CoChat
Sign Up
Cited by 31Ethics and Social Impacts of AIExplainable Artificial Intelligence (XAI)Digitalization, Law, and Regulation

Out of Context: Investigating the Bias and Fairness Concerns of “Artificial Intelligence as a Service”

Open access
Lewicki, Kornel, Lee, Michelle Seng Ah, Cobbe, Jennifer, Singh, Jatinder·2023·Proceedings of the 2023 CHI Conference on Human Factors in Computing Systems
DOI 10.1145/3544548.3581463
Open this paper in CoChat — free accountOpen PDF

Abstract

“AI as a Service” (AIaaS) is a rapidly growing market, offering various plug-and-play AI services and tools. AIaaS enables its customers (users)—who may lack the expertise, data, and/or resources to develop their own systems—to easily build and integrate AI capabilities into their applications. Yet, it is known that AI systems can encapsulate biases and inequalities that can have societal impact. This paper argues that the context-sensitive nature of fairness is often incompatible with AIaaS’ ‘one-size-fits-all’ approach, leading to issues and tensions. Specifically, we review and systematise the AIaaS space by proposing a taxonomy of AI services based on the levels of autonomy afforded to the user. We then critically examine the different categories of AIaaS, outlining how these services can lead to biases or be otherwise harmful in the context of end-user applications. In doing so, we seek to draw research attention to the challenges of this emerging area.

AI snapshot

Written from the paper’s full text.

TL;DR

The paper argues that AI-as-a-Service’s generic “one-size-fits-all” model conflicts with the context-sensitive nature of fairness, creating bias risks across AutoML platforms, AI APIs, and fully-managed AI services. It proposes a taxonomy of AIaaS by user autonomy and uses experiments and examples to show how provider assumptions and opacity can propagate unfair outcomes.

Key findings

  • The paper proposes a taxonomy of AIaaS with three categories reflecting decreasing user control and technical expertise: AutoML Platforms, AI APIs, and Fully-managed AI Services.
  • In an Azure AutoML experiment on Adult, German Credit, and COMPAS, models selected solely by predictive performance were often more biased than alternatives; the paper reports up to a 12% difference for the German dataset split on the sex attribute, while fairer alternatives with only marginally lower accuracy existed.
  • In age-estimation experiments using APIs from Amazon, Baidu, Face++, and Microsoft on a subset of UTKFace, unitary race groups showed no obvious differences, but intersectional age/race disaggregation revealed all classifiers performed worse for the 60+ population; Amazon’s 60+ Asian subgroup had an average error of almost 20 years, and over at least 15.5 years for all remaining groups.
  • For fully-managed services, provider-chosen fairness definitions can conflict with customer contexts or legal requirements; the paper cites COMPAS debates over false positive rate parity versus positive predictive value parity, and notes Raghavan et al. found evaluated hiring providers used the U.S. EEOC 4/5ths rule.

Methods

  • The paper combines a critical review of AIaaS vendors’ claims and practices, empirical experimentation with select AIaaS services, and theoretical analysis; the taxonomy was derived from prior research, market reports, blogs/news, and provider websites/marketing materials.
  • AutoML experiment: used Azure’s AutoML service on Adult, German Credit, and COMPAS datasets, comparing the platform-designated “optimal” model against the model achieving best demographic parity difference.
  • AI API experiment: examined age-estimation APIs from Amazon, Baidu, Face++, and Microsoft using a subset of UTKFace, alongside qualitative analysis of provider disclosures and bias-mitigation support.

Limitations

  • The paper states its experiments are not a comprehensive analysis of bias in any particular AI service and should be viewed as a snapshot of the current AIaaS landscape, since services may be updated or changed at any time.
  • The paper states its taxonomy is not comprehensive, future-proof, or definitive, and it excludes open-source tools, AI model libraries, large language models, and generative AI tools from its scope.
  • Evident from the text: the empirical work is limited to selected providers, datasets, and tasks (AutoML on tabular fairness datasets; age estimation on UTKFace), and the paper does not conduct a user study of AIaaS customers or practitioners, so its findings may not generalize across all AIaaS contexts.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1145/3544548.3581463

Cites

135 of 135 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Claims the paper makes

0 of 124 cited sources checked · 0 supported · 0 flagged

  • Findingown data

    Disaggregating the race groups into under and over 60 years old reveals that all classifiers perform worse on the age estimation task for the 60+ population.

    “As shown in Table 3, disaggregating these subgroups into populations of under and over 60 years old reveals that all classifers perform worse on the age estimation task for the 60+ population.”
    Quote in the paperSubstantial claim
  • Findingown data

    Disaggregating the subgroups further into UN's standard age brackets makes the lack of representation of certain subgroups even more apparent.

    “Further, as seen in Table 4, disaggregating those subgroups even further into UN's standard age brackets [114], makes this lack of representation of certain subgroups even more apparent.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [114]

      Not checked in the free run.

  • Findingown data

    Age estimation performance, as measured by MAE, varies across race groups for the evaluated commercial AI services.

    “Table 3 : Age estimation performance across age and race groups as measured by Mean Absolute Error (MAE) metric.”
    Quote in the paperSubstantial claim
  • Findingown data

    Age estimation performance, as measured by MAE, is worse for older age groups (60+) than for younger age groups (under 60) across race groups for the evaluated commercial AI services.

    “Table 4 : Age estimation performance across age and race groups as measured by Mean Absolute Error (MAE) metric.”
    Quote in the paperSubstantial claim
  • Interpretationreasoning

    Understanding, let alone accounting for all bias concerns in an effort to create a 'generally applicable' AI service appears infeasible.

    “That is, as we discussed previously, one of the key difculties in uncovering and mitigating unintended bias in AI systems is that many fairness issues will manifest only in particular contexts, meaning that understanding, let alone accounting for all bias concerns in an efort to create a 'generally applicable' AI service appears infeasible.”
    Quote in the paperNot a substantial claim
  • May combine two claims. 46
  • Classified as probably not a substantial claim. 47
  • The quote only partly states this claim. 1

  • May combine two claims.

    Findingown data

    Amazon's API age estimates on the 60+ subgroup exhibit an average error of almost 20 years for the Asian subgroup and over at least 15.5 years for all remaining groups.

    “Notably, Amazon's API age estimates on the 60+ subgroup exhibit an average error of almost 20 years for the Asian subgroup and over at least 15.5 years for all remaining groups.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingcitation

    Baidu's beauty estimation service outputs a single 'beauty score' for any image, while Face++ gives two outputs to represent beauty scores 'from both male's and female's perspectives'.

    “Interestingly, we see that Baidu's service outputs a single 'beautyscore' for any image, while Face++ gives two outputs to represent beauty scores”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [100]

      Not checked in the free run.

  • May combine two claims.

    Findingown data

    For age estimation across race groups using APIs from Amazon, Baidu, Face++, and Microsoft on a subset of UTKFace, all APIs perform relatively well with no obvious differences between different race groups.

    “Table 2 shows that all APIs perform relatively well, displaying no obvious diferences between diferent race groups.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    In models created using Azure AutoML on the Adult, German Credit and COMPAS datasets, the model deemed 'optimal' by predictive performance often exhibits more bias than possible alternatives (up to 12% difference for the German dataset split on the sex attribute), while fairer alternatives with only marginally lower accuracy exist.

    “We fnd that it is indeed often the case that as a result of selecting models solely by their predictive performance, the models deemed as 'optimal' might in fact exhibit more bias than possible alternatives (up to 12% diference for German dataset split on the sex attribute), while 'fairer' alternatives with only marginally lower levels of accuracy do exist.”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim.

    Interpretationreasoning

    The AIaaS taxonomy comprises three service categories: AutoML Platforms, AI APIs, and Fully-Managed AI services.

    “Table 1 represents the resulting taxonomy comprising three service categories: (1) AutoML Platforms, (2) AI APIs, and (3) Fully-Managed AI services (which we describe below).”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationreasoning

    Issues of bias in AIaaS can arise both from the nature of the underlying ML model itself and the operating context in which a user employs that service.

    “As such, issues of bias can arise both from the nature of the underlying ML model itself and the operating context in which a user employs that service.”
    Quote in the paperSubstantial claim

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

    Interpretationnone given

    These categories reflect users' decreasing level of involvement, required technical expertise and control over the service and the underlying ML model(s), as well as increasingly specialised types of problems they address.

    “These categories refect the perspective of users' (decreasing) level of involvement, required technical expertise and control over the service and the underlying ML model(s), as well as the increasingly specialised types of problems they address.”
    Quote in the paperNot a substantial claim

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

    Interpretationreasoning

    Identifying the service types helps indicate and highlight certain properties and characteristics that can have fairness implications in practice.

    “however, identifying the service types helps indicate and highlight certain properties and characteristics that can have fairness implications in practice.”
    Quote in the paperNot a substantial claim

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

    Interpretationreasoning

    Fully-managed AI services model resembles traditional SaaS model, which removes the user from low-level technical complexities.

    “As such, the fully managed AI services model resembles that of a traditional SaaS (Software-as-a-Service) model, which removes the user from the low-level, technical complexities of these systems.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Premisecitation

    AIaaS enables customers who may lack expertise, data, and/or resources to develop their own systems to easily build and integrate AI capabilities into their applications.

    “Marketed as”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [7]

      Not checked in the free run.

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

    Interpretationnone given

    The categories are not mutually exclusive; some services may have characteristics that overlap the service types.

    “These categories are not mutually exclusive in that some services may have characteristics that overlap the service types; however, identifying the service types helps indicate and highlight certain properties and characteristics that can have fairness implications in practice.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationreasoning

    Through their services, AIaaS providers might propagate not only unintended bias, but also their own set of views, values and priorities.

    “As such, there is a risk that through their services, AIaaS providers might propagate not only unintended bias, but also their own set of views, values and priorities.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationreasoning

    AIaaS can amplify AI fairness problems and introduce additional risks because it makes widely-accessible, scalable, readily-available AI capable of underpinning a broad range of applications.

    “Though issues of fairness pervade AI in general, there is the potential for AIaaS-in making widelyaccessible, scalable and readily-available AI that is capable of underpinning a broad range of applications-to amplify AI fairness problems and introduce additional risks.”
    Quote in the paperSubstantial claim

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

    Interpretationreasoning

    AI APIs offerings can be grouped into four key categories: Data Analytics, Text, Speech, and Vision services.

    “and broadly many oferings can be grouped into four key categories of service:”
    Quote in the paperNot a substantial claim

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

    Premisecitation

    Commercial AutoML offerings can be categorised broadly into two groups: those offered by major cloud providers and those offered by dedicated AutoML organisations.

    “The commercial Au-toML oferings can be categorised broadly into two groups: those ofered by major cloud providers (e.g., Google, Microsoft), and those ofered by dedicated AutoML organisations (often startups, e.g., DataRobot, H2O.ai, 4Paradigm) [157].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [157] source

      Not checked in the free run.

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

    Interpretationnone given

    Beauty is an inherently context-dependent and highly-subjective concept, which naturally varies between individuals, cultures, and regions.

    “Beauty, however, is an inherently context-dependent and highly-subjective concept, which naturally varies between individuals, cultures, and regions.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisereasoning

    It is generally not possible to fully debias an AI system or guarantee its fairness; rather, the aim is to mitigate fairness-related harms as much as possible.

    “Accordingly, given the many complexities of fairness and its contextual nature, it is generally not possible to fully debias an AI system or to guarantee its fairness [80,101,122]; rather, the aim is to mitigate fairness-related harms and other unwanted consequences as much as possible [101,136,143].”
    Quote in the paperSubstantial claim0 of 5 sources checked
    • cites [80]

      Not checked in the free run.

    • cites [101]

      Not checked in the free run.

    • cites [122]

      Not checked in the free run.

    • cites [136] source

      Not checked in the free run.

    • cites [143] source

      Not checked in the free run.

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

    Interpretationnone given

    Providers' views are encapsulated in their models (to varying degrees) for all services.

    “While it is generally the case that providers' views are encapsulated in their models (to varying degrees) for all services”
    Quote in the paperNot a substantial claim

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

    Interpretationnone given

    Fully-managed services tend to deal with more numerous and/or complex sets of functionalities-targeting niche industry-specific needs, in which contextual-specifics may be of even greater importance-potentially amplifying existing or incurring new kinds of risks and tensions.

    “Moreover, fullymanaged services tend to deal with more numerous and/or complex set of functionalities-targeting niche industry-specifc needs, in which contextual-specifcs may be of even greater importancepotentially amplifying existing or incurring new kinds of risks and tensions.”
    Quote in the paperNot a substantial claim

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

    Interpretationreasoning

    Aiming to provide 'universal' functionality, which abstracts away or otherwise does not account for real-world differences, will inevitably lead to problems.

    “Aiming to provide 'universal' functionality, which abstracts away or otherwise does not account for real-world diferences will inevitably lead to problems.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisereasoning

    AI services can compound existing inequities by producing unfair outcomes, reinforcing pernicious stereotypes and disproportionately distributing negative consequences to those already marginalised.

    “Like any AI system, AI services can compound existing inequities by producing unfair outcomes, reinforcing pernicious stereotypes and disproportionately distributing negative consequences of technology to those already marginalised.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisecitation

    There is some work on fairness relating to AI services, though this tends to focus on either exposing or 'auditing' certain AI services, or devising tools, methods and interventions relevant for those building (rather than using) such services.

    “There is some work on fairness relating to AI services, though this tends to focus on either exposing or 'auditing' certain AI services [22,38,126,133], or devising tools, methods and interventions relevant for those building (rather than using) such services [12,96,106,128].”
    Quote in the paperSubstantial claim0 of 8 sources checked
    • cites [22]

      Not checked in the free run.

    • cites [38]

      Not checked in the free run.

    • cites [126] source

      Not checked in the free run.

    • cites [133] source

      Not checked in the free run.

    • cites [12] source

      Not checked in the free run.

    • cites [96] source

      Not checked in the free run.

    • cites [106] source

      Not checked in the free run.

    • cites [128] source

      Not checked in the free run.

  • May combine two claims.

    Interpretationreasoning

    AIaaS providers are primarily cloud technology giants and are in a powerful position.

    “This places AIaaS providers-primarily the cloud technology giants-in a powerful position.”
    Quote in the paperSubstantial claim

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

    Interpretationreasoning

    There will be many situations where a provider will not have considered the particulars of the social, economic, environment or other context of use; where a provider decides, judges or takes a position on particular concepts that are not or cannot be universally agreed upon; or where it is challenging, infeasible or inappropriate to reduce a range of distinct concepts, definitions and groups into

    “There will be many situations where a provider will not have considered the particulars of the social, economic, environment or other context of use; where a provider decides, judges or takes a position on particular concepts that are not or cannot be universally agreed upon; or where it is challenging, infeasible or inappropriate to reduce a range of distinct concepts, defnitions and groups into”
    Quote in the paperNot a substantial claim

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

    Premisecitation

    The nature of AI development-requiring volumes of training data, specialised hardware, and technical expertise-presents a significant challenge for organisations wishing to integrate AI into their business processes.

    “However, the nature of AI development-requiring volumes of training data, specialised hardware for building and training models, and technical expertise in machine learning-presents a signifcant challenge for organisations wishing to integrate AI into their business processes [35,44].”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [35] source

      Not checked in the free run.

    • cites [44] source

      Not checked in the free run.

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

    Premisereasoning

    AIaaS services reduce overheads and barriers to entry by offering and making accessible state-of-the-art tools/models.

    “given that such services reduce overheads and barriers to entry through providers ofering and making accessible state-of-the-art tools/models.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationreasoning

    The ML models underpinning fully managed services are often offered on an 'as-is' basis where no further user-led model retraining or larger customisation is possible, and are thus vulnerable to many of the challenges described for AI APIs.

    “Similar to the AI APIs, the ML models underpinning fully managed services are often ofered on an 'as-is' basis where no further user-led model retraining or larger customisation is possible, and are thus vulnerable to many of the challenges we have described in the previous section.”
    Quote in the paperSubstantial claim

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

    Interpretationreasoning

    Certain assumptions, values and ideals may hold in some contexts and situations but not in others.

    “The broader overarching point (inappropriate services aside) is that certain assumptions, values and ideals may hold in some contexts and situations but not in others.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Premisecitation

    Publicly available object-recognition services are less effective at recognising household items that are common in particular geographies and low-income communities.

    “For example, Terrance et al. [38] found that publicly available object-recognition services are less efective at recognising household items that are common in particular geographies and lowincome communities.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [38]

      Not checked in the free run.

  • May combine two claims.

    Premisecitation

    Today's AutoML systems tend to limit customers' interaction with the system to a few specific decision points and present only limited to no information on how these systems work in operation and the complex processes behind the ML models generation and selection.

    “Today's AutoML systems tend to limit customers' interaction with the system to a few specifc decision points and present only limited to no information on how these systems work in operation and the complex processes behind the ML models generation and selection [88,[150][151][152]157].”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [88,[150][151][152]157]

      Not checked in the free run.

  • May combine two claims.

    Premisecitation

    Machine learning models, such as those offered by API providers, are prone to exhibiting various kinds of discriminatory behaviours, including different error rates across demographic groups, stereotyping, or minority groups under-representation.

    “It is well acknowledged that machine learning models, such as those ofered by the API providers, are prone to exhibiting various kinds of discriminatory behaviours, including diferent error rates across demographic groups [22,126,127], stereotyping [64,163] or minority groups under-representation [134,137].”
    Quote in the paperSubstantial claim0 of 7 sources checked
    • cites [22]

      Not checked in the free run.

    • cites [126] source

      Not checked in the free run.

    • cites [127] source

      Not checked in the free run.

    • cites [64] source

      Not checked in the free run.

    • cites [163] source

      Not checked in the free run.

    • cites [134] source

      Not checked in the free run.

    • cites [137]

      Not checked in the free run.

  • May combine two claims.

    Interpretationreasoning

    These proposed laws do not address information, skills, and power imbalances between AIaaS providers and users, which may make it difficult for users to identify and mitigate problems with systems not under their control.

    “Moreover, these proposed laws do not address the information, skills, and power imbalances often present between AIaaS providers (again, typically large technology companies) and their service users (who may be much smaller organisations and/or lacking certain degrees of technical expertise), which may make it difcult if not impossible in practice for AIaaS users to identify and mitigate problems”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationown data

    The current reliance of AutoML frameworks on fairness-blind optimisation metrics combined with the inability to choose between the various models created in the AutoML process abstracts away the complex, domain-specific notion of 'optimality' and its relational nature, restricting users' ability to build a model that best reflects the values and objectives of their particular use case.

    “The current reliance of AutoML frameworks on fairness-blind optimisation metrics combined with the inability to choose between the various models created in the AutoML process abstract away the complex, domain-specific notion of 'optimality' and its relational nature.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Premisecitation

    The AIaaS market is consolidating predominantly around a few already dominant 'Big Tech' companies, primarily from the U.S. and China, which further entrenches their position and extends their societal influence.

    “Indeed, we already see a consolidation of the AIaaS market predominantly around the few already dominant 'Big Tech' companies (primarily from two regions: U.S. and China), which further entrenches their position and extends their societal infuence [4,29].”
    Quote in the paperSubstantial claim0 of 2 sources checked
    • cites [4]

      Not checked in the free run.

    • cites [29] source

      Not checked in the free run.

  • May combine two claims.

    Premisereasoning

    Access to fully managed AI services is by and large strictly limited to those directly engaging with the firm, or through provider-managed demos which often come with limits or restrictions on functionality and use, and thus might effectively prevent interrogation.

    “the access to fully managed AI services is by and large strictly limited to those directly engaging with the frm, or through provider-managed 'demos' which often come with limits or restrictions on functionality and use (and thus might efectively prevent interrogation).”
    Quote in the paperSubstantial claim

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

    Premisecitation

    Algorithmic recruitment services provide users (customers) with access to a suite of candidate assessments (questions, video interviews, games etc.) which are then algorithmically analysed by the provider and their outputs combined to score and rank candidates.

    “As an example consider algorithmic recruitment services (e.g.[65,123]), which provide users (customers) with access to a suite of candidate assessments (questions, video interviews, games etc.) which are then algorithmically analysed by the provider and their outputs combined to score and rank candidates.”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [65]

      Not checked in the free run.

    • cites [123]

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S1 (asserted). May combine two claims.

    Premisecitation

    These discriminatory issues are primarily a reflection of training set deficiencies, namely the underrepresentation of certain parts of the input space and the lack of awareness, consideration and attention of the people building these systems.

    “Many of these works conclude that these issues are primarily a refection of training set defciencies, namely the underrepresentation of certain parts of the input space [15,22] and the lack of awareness, consideration and attention of the people building these systems [97,136,146].”
    Quote in the paperNot a substantial claim0 of 5 sources checked
    • cites [15] source

      Not checked in the free run.

    • cites [22]

      Not checked in the free run.

    • cites [97] source

      Not checked in the free run.

    • cites [136] source

      Not checked in the free run.

    • cites [146] source

      Not checked in the free run.

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

    Interpretationreasoning

    Beauty scoring is something that cannot be generally modelled.

    “Naturally this is the case for beauty scoring, which is something that cannot be generally modelled”
    Quote in the paperNot a substantial claim

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

    Interpretationreasoning

    Algorithmic recruitment systems are concerned with allocating or withholding opportunities or resources; therefore, fairness is of key concern.

    “These systems are concerned with allocating or withholding opportunities or resources; therefore, fairness is of key concern.”
    Quote in the paperSubstantial claim

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

    Interpretationreasoning

    This limits the accessibility and usability of the AI service and presents a risk that these services and any supporting tools might be designed in a way that fails to align with user needs.

    “This limits the accessibility and usability of the AI service and presents a risk that these services and any supporting tools might be designed in a way that fails to align with user needs.”
    Quote in the paperSubstantial claim

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

    Interpretationcitation

    Transparency which supports broader accountability processes on training sets, models and internal processes behind these services, as well as acknowledgement of known limitations, past failures, and lessons learnt, can help move AIaaS from opaque, unknown and potentially harmful 'black-boxes' to more understandable, accepting of their own limitations, building blocks for affordable and widely acc

    “Transparency which supports broader accountability processes on training sets, models and internal processes behind these services, as well as in the widersense of acknowledgement of known-limitations, past failures, and lessons learnt [27,131] can help move AIaaS from the opaque, unknown and potentially harmful 'black-boxes' of today, to a more understandable, accepting of their own limitations,”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [27] source

      Not checked in the free run.

    • cites [131] source

      Not checked in the free run.

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

    Interpretationreasoning

    The majority of these providers are located in the U.S. and might thus be legally obliged to adhere to the 4/5ths rule.

    “This is understandable considering that majority of these providers are located in the U.S. and might thus be legally obliged to adhere to this rule.”
    Quote in the paperNot a substantial claim

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

    Interpretationreasoning

    There will be benefits in including AIaaS in discussions of everyday algorithm audits.

    “We argue that there will be benefts in including AIaaS in those discussions; for example, by exploring designs and methods that better support users in conducting more efective service audits for bias, and means for meaningfully presenting and communicating such fndings back to service providers.”
    Quote in the paperNot a substantial claim

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

    Interpretationreasoning

    By conducting their own experimentation as part of their provider/service vetting process, the users of AI services could themselves be well-placed to validate service performance in context of their specific use case.

    “However, by conducting their own experimentation-not unlike the experiments we presented in §5.1 and §5.2-as part of their provider/ service vetting process, the users of AI services could themselves be well-placed to validate service performance in context of their specifc use case.”
    Quote in the paperNot a substantial claim

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

    Interpretationnone given

    It is the users of AIaaS that have the advantage of being placed to know the precise context in which the AI service will be used, something the AIaaS providers are currently only generally able to try to anticipate.

    “Notably, it is the users of AIaaS that have the advantage of being placed to know the precise context in which the AI service will be used, something the AIaaS providers are currently only generally able to try to anticipate.”
    Quote in the paperNot a substantial claim

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

    Interpretationreasoning

    For some AI services-particularly those of an AutoML or AI API type-there might be scope for potential customers to take on a more proactive role and examine the suitability of a given service through their own algorithmic investigations.

    “Additionally, for some AI services-particularly those of an Au-toML or AI API type-there might be scope for potential customers to take on a more proactive role and examine the suitability of a given service through their own algorithmic investigations.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationreasoning

    AIaaS challenges established understandings of the roles and responsibilities of providers and users; in some cases, the users (customers) of those services may themselves become considered as the 'provider' under the Regulation for the underlying AI system in relation to their use of the service.

    “As we note in §2.2, however, AIaaS challenges established understandings of the roles and responsibilities of providers and users; in some cases, the users (customers) of those services may themselves become considered as the”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Premisecitation

    Studies exploring user (ML practitioner) attitudes and expectations regarding fairness toolkits showed a misalignment of tools with user expectations and needs, and generally being difficult to use, which not only makes such tooling less effective, but can potentially lead to increased bias risks.

    “These issues have been shown in more general ML-contexts; for example, studies exploring user (ML practitioner) attitudes and expectations regarding fairness toolkits showed a misalignment of tools with user expectations and needs, and generally being difcult to use (among other concerns), which not only makes such tooling less-efective, but can potentially lead to increased bias risks [40,91].”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [40] source

      Not checked in the free run.

    • cites [91] source

      Not checked in the free run.

  • May combine two claims.

    Premisenone given

    The customer base of these services is global and subject to a wide range of different jurisdictions which may or may not follow similar guidelines.

    “However, the customer base of these services is global and subject to wide range of diferent jurisdictions which may or may not follow similar guidelines.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationreasoning

    This can lead to potential tensions, where the definition of fairness chosen by the service provider is incompatible with the notion of fairness held by the customer, or the legal requirements of the market they operate in.

    “This can lead to potential tensions, where the defnition of fairness chosen by the service provider is incompatible with the notion of fairness held by the customer, or the legal requirements of the market they operate in.”
    Quote in the paperSubstantial claim

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

    Interpretationcitation

    There appears a clear role for adopting and extending various documentation approaches, be they relating to data and models (such as datasheets, model cards and factsheets), their interactions and interconnections (e.g. decision provenance), and broader, holistic socio-technical system reviews (e.g. reviewability and traceability).

    “There appears a clear role for adopting and extending various documentation approaches, be they relating to data and models (such as datasheets [53], model cards [106] and factsheets [12]), their interactions and interconnections (e.g.decision provenance [139]), and broader, holistic socio-technical system reviews (e.g.reviewability [27] and traceability [86]).”
    Quote in the paperNot a substantial claim0 of 3 sources checked
    • cites [53]

      Not checked in the free run.

    • cites [106] source

      Not checked in the free run.

    • cites [12] source

      Not checked in the free run.

  • May combine two claims.

    Interpretationreasoning

    By taking away the control of such context-dependent aspects from the customer, the fully managed AI service model risks imposing providers' world views and values onto their clients, which could lead to potential unintended consequences and tensions.

    “Again, by taking away the control of such context-dependent aspects from the customer, this fully managed AI service model risks imposing providers' world views and values onto their clients, which could lead to potential unintended consequences and tensions.”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim.

    Interpretationnone given

    In aiming to offer AI services generically to potentially millions of customers, AIaaS providers need to make certain assumptions or choices about where and how their services will be used and by whom.

    “In aiming to ofer AI services generically to potentially millions of customers, AIaaS providers need to make certain assumptions or choices about where and how their services will be used and by whom.”
    Quote in the paperSubstantial claim

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

    Interpretationreasoning

    Similar issues can arise in other fully managed AI services, where user needs and expectations may not align with, or indeed, be incompatible with the assumptions and approaches taken by the provider.

    “Similar issues can arise in other fully managed AI services, where user needs and expectations may not align with, or indeed, be incompatible with the assumptions and approaches taken by the provider.”
    Quote in the paperNot a substantial claim

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

    Premisenone given

    There are increasing moves towards regulating AI systems, including the EU's proposed AI Act which will apply to 'providers' of AI services.

    “As we have discussed, there are increasing moves towards regulating AI systems, including the EU's proposed AI Act which will apply to”
    Quote in the paperNot a substantial claim

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

    Premisecitation

    Users expressed interest in a more 'human-in-the-loop' AutoML that aims to empower users in undertaking ML development, rather than replace them completely.

    “Instead, they found that users expressed interest in a more”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [157] source

      Not checked in the free run.

  • Classified as probably not a substantial claim.

    Interpretationreasoning

    By embracing transparency at every stage of the AIaaS development process, including being upfront about the design contexts considered and any known limitations of the service while marketing it to potential users, some of the risks cautioned against in this paper could be reduced.

    “However, by embracing transparency at every stage of the AIaaS development process, including being upfront about the design contexts considered and any known limitations of the service while marketing it to potential users, some of the risks we have cautioned against in this paper could be reduced.”
    Quote in the paperSubstantial claim

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

    Premisecitation

    'Fair' can mean different things in different contexts to different people.

    “But 'fair' can mean diferent things in diferent contexts to diferent people [109].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [109] source

      Not checked in the free run.

  • May combine two claims. The quote only partly states this claim.

    Interpretationreasoning

    In a service context, such user-providers will lack anything like the information needed to meet their obligations; may not be in a position to understand or deal with such concerns, despite provider claims of 'no expertise required'; nor will they have the technical influence over such systems to be able to do so.

    “As this paper makes clear, in a service context such-user providers will therefore lack anything like the information needed to meet their obligations; may not be in a position to understand or deal with such concerns, despite provider claims of”
    Quote in the paperSubstantial claim

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

    Interpretationnone given

    Poorly implemented or poorly explained AI services risk that a customer would select a suboptimal or inappropriate tool for their use case, or simply decide against adopting AIaaS at all.

    “On the other hand, poorly implemented or poorly explained AI services risk that a customer would select a suboptimal or inappropriate tool for their use case, or simply decide against adopting AIaaS at all.”
    Quote in the paperNot a substantial claim

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

    Interpretationnone given

    The AIaaS model has the potential to make AI more widely accessible, making it easier, faster and cheaper to deliver state-of-the-art AI and its benefits to a much wider range of applications and groups than might otherwise be possible.

    “The AIaaS model has the potential to make AI more widely accessible, making it easier, faster and cheaper to deliver state-of-the-art AI and its benefts to a much wider range of applications and groups than might otherwise be possible.”
    Quote in the paperNot a substantial claim

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

    Premisecitation

    A recent line of HCI research has explored the idea of 'everyday algorithm audits' whereby users detect, understand, and interrogate problematic machine behaviours via their day-to-day interactions with algorithmic systems.

    “Indeed, a recent line of HCI research has explored the idea of”
    Quote in the paperNot a substantial claim0 of 3 sources checked
    • cites [39]

      Not checked in the free run.

    • cites [43] source

      Not checked in the free run.

    • cites [138] source

      Not checked in the free run.

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

    Interpretationreasoning

    The need to select a fairness definition also poses a challenge to the service user, who may not have all the information to assess for fairness.

    “It also poses a challenge to the service user, who may not have all the information to assess for fairness.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationreasoning

    Inevitably there will be situations in which the 'generic' assumptions underpinning AI services will fail to address or come into conflict with the specific, context-dependent needs of a specific user's use case, and in turn may render these services and their end-users' systems biased.

    “As we have shown in this work, inevitably there will be situations in which the 'generic' assumptions underpinning AI services will fail to address or come into confict with the specifc, context-dependent needs of a specifc user's use case (which the AIaaS providers are unlikely to have considered for all their potential customers) and in turn may render these services and their end-users' systems”
    Quote in the paperNot a substantial claim

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

    Premisenone given

    In the typical machine learning development process, the creator of the system selects which definition of fairness to apply based on their own understanding of the specific contextual needs of a given use case.

    “In the typical machine learning development process, the creator of the system selects which defnition of fairness to apply based on own understanding of the specifc contextual needs of a given use case.”
    Quote in the paperNot a substantial claim

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

    Interpretationreasoning

    The AIaaS space would benefit greatly from the attention of the HCI community, by studying how practitioners actually attempt to use these services in the context of real-world tasks.

    “As such, we argue that the AIaaS space would beneft greatly from the attention of the HCI community, by studying how practitioners actually attempt to use these services in the context of real-world tasks.”
    Quote in the paperNot a substantial claim

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

    Premisenone given

    Little is known about who the typical AIaaS user is, what level of ML-proficiency they possess in practice, or in which markets they operate.

    “Similarly, little is known about who the typical AIaaS user is, what level of ML-profciency do they possess in practice, or in which markets they operate in.”
    Quote in the paperNot a substantial claim

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

    Interpretationreasoning

    Without proper care, the benefits of 'turn-key' AI availability and virtually limitless scalability can quickly risk potentially exacerbating the many bias problems and the broader ethical concerns around AI.

    “However, without proper care, these seeming benefts of 'turn-key' AI availability and virtually limitless scalability can quickly risk potentially exacerbating the many bias problems and the broader ethical concerns around AI that various research communities have repeatedly cautioned against.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisecitation

    Scholars have proposed a number of competing ways to quantify fairness, some of which are mathematically incompatible with each other.

    “As we discussed in §2.3 scholars have proposed a number of competing ways to quantify fairness [46,75,93,148], some of which are mathematically incompatible with each other [26,51,80].”
    Quote in the paperSubstantial claim0 of 7 sources checked
    • cites [46] source

      Not checked in the free run.

    • cites [75]

      Not checked in the free run.

    • cites [93]

      Not checked in the free run.

    • cites [148] source

      Not checked in the free run.

    • cites [26]

      Not checked in the free run.

    • cites [51]

      Not checked in the free run.

    • cites [80]

      Not checked in the free run.

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

    Premisenone given

    Despite growth in the development and dissemination of AIaaS services, there has been little research investigating how AIaaS customers actually use these services in practice, nor on practitioners' experiences and desires around AIaaS and the gaps between the capabilities of existing AI services and the needs of their users.

    “Despite growth in the development and dissemination of AIaaS services, there has been little research investigating how AIaaS customers actually use these services in practice, nor on practitioners' experiences and desires around AIaaS and the gaps between the capabilities of existing AI services and the needs of their users.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisecitation

    In the EU and UK, no equivalent of the 4/5ths rule exists in either statute or case law, and concepts of direct and indirect discrimination raise different considerations around algorithmic fairness.

    “For instance, in the EU and UK, where no equivalent of the 4/5ths rule exists in either statute or case law and where concepts of direct and indirect discrimination raise diferent considerations around algorithmic fairness [2,132]”
    Quote in the paperSubstantial claim0 of 2 sources checked
    • cites [2] source

      Not checked in the free run.

    • cites [132] source

      Not checked in the free run.

  • May combine two claims.

    Interpretationcitation

    The concern about appropriateness of an AI service is relevant for any ML model development, but it is particularly pertinent in the AIaaS context where potential harms might achieve a far larger scale.

    “While this concern is relevant for any ML model development, it is particularly pertinent in the AIaaS context where potential harms might achieve a far larger scale [29,72].”
    Quote in the paperSubstantial claim0 of 2 sources checked
    • cites [29] source

      Not checked in the free run.

    • cites [72] source

      Not checked in the free run.

  • May combine two claims.

    Premisecitation

    While there is a growing interest in auditing algorithmic systems, this field generally lacks defined standards and records could be incomplete or not relating to information needed for accountability purposes.

    “While there is a growing interest in auditing algorithmic systems [125,126,128,138,154], this feld generally lacks defned standards and records could be incomplete or not relating to information needed for accountability purposes [15,127].”
    Quote in the paperSubstantial claim0 of 7 sources checked
    • cites [125] source

      Not checked in the free run.

    • cites [126] source

      Not checked in the free run.

    • cites [128] source

      Not checked in the free run.

    • cites [138] source

      Not checked in the free run.

    • cites [154] source

      Not checked in the free run.

    • cites [15] source

      Not checked in the free run.

    • cites [127] source

      Not checked in the free run.

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

    Interpretationcitation

    Given current regulatory directions, there appears both scope and appetite for oversight bodies to set directions regarding various issues relating to AIaaS.

    “Indeed, given current regulatory directions [31,49,74], there appears both scope and appetite for oversight bodies to set directions regarding various issues relating to AIaaS.”
    Quote in the paperNot a substantial claim0 of 3 sources checked
    • cites [31]

      Not checked in the free run.

    • cites [49]

      Not checked in the free run.

    • cites [74] source

      Not checked in the free run.

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

    Interpretationnone given

    The consolidation of AI services around a few providers and the potential future widespread use of those services may mean that actually targeting specific regulation, oversight, and enforcement mechanisms at certain AI-related problems becomes easier than if many companies were developing their own AI systems in house.

    “The consolidation of AI services around a few providers and the potential future widespread use of those services-although bringing challenges, as we have argued-may also mean that actually targeting specifc regulation, oversight, and enforcement mechanisms at certain AI-related problems becomes easier than if many companies were developing their own AI systems in house.”
    Quote in the paperNot a substantial claim

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

    Interpretationnone given

    The adoption of AIaaS will likely only continue to grow.

    “The adoption of AIaaS will likely only continue to grow.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationreasoning

    Any biases and other undesirable characteristics ingrained in these services by the AIaaS provider will inevitably propagate downstream and there may be little downstream user-providers can do to address this.

    “Any biases and other undesirable characteristics ingrained in these services (by the AIaaS provider) will thus inevitably propagate downstream and there may be little downstream user-providers can do to address this;”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationreasoning

    Downstream user-providers are given no help in scrutinising or holding to account upstream providers and instead are tasked with impossible compliance tasks.

    “As it stands, the downstream user-providers are given no help in scrutinising or holding to account upstream providers and instead are tasked with impossible compliance tasks.”
    Quote in the paperSubstantial claim
78 of 78

  • The paper relies on this without support.

    AIaaS services are offered as generic 'AI building blocks' aiming to underpin a wide range of customers' applications in a variety of different contexts.

    “Moreover, with an aim to appeal to the widest possible group of customers, many of these AI services are ofered as generic 'AI building blocks', aiming to underpin a wide range of customers' applications in a variety of diferent contexts.”

  • The paper relies on this without support.

    AIaaS products are developed and sold by private for-profit companies.

    “Further, with AIaaS products developed and sold by private for-proft companies, the inner workings of these commercial algorithms are often opaque or hidden (for various reasons), making it difcult for potential users to determine if, when and why bias manifests itself within these services.”

  • The paper relies on this without support.

    The inner workings of commercial AIaaS algorithms are often opaque or hidden.

    “Further, with AIaaS products developed and sold by private for-proft companies, the inner workings of these commercial algorithms are often opaque or hidden (for various reasons), making it difcult for potential users to determine if, when and why bias manifests itself within these services.”

  • The paper relies on this without support.

    The AIaaS services and platforms examined (Azure AutoML, IBM Watson OpenScale, Baidu and Face++ beauty estimation, etc.) are representative examples of their respective service categories.

    “we use practical, tangible exemplars to expose some fundamental issues concerning bias and fairness for the diferent AIaaS categories”
4 of 4

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?

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

  • Face++ performs better for the White 60+ group

    Minor

    The claim that all classifiers perform worse for the 60+ population does not hold for every race group: Table 4 reports Face++ errors of 9.08 for White people under 60 and 6.99 for White people aged 60+, where lower error is better.

1 of 1

Cites

Flagged first, then by citations

10 of the 145 references couldn’t be matched to a known paper.

Research works printed but not linked: 11

  1. Alekh Agarwal, Alina Beygelzimer, Miroslav Dudík, John Langford, and Hanna Wallach. 2018. A reductions approach to fair classification. In International Conference on Machine Learning. PMLR, 60–69.
  2. Julia Angwin, Jeff Larson, Surya Mattu, and Lauren Kirchner. 2019. Machine bias: There’s software used across the country to predict future criminals. and it’s biased against blacks. 2016. URL https://www. propublica. org/article/machine-bias-risk-assessments-in-criminal-sentencing(2019).
  3. Tolga Bolukbasi, Kai-Wei Chang, James Y Zou, Venkatesh Saligrama, and Adam T Kalai. 2016. Man is to computer programmer as woman is to homemaker? debiasing word embeddings. Advances in neural information processing systems 29 (2016).
  4. McKinsey & Company. 2020. The state of AI in 2020. https://www.mckinsey.com/Business-Functions/McKinsey-Analytics/Our-Insights/Global-survey-The-state-of-AI-in-2020.
  5. Allan Dafoe. 2018. AI governance: a research agenda. Governance of AI Program, Future of Humanity Institute, University of Oxford: Oxford, UK (2018).
  6. 10.1016/j.ijinfomgt.2019.08.002 · DOI · Paper page
  7. Patrick Grother, Mei Ngan, and Kayee Hanaoka. 2019. Face Recognition Vendor Test (FVRT): Part 3, Demographic Effects. National Institute of Standards and Technology.
  8. Joshua A. Kroll, Joanna Huey, Solon Barocas, Edward W. Felten, Joel R. Reidenberg, David G. Robinson, and Harlan Yu. 2017. Accountable algorithms. University of Pennsylvania Law Review 165, 3 (Feb. 2017), 633–705.
  9. Gustavo Malkomes, Chip Schaff, and Roman Garnett. 2016. Bayesian optimization for automated model selection. In Workshop on Automatic Machine Learning. PMLR, 41–47.
  10. Jakob Mökander, Prathm Juneja, David S Watson, and Luciano Floridi. 2022. The US Algorithmic Accountability Act of 2022 vs. The EU Artificial Intelligence Act: what can they learn from each other?Minds and Machines (2022), 1–8.
  11. United Nations. Statistical Office. 1982. Provisional guidelines on standard international age classifications. New York: United Nations.

Cited by

Sorted by citations

Related

Authors
Lewicki, K., Lee, M. S. A., Cobbe, J., & Singh, J.
Affiliations
University of Cambridge (GB)
ORCID
Venue
Proceedings of the 2023 CHI Conference on Human Factors in Computing Systems
Publisher
ACM
Published
2023-04-19
Pages
1-17
DOI
10.1145/3544548.3581463
OpenAlex
W4319323326
Semantic Scholar
3200e16772cd05c63f08027adf842b24a4f1f64d
Cited by
30
References
145
Citations by year
2026: 7 · 2025: 12 · 2024: 6 · 2023: 5
FWCI
7.51
Type
conference paper
Language
en
Open access
Licence
Locations
  • CHI Conference on Human Factors in Computing Systems (publishedVersion, cc-by)
  • arXiv (Cornell University) (submittedVersion, cc-by)
  • Apollo (University of Cambridge) (submittedVersion, cc-by)
  • arXiv (Cornell University) (cc-by)
Funders
  • Microsoft (Microsoft Cloud Computing Research Centre)
  • Engineering and Physical Sciences Research Council (EP/P024394/1)
  • Aviva (Cambridge PhD studentship)
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
AIAA, algorithmic bias, fairness, user autonomy, societal impact of AI, end-user applications

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 2026-10-09.

Classified by jev-1.13-20260917 (gold-guideline-v0.2) on 2026-10-09.

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