failing both to produce outputs that meet the objectives of their customers and to produce outputs at all.

    “Because software systems are developed without the process rigidities of other types of engineering work or the constraints of physical reality, large software development projects are more prone to failing both to produce outputs that meet the objectives of their customers and to produce outputs at all [22].”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [22]

      Not checked in the free run.

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

    Interpretationnone given

    The presence of such defects is a choice, not an inevitability.

    “But the presence of such defects is a choice, not an inevitability.”
    Quote in the paperNot a substantial claim
  • May combine two claims.

    Premisenone given

    In domains where perfect foresight is impossible and outcomes are not foreseeable, such as medicine, practitioners are held to a standard of practice and are punished when their actions do not rise to a sufficient standard of care.

    “In many domains where perfect foresight is impossible and where the outcomes from interventions are not foreseeable, such as medicine, we nonetheless hold practitioners to a standard of practice and are comfortable punishing practitioners when their actions do not rise to a sufficient standard of care.”
    Quote in the paperSubstantial claim
  • Classified as probably not a substantial claim. May combine two claims.

    Premisereasoning

    Testing measures the functionality of software both in the sense that it reveals defects and also in the sense that it verifies that the software acts according to its goals.

    “Testing measures the functionality of software both in the sense that it reveals defects and also in the sense that it verifies that the software acts according to its goals.”
    Quote in the paperSubstantial claim
  • Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

    Interpretationreasoning

    Any inability to describe how or why a particular advertisement was presented in a particular situation is merely a design choice, not an inevitability of the complexity of large systems.

    “Any inability to describe how or why a particular advertisement was presented in a particular situation is merely a design choice, not an inevitability of the complexity of large systems.”
    Quote in the paperNot a substantial claim
  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

    Interpretationreasoning

    Because iterative process focuses on iterative improvement of an early prototype, outcomes can be path dependent and it is common for only the most obvious requirements to be addressed, leaving little time for features which are important but only affect a small part of a system's user base.

    “However, this flexibility comes at a cost: because the process focuses on iterative improvement of an early prototype, outcomes can be path dependent and it is common for only the most obvious requirements to be addressed, leaving little time for features which are important, but only affect a small part of a system's user base.”
    Quote in the paperNot a substantial claim
  • May combine two claims.

    Interpretationnone given

    Agile software development can improve the speed at which functional software is delivered, but it does so by creating incentives to avoid thinking deeply about the consequences of technical requirements or engaging stakeholders outside the day-to-day development process.

    “While agile software development can improve the speed at which functional software is delivered, it does so by creating incentives to avoid thinking deeply about the consequences of technical requirements or engaging stakeholders outside the day-today development process.”
    Quote in the paperNot a substantial claim
  • Probably not a substantial claim: may not pass S2 (checkable), S4 (based). May combine two claims.

    Interpretationreasoning

    Even here, systems can be designed to sidestep these limits and to support interrogation-inscrutability remains a choice that can be accommodated for in system design.

    “But even here, systems can be designed to sidestep these limits and to support that interrogation-inscrutability remains a choice that can be accommodated for in system design.”
    Quote in the paperNot a substantial claim
  • Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

    Interpretationnone given

    Systems can also be understood in terms of their inputs and outputs and the outcomes that result from their application in a particular context.

    “Additionally, systems can also be understood in terms of their inputs and outputs and the outcomes that result from their application in a particular context.”
    Quote in the paperNot a substantial claim
  • Probably not a substantial claim: may not pass S2 (checkable), S4 (based). May combine two claims.

    Interpretationnone given

    Explanations must speak to the decisions made during the design of a computer system, as such information is always available and always fulfils the key requirements of a meaningful explanation.

    “We offer that explanations must speak to the decisions made during the design of a computer system, as such information is always available and always fulfils the key requirements of a meaningful explanation.”
    Quote in the paperNot a substantial claim
  • Classified as probably not a substantial claim.

    Interpretationnone given

    Even systems that are carefully designed can fail to capture important facts about the world.

    “Even carefully designed systems can miss important facts about the world, and it is important to verify that systems are operating as intended and satisfying their articulated requirements.”
    Quote in the paperSubstantial claim
  • Classified as probably not a substantial claim.

    Interpretationreasoning

    If a system's specification properly reflects its requirements and design goals, then fidelity to that specification demonstrates fitness for purpose.

    “If a system's specification properly reflects the system's requirements and design goals-that is, if the system is the result of proper requirements engineering-then fidelity to that specification demonstrates fitness for purpose.”
    Quote in the paperSubstantial claim
  • May combine two claims.

    Interpretationcitation

    Audits cannot, for technical reasons, ever guarantee that they will completely suss out the mechanisms of a computer system, but they can, in many cases, provide useful information about when a system is causing harm or when it is behaving in a way different from what is expected.

    “While audits cannot, for technical reasons, ever guarantee that they will completely suss out the mechanisms of a computer system [15,16], they can, in many cases, provide useful information about when a system is causing harm or when it is behaving in a way different from what is expected.”
    Quote in the paperSubstantial claim0 of 2 sources checked
    • cites [15]

      Not checked in the free run.

    • cites [16]

      Not checked in the free run.

  • May combine two claims.

    Interpretationreasoning

    Transparency neither requires nor is provided by open source or source availability alone.

    “However, transparency neither requires nor is provided by either open source or source availability alone.”
    Quote in the paperSubstantial claim
  • Probably not a substantial claim: may not pass S2 (checkable).

    Interpretationreasoning

    Failing to make a system reproducible when it should be constitutes malpractice.

    “Because the reproducibility of computer outputs is a design choice, we can again say that failing to make a system reproducible when it should be constitutes malpractice.”
    Quote in the paperNot a substantial claim
  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Interpretationcitation

    Sufficient transparency may simply mean disclosing the fact and scope of data processing in a computer system.

    “To that end, sufficient transparency may simply mean disclosing the fact and scope of data processing in a computer system, as is required by the GDPR in the EU.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites GDPR

      Not checked in the free run.

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

    Interpretationnone given

    Explicitly debiasing data risks inventing counterfactuals that are not supportable or testable during later validation phases.

    “Ideally, bias is accounted for as assumptions are passed through a system; explicitly debiasing data risks inventing counterfactuals which are not supportable or testable during later validation phases.”
    Quote in the paperNot a substantial claim
  • Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based).

    Premisenone given

    Examining group-level system outcomes is a cornerstone of quantitative evaluation for policy interventions of all kinds.

    “Beyond individual decisions, examination of group-level system outcomes is a cornerstone of quantitative evaluation for policy interventions of all kinds.”
    Quote in the paperNot a substantial claim
  • May combine two claims.

    Interpretationreasoning

    Understanding the design process that led to a software system helps elucidate the mechanisms by which the system operates, in turn providing insight into why particular outcomes or outputs were produced.

    “Understanding the design process that led to a software system helps elucidate the mechanisms by which the system operates, in turn providing insight into why particular outcomes or outputs were produced.”
    Quote in the paperNot a substantial claim
  • May combine two claims.

    Premisecitation

    Similar audits have turned up issues of bias reinforcement in systems that guide police to portions of a city to patrol, and in e-commerce platforms.

    “Similar audits have turned up issues of bias reinforcement in systems that guide police to portions of a city to patrol [36], and in e-commerce platforms [35].”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [36] source

      Not checked in the free run.

    • cites [35] source

      Not checked in the free run.

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

    Interpretationreasoning

    Viewing computer system behaviours as uncontrollable ignores the fact that these systems are human artefacts, built to a purpose by some human agency that must be accountable for the behaviours of those artefacts.

    “Viewing computer system behaviours as though they are uncontrollable ignores the fact that these systems are human artefacts, built to a purpose by some human agency that must be accountable for the behaviours of those artefacts.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: Human agents who design, build, and field computer systems are accountable for the behaviours of those systems.
  • Probably not a substantial claim: may not pass S4 (based). May combine two claims.

    Interpretationreasoning

    All software is designed and built by people, and responsible organizations will create traces of their decision-making processes which provide insight into the meaning and function of the artefacts they produce.

    “All software is designed and built by people, and responsible organizations will create traces of their decision-making processes which provide insight into the meaning and function of the artefacts they produce.”
    Quote in the paperNot a substantial claim
  • May combine two claims.

    Interpretationreasoning

    In turn, black arrestees find themselves systematically less able to mount a defence to the case against them, reinforcing historical and structural discrimination.

    “In turn, black arrestees find themselves systematically less able to mount a defence to the case against them, reinforcing historical and structural discrimination.”
    Quote in the paperSubstantial claim
  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Interpretationreasoning

    The extent to which particular details are or are not foreseeable may differ depending on the discipline and background of the viewer.

    “The extent to which particular details are or are not foreseeable may differ depending on the discipline and background of the viewer.”
    Quote in the paperNot a substantial claim
  • Probably not a substantial claim: may not pass S4 (based).

    Interpretationnone given

    Calls for requiring explanations of computer system behaviour as a path to governing those systems often do not address the question of who is meant to receive what knowledge through the explanation.

    “Yet, calls for requiring explanations of computer system behaviour as a path to governing those systems often do not address the question of who is meant to receive what knowledge through the explanation.”
    Quote in the paperNot a substantial claim
  • Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

    Interpretationnone given

    Malpractice involves situations where bad outcomes could have been avoided by more responsible behaviour on the part of a system's controllers.

    “Importantly, the mere fact of a mistake is insufficient to define malpractice; rather, malpractice involves situations where bad outcomes could have been avoided by more responsible behaviour on the part of a system's controllers.”
    Quote in the paperNot a substantial claim
  • Probably not a substantial claim: may not pass S2 (checkable). May combine two claims.

    Interpretationreasoning

    Explanations must create understanding by humans of computer systems in context, though they are not on their own necessary or sufficient to create that understanding.

    “In either case, explanations must create understanding by humans of computer systems in context, though they are not on their own necessary or sufficient to create that understanding.”
    Quote in the paperNot a substantial claim
  • Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

    Premisenone given

    Any system is a model of the world.

    “Any system is a model of the world, and too often the called-for explanation is only a description of that model.”
    Quote in the paperNot a substantial claim
  • Probably not a substantial claim: may not pass S2 (checkable). May combine two claims.

    Interpretationcitation

    Good explanations must not only explain why a behaviour occurred or why some other behaviour did not occur as a product of a system, but also capture and transmit the knowledge of the context underlying a system's mechanism of action as a process within a broad sociotechnical system.

    “Good explanations must not only explain why a behaviour occurred or why some other behaviour did not occur as a product of a system, but also capture and transmit the knowledge of the context underlying a system's mechanism of action as a process within a broad sociotechnical system [30].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [30]

      Not checked in the free run.

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

    Interpretationnone given

    Not all computer systems are easily analysable.

    “Of course, this does not mean that all computer systems are easily analysable.”
    Quote in the paperNot a substantial claim
  • Classified as probably not a substantial claim.

    Interpretationreasoning

    Open-source software strengthens transparency by combining source code disclosure with the ability to derive additional knowledge by recompiling or modifying the software.

    “The concept of 'open-source' software, or software which has source code that is both freely available and permissively licensed, strengthens transparency by combining detailed source code disclosure with the ability to derive additional knowledge by recompiling or modifying the software.”
    Quote in the paperSubstantial claim
  • May combine two claims.

    Interpretationcitation

    Because systems are designed to achieve particular ends, we cannot disentangle the environmental contributions to a particular behaviour from the designed mechanism-the mechanism must compensate for its environment if it correctly embodies its goals.

    “Because systems are designed to achieve particular ends, we cannot disentangle the environmental contributions to a particular behaviour from the designed mechanism [31]-the mechanism must, after all, compensate for its environment if it correctly embodies its goals.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [31] source

      Not checked in the free run.

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

    Interpretationreasoning

    No artefact is properly comprehended without reference to its human context, and software systems are no different.

    “No artefact is properly comprehended without reference to its human context, and software systems are no different.”
    Quote in the paperNot a substantial claim
  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    The theoretical potential for discrimination in automated systems is well established.

    “The theoretical potential for discrimination in automated systems is well established [15,41,42].”
    Quote in the paperNot a substantial claim0 of 3 sources checked
    • cites [15]

      Not checked in the free run.

    • cites [41]

      Not checked in the free run.

    • cites [42]

      Not checked in the free run.

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

    Interpretationnone given

    It can be difficult to understand at the design stage how a system will interact with the world.

    “Additionally, it can be difficult to understand at the design stage how a system will interact with the world.”
    Quote in the paperNot a substantial claim
  • Classified as probably not a substantial claim.

    Interpretationcitation

    Fidelity to a system's specification is one aspect of determining whether that system is correct.

    “Fidelity to a system's specification is one aspect of determining whether that system is correct, although we might also demand that the specification itself satisfy design goals or relate properly to the real world [16].”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [16]

      Not checked in the free run.

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

    Premisenone given

    Design efforts that chose a tool as right for a context exist at least informally for all systems.

    “The design efforts that chose that tool as the right one for that context-and such efforts exist at least at an informal level for all systems-explain who believes that tool is the proper one and why it is correct to make use of it.”
    Quote in the paperNot a substantial claim
  • Probably not a substantial claim: may not pass S2 (checkable).

    Interpretationreasoning

    Good explanations require careful consideration of who receives the explanation and what facts those people must extract from it.

    “While these technologies promise much in the way of making systems more trustworthy, good explanations require careful consideration of who receives the explanation and what facts those people must extract from the explanations.”
    Quote in the paperNot a substantial claim
  • Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing).

    Interpretationreasoning

    In other cases, an unrelated justification of the result or a general description of how it was reached is sufficient.

    “In other cases, as with human decision-makers, it is sufficient to produce an unrelated justification of the result or a general description of how it was reached.”
    Quote in the paperNot a substantial claim
  • May combine two claims.

    Interpretationcitation

    Disclosure of source code or data is neither necessary nor sufficient for improving understanding of a system.

    “While transparency is often taken to mean the disclosure of source code or data, possibly to a trusted entity such as a regulator, this is neither necessary nor sufficient for improving understanding of a system [15], and it does not capture the full meaning of transparency (though transparency is often equated with disclosure of system internals in technical communities and scholarship).”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [15]

      Not checked in the free run.

  • Classified as probably not a substantial claim.

    Interpretationreasoning

    The choice of task can dispel inscrutability by focusing analysis on the sociotechnical system involving humans and their tools.

    “The choice of this task can also dispel inscrutability, as it takes the focus off of possibly complex software systems and puts it on the sociotechnical system involving humans and their tools, concentrating analysis on where trustworthiness is necessary.”
    Quote in the paperSubstantial claim
  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Interpretationreasoning

    Impact assessments help the controllers of a system understand how to set their goals, or how to achieve their goals in light of the world as it is.

    “Further, they help the controllers of a system understand how to set their goals, or how to achieve their goals in light of the world as it is.”
    Quote in the paperNot a substantial claim
  • Probably not a substantial claim: may not pass S4 (based).

    Interpretationnone given

    Even well-designed systems may in some cases err or cause negative outcomes for their subjects.

    “For example, even well-designed systems following the pattern above may in some cases err or cause negative outcomes for their subjects.”
    Quote in the paperNot a substantial claim
  • Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

    Interpretationnone given

    The validation of non-causal reasoning systems is limited by their lack of undergirding mechanism and the extent to which their constructs are valid.

    “Non-causal reasoning is possible, but the way in which such systems can be validated is limited by their lack of undergirding mechanism and the extent to which their constructs are valid.”
    Quote in the paperNot a substantial claim
  • Probably not a substantial claim: may not pass S4 (based). May combine two claims.

    Interpretationnone given

    Information about the design process almost always exists and can improve understanding of a computer system, but such information generally is not available to the people who need it for review.

    “While information about the design process almost always exists and can improve understanding of a computer system, such information generally is not available to the people who need it for review.”
    Quote in the paperNot a substantial claim
  • Probably not a substantial claim: may not pass S4 (based).

    Interpretationnone given

    In some cases, a mechanistic description of how a system computed a particular output sufficiently explains that output.

    “In some cases, a mechanistic description of how a system computed a particular output will sufficiently explain that output.”
    Quote in the paperNot a substantial claim
  • Probably not a substantial claim: may not pass S2 (checkable).

    Interpretationreasoning

    Impact assessment processes are useful tools for governing systems when precise consequences are not readily foreseeable.

    “Several scholars propose the use of an impact assessment process to aid in governing those systems [17][18][19]33], and assessments can even be legally required, as in Article 35 of the General Data Protection Regulation (GDPR) in Europe.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [17][18][19]33]

      Not checked in the free run.

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

    Interpretationreasoning

    Computer system behaviours should not be held to a 'best-effort' standard because they are unforeseeable.

    “We further reject that computer system behaviours should be held to a 'best-effort' standard because they are unforeseeable.”
    Quote in the paperNot a substantial claim
  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation

    In safety-critical applications, software systems are presented together with assurance cases, documents that are meant to convince a sceptical expert that the system operates as intended, and which detail the steps taken to assure fidelity of the system to its stated goals.

    “In safety-critical applications, software systems are presented together with assurance cases, documents that are meant to convince a sceptical expert that the system operates as intended, and which detail the steps taken to assure fidelity of the system to its stated goals [28,29].”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [28]

      Not checked in the free run.

    • cites [29]

      Not checked in the free run.

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

    Interpretationreasoning

    Failure to understand why relationships exist between data elements risks designing systems that codify existing unfairness and reinforce problematic power dynamics.

    “Failure to understand why relationships exist between data elements risks designing systems which codify existing unfairness and which reinforce problematic power dynamics.”
    Quote in the paperNot a substantial claim
  • Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based).

    Interpretationnone given

    Inscrutability is a difficult problem in practice, in no small part because software systems have become incredibly powerful.

    “Inscrutability is a difficult problem in practice, in no small part because software systems have become incredibly powerful.”
    Quote in the paperNot a substantial claim
74 of 74
  • The paper relies on this without support.

    Human agents who design, build, and field computer systems are accountable for the behaviours of those systems.

    “built to a purpose by some human agency that must be accountable for the behaviours of those artefacts.”
    “Responsibility and ethics attach not to the specifics of a technical tool, but rather to the ways that tool is used in a sociotechnical context, which are always considered when tools are created.”
  • The paper relies on this without support.

    Interventions to protect users or subjects of computer systems should engage power relationships between controllers and affected parties.

    “Interventions that aim to protect the subjects or users of a computer system must therefore engage these power relationships.”
2 of 2

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

  • It is not impossible to close the gap between principles and practice if the right questions are asked and closer attention is paid to how the design process can influence whether an algorithm is more or less 'ethically-aligned.'

    “Yet, it is not impossible if the right questions are asked (Green 2018; Wachter et al. 2017) and closer attention is payed to how the design process can influence (Kroll 2018) whether an algorithm is more or less ‘ethically-aligned.’”
  • The thesis that the understanding and transparency of algorithmic-assisted decision-making is limited by the inscrutability of the machine learning models and their algorithms is criticizable.

    “As showed by Kroll (2018), the thesis that the understanding and transparency of algorithmic-assisted decision-making is limited by the inscrutability of the machine learning models and their algorithms (i.e. the fact that they are opaque or “black boxes”) is criticizable.”

Cites

Flagged first, then by citations

7 of the 53 references couldn’t be matched to a known paper.

Research works printed but not linked: 4

  1. Selbst AD Barocas S. The intuitive appeal of explainable machines. Preprint available at SSRN. (https://ssrn.com/abstract=3126971)
  2. Jones C, 2010, Software engineering best practices: lessons from successful projects in top companies
  3. Argyris C, 1977, Double loop learning in organizations, Harv. Bus. Rev., 55, 115
  4. Myers GJ, 1979, The art of software testing

Cited by

Sorted by citations

Authors
Kroll, J. A.
Affiliations
University of California, Berkeley (US)
Corresponding author
Kroll, Joshua A.
Venue
Philosophical Transactions of the Royal Society A: Mathematical, Physical and Engineering Sciences
Published
2018-10-15
Volume
376
Issue
2133
Pages
20180084
ISSN
  • 1364-503X (ISSN-L)
  • 1471-2962
OpenAlex
W2897439270
Cited by
132
References
53
Citations by year
2026: 9 · 2025: 13 · 2024: 15 · 2023: 22 · 2022: 27 · 2021: 19 · 2020: 17 · 2019: 8 · 2018: 2
FWCI
22.85
Type
journal article
Language
en
Open access
Locations
  • Philosophical Transactions of the Royal Society A Mathematical Physical and Engineering Sciences (publishedVersion)
  • Europe PMC (PubMed Central) (publishedVersion)
  • PubMed Central (publishedVersion)
  • PubMed (publishedVersion)
  • eScholarship (California Digital Library) (submittedVersion)
Funders
Berkeley Center for Law and Technology
Topics
  • Social Sciences › Safety Research › Ethics and Social Impacts of AI
  • Physical Sciences › Computer Science › Artificial Intelligence › Explainable Artificial Intelligence (XAI)
  • Social Sciences › Law › Digitalization, Law, and Regulation
Keywords
algorithmic opacity, algorithmic governance, algorithmic accountability, computer ethics, privacy by design, algorithmic transparency, algorithmic decision-making

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Claims read by deepseek-v4.1-flash under definition version 3 (extract-v3) on 2026-10-05.

Classified by jev-1.13-20260917 (gold-guideline-v0.2) on 2026-10-05.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review