Trustworthy artificial intelligence

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

Within the eight frameworks and guidelines considered, explicability can be found under different terms and to varying degrees.

“Within the eight frameworks and guidelines considered in this work, explicability can be found under different terms and to varying degrees.”

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

Justice is a key aspect of all eight frameworks and guidelines discussed in this article, though also referred to as fairness by some.

“Like non-maleficence, justice is as key aspect of all eight frameworks and guidelines discussed in this article, albeit it is also referred to as fairness by some.”

May combine two claims.

All frameworks and guidelines exhibit similar but slightly distinctive views on justice, summarizable as (1) utilization of AI to amend past inequities, (2) creation and distribution of benefits, and (3) thwarting new harms and inequities.

“As such, all frameworks and guidelines exhibit similar but slightly distinctive views on justice, which can be summarized as (1) the utilization of AI to amend past inequities like discrimination, (2) the creation of shareable and subsequent distribution of benefits through AI, and (3) thwarting the creation of new harms and inequities by AI (Floridi et al. 2018).”

• cites Floridi et al. 2018Not checked yet

  Not checked in the free run.

May combine two claims.

Such acts undermine users' trust in AI-based systems, creating a tension between output data and the non-maleficence principle.

“In the long run, such acts undermine users' trust in those AI-based systems, which is why we identify a tension between output data and the non-maleficence principle (tension: invasion of privacy).”

May combine two claims.

The concept of using DLT to create explainable AI is at the idea stage at most, and it remains unknown how DLT can support the attainment of explainable AI.

“However, looking at the recent literature, we see that the concept of using DLT to create explainable AI is at the idea stage at most and that it remains unknown how DLT can support the attainment of explainable AI.”

Classified as probably not a substantial claim.

Trust persists to play a pivotal role in any buyer-seller relationship.

“Although cost-and time-saving opportunities have triggered a widespread implementation of AI-based systems and services in electronic markets, trust persists to play a pivotal role in any buyer-seller relationship (Bauer et al. 2019;Marella et al. 2020).”

• cites Bauer et al. 2019Not checked yet

  Not checked in the free run.

• cites Marella et al. 2020 sourceNot checked yet

  Not checked in the free run.

May combine two claims.

Specific trust in people and trust in technology differ in terms of the nature of the object of dependence and on important trusting beliefs.

“Specific trust in people and trust in technology not only differ in terms of the nature of the object of dependence but also on important trusting beliefs.”

• cites Mayer et al. 1995 sourceNot checked yet

  Not checked in the free run.

• cites McKnight et al. 2011 sourceNot checked yet

  Not checked in the free run.

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

TAI is of increasing relevance to electronic markets and its research community.

“Consequently, TAI is of increasing relevance to electronic markets and its research community.”

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

AI is not the 'magic bullet' some would like to believe it is and that AI, just like any other technology, will not only bring forth many benefits but will also be accompanied with a variety of novel ethical, legal, and social challenges.

“However, it is becoming increasingly evident that AI is not the 'magic bullet' some would like to believe it is and that AI, just like any other technology, will not only bring forth many benefits but will also be accompanied with a variety of novel ethical, legal, and social challenges (Floridi 2019;Floridi et al. 2018).”

• cites Floridi 2019 sourceNot checked yet

  Not checked in the free run.

• cites Floridi et al. 2018Not checked yet

  Not checked in the free run.

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

TAI is a highly interdisciplinary and dynamic field of research, with knowledge on technical and non-technical means to realize TAI being scattered across research disciplines, thus making it challenging to grasp the status quo on its realization.

“Overall, TAI is a highly interdisciplinary and dynamic field of research, with knowledge on technical and non-technical means to realize TAI being scattered across research disciplines, thus making it challenging to grasp the status quo on its realization.”

• Also stated in: § Toward a definition of trustworthy artificial intelligence · plain

  “Yet, TAI is a highly interdisciplinary and dynamic field of research, comprising multifarious research discussions and streams that are scattered across disciplines, including psychology, sociology, economics, management, computer science, and IS.”

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

Various perspectives on trust exist in literature, comprising different dimensions and (partially opposing) interpretations.

“Nevertheless, various perspectives on trust exist in literature, comprising different dimensions and (partially opposing) interpretations (McKnight et al. 2002).”

• cites McKnight et al. 2002 sourceNot checked yet

  Not checked in the free run.

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

The Ethics Guidelines for Trustworthy AI published by the independent High-Level Expert Group on Artificial Intelligence of the European Commission in early 2019 have quickly gained traction in research and practice and have laid the foundation for the adoption of the term trustworthy AI in other guidelines and frameworks like the OECD principles on AI or the White House AI principles.

“These guidelines have quickly gained traction in research and practice and have laid the foundation for the adoption of the term trustworthy AI in other guidelines and frameworks like the OECD principles on AI (OECD 2019) or the White House AI principles (Vought 2020).”

• cites OECD 2019Not checked yet

  Not checked in the free run.

• cites Vought 2020Not checked yet

  Not checked in the free run.

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

The conceptualization of trust in AI and what makes AI trustworthy remains inconclusive and highly discussed in research and practice.

“Considering that 'trust' in general is a complex phenomenon that has sparked many scholarly debates in recent decades, it is not surprising that the conceptualization of trust in AI and what makes AI trustworthy-as of today-remains inconclusive and highly discussed in research and practice.”

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

Trustworthy AI (TAI) is based on the idea that trust builds the foundation of societies, economies, and sustainable development, and that individuals, organizations, and societies will only ever be able to realize the full potential of AI if trust can be established in its development, deployment, and use.

“Trustworthy AI (TAI) bases on the idea that trust builds the foundation of societies, economies, and sustainable development, and that individuals, organizations, and societies will therefore only ever be able to realize the full potential of AI, if trust can be established in its development, deployment, and use.”

• Also stated in: § Introduction · plain

  “the concept of trustworthy AI (TAI) promotes the idea that individuals, organizations, and societies will only ever be able to achieve the full potential of AI if trust can be established in its development, deployment, and use (Independent High-Level Expert Group on Artificial Intelligence 2019).”

• Also stated in: § Toward a definition of trustworthy artificial intelligence · plain

  “In its essence, TAI is based on the idea that trust builds the foundation of societies, economies, and sustainable development, and that therefore the global society will only ever be able to realize the full potential of AI if trust can be established in it (Independent High-Level Expert Group on Artificial Intelligence 2019).”

• cites Independent High-Level Expert Group on Artificial Intelligence 2019Not checked yet

  Not checked in the free run.

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

Extant research has commonly agreed that individuals express expectations about a person's competence, benevolence, and integrity.

“Extant research has commonly agreed that individuals express expectations about a person's competence (i.e., its ability to do what the individual needs), benevolence (i.e., its care and motivation to act in the individual's interests), and integrity (i.e., its honesty and promise-keeping) (McKnight et al. 2002).”

• cites McKnight et al. 2002 sourceNot checked yet

  Not checked in the free run.

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

There is no commonly accepted definition of trust but rather a need for contextualized trust conceptualizations.

“As a result of the plurality of perspectives on this concept, there is no commonly accepted definition of trust (Lansing and Sunyaev 2016;Söllner et al. 2016) but rather a need for contextualized trust conceptualizations (Jarvenpaa et al. 2004).”

• cites Lansing and Sunyaev 2016 sourceNot checked yet

  Not checked in the free run.

• cites Söllner et al. 2016 sourceNot checked yet

  Not checked in the free run.

• cites Jarvenpaa et al. 2004 sourceNot checked yet

  Not checked in the free run.

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

Trust plays a particularly important role in almost any IS-enabled situation in which either uncertainty prevails or undesirable outcomes are possible.

“Trust plays a particularly important role in almost any ISenabled situation in which either uncertainty prevails or undesirable outcomes are possible (McKnight et al. 2011).”

• cites McKnight et al. 2011 sourceNot checked yet

  Not checked in the free run.

May combine two claims.

Interpersonal trusting beliefs reflect judgments that the other party has suitable attributes and motives for performing as expected in a risky situation, whereas technology-related trust necessarily reflects beliefs about a technology's characteristics rather than its motives.

“Interpersonal trusting beliefs reflect judgments that the other party has suitable attributes and motives for performing as expected in a risky situation (Mayer et al. 1995), whereas technology-related trust necessarily reflects beliefs about a technology's characteristics rather than its motives (McKnight et al. 2011).”

• cites Mayer et al. 1995 sourceNot checked yet

  Not checked in the free run.

• cites McKnight et al. 2011 sourceNot checked yet

  Not checked in the free run.

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

Both types of trust are highly relevant in the context of AI.

“Both types of trust are highly relevant in the context of AI.”

May combine two claims.

AI-based systems' autonomous and intelligence-based capabilities allow them to have a great degree of self-governance, which enables them to respond to situations that were not pre-programmed or explicitly anticipated during their development, and to make independent decisions and action selection with little or no control by their users.

“AI-based systems' autonomous and intelligence-based capabilities allow them to have a great degree of self-governance, which enables them to respond to situations that were not pre-programmed or explicitly anticipated during their development, and to make independent decisions and action selection with little or no control by their users (Parasuraman et al. 2000).”

• cites Parasuraman et al. 2000 sourceNot checked yet

  Not checked in the free run.

May combine two claims.

Training data that are already available are routinely biased toward certain groups of people, leading to implicit discrimination of underrepresented groups.

“Next to the limited availability of high-quality training data, training data that are already available are routinely biased toward certain groups of people, which in the past has led to the implicit discrimination of those groups of people that are underrepresented in the training data (Bellamy et al. 2019).”

• cites Bellamy et al. 2019Not checked yet

  Not checked in the free run.

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

Input data plays a dual role: it serves as the main resource for training AI models, and it is transferred into output data via a trained AI model once deployed.

“Input data plays an important dual role in AI-based systems.On the one hand, it serves as the main resource for the training of AI models.foot_0On the other hand, input data is transferred into output data via a trained AI model, once an AI has been deployed.”

May combine two claims.

Two research streams emerged that either adapted the concept of interpersonal trust to conceptualize trust in human-like IT artifacts, or revised the concept of trust in technology to account for automation technology and autonomous systems.

“To understand trust relationships in such scenarios, two research streams emerged that either adapted the concept of interpersonal trust to conceptualize trust in human-like IT artifacts (Lankton et al. 2015), or revised the concept of trust in technology to account for automation technology and autonomous systems (J.D. Lee and See 2004).”

• cites Lankton et al. 2015 sourceNot checked yet

  Not checked in the free run.

• cites J.D. Lee and See 2004 sourceNot checked yet

  Not checked in the free run.

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

The non-maleficence principle is highly important for electronic markets due to the exchange and analysis of highly sensitive consumer and intellectual property data.

“the non-maleficence principle is highly important for electronic markets due to the exchange and analysis of highly sensitive consumer and intellectual property data.”

May combine two claims.

Non-maleficence especially concerns the protection of people's privacy and security, as well as their safety.

“Non-maleficence especially concerns the protection of people's privacy (expressed by the Asilomar AI Principles, Montreal Declaration, UK AI Code, AI4People, EU TAI Guidelines, Chinese AI Principles) and security (expressed by the AI4People, EU TAI Guidelines, OECD Principles on AI, White House AI Principles), as well as their safety (expressed by the Asilomar AI Principles, UK AI Code, AI4People,”

May combine two claims.

Some TAI frameworks mainly focus on the promotion of human autonomy, agency, and oversight, while others also consider the restriction of AI-based systems' autonomy, where necessary.

“While some mainly focus on the promotion of human autonomy, agency, and oversight (e.g., EU TAI Guidelines), others also consider the restriction of AI-based systems' autonomy, where necessary (e.g., the Montreal Declaration) (Floridi and Cowls 2019).”

• cites Floridi and Cowls 2019Not checked yet

  Not checked in the free run.

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

These trust in automation beliefs take a more technical focus, yet they still relate to prevalent beliefs of trust in human-like IT artifacts; purpose corresponds to helpfulness and benevolence and reflects the perception that an automated system has a positive orientation toward the user.

“These trust in automation beliefs therefore take a more technical focus, yet they still relate to prevalent beliefs of trust in human-like IT artifacts.Purpose, for example, corresponds to helpfulness and benevolence and reflects the perception that an automated system has a positive orientation toward the user.”

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

Unintended or malicious behavior of an AI-based system will most likely exert a negative impact on users' trust in the AI.

“We discern this as a tension between input data and the non-maleficence TAI principle (tension: malicious training data), since unintended or malicious behavior of an AI-based system will most likely exert a negative impact on users' trust in the AI.”

May combine two claims.

AI-based systems' actions necessarily become more indeterminate across time and are thus more challenging to predict, making trust interactions between humans and AI-based systems more complex and difficult to understand than trust interactions between humans and non-AI technologies.

“AI-based systems' actions necessarily become more indeterminate across time and are thus more challenging to predict (Hancock 2017), making trust interactions between humans and AI-based systems more complex and difficult to understand than trust interactions between humans and non-AI technologies.”

• cites Hancock 2017 sourceNot checked yet

  Not checked in the free run.

May combine two claims.

A widely used AI in US hospitals is negatively biased toward Black patients because it uses health costs as a proxy for health needs and less money is spent on Black patients.

“Obermeyer et al. (2019), for example, found that a widely used AI in US hospitals is negatively biased toward Black patients, since it uses health costs spent on patients as a proxy for health needs and since on average less money is spent on Black patients in the US.”

• cites Obermeyer et al. 2019 sourceNot checked yet

  Not checked in the free run.

May combine two claims.

AI models themselves constitute an important form of data, and there are tensions between the model and the five TAI principles.

“In line with our guiding notion that data is the single, most important resource for contemporary AI-based systems, we argue that AI models themselves constitute an important form of data and identify several tensions between the model and the five TAI principles.”

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

AI can be employed for predictive policing and benefit society by allowing better allocation of police staff and reducing crime rates.

“Extant research shows that AI can be employed for purposes of predictive policing (i.e., using mathematical models to forecast what crimes will happen when and where) and therefore benefit society by allowing for a better allocation of police staff and reducing crime rates (Courtland 2018).”

• cites Courtland 2018 sourceNot checked yet

  Not checked in the free run.

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

Non-maleficence represents a distinct principle that represents a key aspect of all considered frameworks and guidelines.

“Although similar to beneficence, which emphasizes the creation of AI that actively acts towards the wellbeing of humanity, non-maleficence represents a distinct principle that represents a key aspect of all considered frameworks and guidelines.”

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

An individual will tend to trust the AI if its algorithms can be understood and seem capable of achieving the individual's goals in the current situation.

“Thus, an individual will tend to trust the AI if its algorithms can be understood and seem capable of achieving the individual's goals in the current situation.”

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

Explicability relates to the trusting beliefs competence, functionality, and performance in that explainable and interpretable AI proves it has the capability, functionality, or features to do what needs to be done.

“Explicability relates also to the trusting beliefs competence, functionality, and performance in the sense that explainable and interpretable AI proves that it has the capability, functionality, or features to do what needs to be done.”

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

The autonomy principle is not directly related to extant trusting beliefs but reflects a means to mitigate integrity and reliability risks by balancing between human- and machine-led decision-making.

“The autonomy principle is not directly related to extant trusting beliefs but reflects a means to mitigate integrity and reliability risks by balancing between human-and machineled decision-making.”

May combine two claims.

In the IS domain, explicability of AI is of major importance because it allows organizations to meet compliance requirements and is a key driver for acceptance of AI by managers, workforce, and consumers.

“In the IS domain, explicability of AI is of major importance since it will not only allow organizations to meet compliance requirements when employing AI (e.g., by means of enabling independent third-party audits) but will also be a key driver for acceptance of AI by managers, the general workforce, and consumers (Hagras 2018;Rai 2020).”

• cites Hagras 2018 sourceNot checked yet

  Not checked in the free run.

• cites Rai 2020 sourceNot checked yet

  Not checked in the free run.

May combine two claims.

Although beneficence is found in all of the frameworks and guidelines discussed here, it is taken into account to varying degrees.

“Although beneficence is found in all of the frameworks and guidelines discussed here, it is taken into account to varying degrees.”

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

These tensions form the backbone of DaRe4TAI and should be understood as illustrations of challenges to overcome to achieve TAI, not as contributing to mistrust.

“These tensions, which form the backbone of DaRe4TAI, should thereby not to be understood from a negative point of view, as in contributing to mistrust in AI, but instead as illustrations of challenges that need to be overcome in certain scenarios to achieve TAI from a user (i.e., consumers, organizations, society) perspective.”

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

Non-maleficence relates to the trusting beliefs integrity, reliability, and process because it requires AI-based systems to act honestly and consistently, and to sincerely adhere to ethical and other pre-defined principles.

“Nonmaleficence relates to the trusting beliefs integrity, reliability, and process because it requires AI-based systems to act honestly and consistently, and to sincerely adhere to ethical and other pre-defined principles.”

May combine two claims.

Such insights could be used to identify individuals, representing a privacy infringement that could undermine trust.

“In extreme cases, such insights could be used to identify individuals who contributed their data, which in turn represents a privacy infringement that could undermine those very individuals' trust in AI-based systems.”

May combine two claims.

Autonomy aligns with openness, a sub-dimension of the process belief of automation technologies, and openness increases trust into another party.

“addition, autonomy aligns with openness, a sub dimension of the process belief of automation technologies (J.D. Lee and See 2004), that refers to the willingness to give and receive ideas, which will increase trust into another party (Mishra 1992;Schindler and Thomas 1993).”

• cites J.D. Lee and See 2004 sourceNot checked yet

  Not checked in the free run.

• cites Mishra 1992Not checked yet

  Not checked in the free run.

• cites Schindler and Thomas 1993 sourceNot checked yet

  Not checked in the free run.

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

Justice aligns with the trusting beliefs integrity, reliability, and process, ensuring ethical principles are fulfilled by an AI-based system.

“Similar to non-maleficence, justice aligns with the trusting beliefs integrity, reliability, and process, ensuring that ethical principles are fulfilled by an AI-based system.”

May combine two claims.

Attempts to protect competitive advantages can contribute to promising AI models not being shared and AI being perceived as not beneficent enough.

“Attempts to protect competitive advantages can thereby contribute to the fact that particularly promising AI models are not shared and that AI as a specific class of technology are perceived as not beneficent (enough) by the society (i.e., the whole of AI-based systems not acting in societies best interest).”

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

Given that extant TAI frameworks and guidelines provide slightly different understandings of autonomy, it lacks a precise definition.

“Given that extant TAI frameworks and guidelines provide slightly different understandings of this principle, it lacks a precise definition.”

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

AI-based systems are based on an abstract AI co-creation process comprising three stages: input, modeling, and output.

“They are thus based on an abstract AI co-creation process, comprising three important stages: input, modeling, and output.”

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

Several frameworks and guidelines that promote (ethical) principles for TAI have been developed and published by researchers, industry, and policymakers in the recent past.

“Several frameworks and guidelines that promote (ethical) principles for TAI have been developed and published by researchers, industry, and policymakers in the recent past.”

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

The question of how to effectively design token economies remains a focal theme of contemporary DLT research.

“However, despite first technical solutions being developed by researchers from the IS, computer science, and related disciplines (Ozercan et al. 2018;Özyilmaz et al. 2018;Xiong and Xiong 2019;Zhao et al. 2019), the question of how to effectively design token economies (e.g., to democratize data access or to encourage the generation of more diverse data sets) remains a focal theme of contemporary D”

• cites Ozercan et al. 2018 sourceNot checked yet

  Not checked in the free run.

• cites Özyilmaz et al. 2018Not checked yet

  Not checked in the free run.

• cites Xiong and Xiong 2019 sourceNot checked yet

  Not checked in the free run.

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

The last avenue concerns achieving accountability and explainability of AI through DLT.

“The last avenue of DLT-related research on the realization of TAI that we discuss in this article concerns achieving accountability and explainability of AI through DLT.”

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

DLT-based data markets, DLT-based federated learning, and DLT-based transparency, accountability, and explainability are fruitful avenues for further research to address these tensions and realize TAI.

“Especially DLT-based data markets, DLT-based federated learning, and DLT-based transparency, accountability, and explainability are fruitful avenues for further research to better address these tensions and eventually realize TAI.”

May combine two claims.

DLT-based data markets provide economic incentives that could stimulate democratization of access to high-quality AI training data and encourage generation of more diverse data sets.

“For example, DLT-based data markets provide the ability to create economic incentives, which could not only stimulate the democratization of access to extant, high-quality AI training data (i.e., addressing the training data availability tension) but as well encourage greater participation by the general public to drive the generation of new, more diverse data sets (i.e., addressing the training”

May combine two claims.

The privacy invasion on the output stage occurs because the AI's purpose is malicious, unlike input/model stages where purpose might be benevolent.

“However, in contrast to the invasion of privacy tension on the input stage and the model stage, where the AI-based system's purpose might have been benevolent but the collection, processing, and analysis of the users' data without their consent stipulated a privacy infringement, the privacy invasion on the output stage occurs simply due to the fact that the AI's purpose is malicious and an”

May combine two claims.

The modeling of assets as tokens is a topic of ongoing research, and we lack knowledge on how AI-related assets can be modeled and represented adequately as tokens.

“Furthermore, the modeling of assets as tokens is a topic of ongoing research (Kim and Chung 2018;Laskowski et al. 2019) and, thus far, we lack knowledge on what and how AI-related assets (e.g., training data, model data, algorithms) can be modeled and represented adequately as tokens.”

• cites Kim and Chung 2018 sourceNot checked yet

  Not checked in the free run.

• cites Laskowski et al. 2019Not checked yet

  Not checked in the free run.

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

TAI is a promising research topic for IS research.

“In this article, we introduced the concept of TAI as a promising research topic for IS research, delineated its background, positioned it in related trust conceptualizations, and contextualized the five TAI principles beneficence, non-maleficence, autonomy, justice, and explicability to the IS context.”

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

We still lack substantive knowledge on how to successfully design a token economy for DLT-based federated learning.

“However, also similar to DLT-based data markets, we still lack substantive knowledge on how to successfully design such a token economy for DLT-based federated learning.”

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

The development and deployment of AI are highly dynamic, with training data and algorithms rapidly changing and constantly evolving.

“However, the development and deployment of AI are highly dynamic, with training data and algorithms (and thus model uncertainty) rapidly changing and constantly evolving.”

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

It is not clear whether promises of increased privacy from DLT-based federated learning may strengthen data providers' trust and willingness to contribute data.

“It is, for example, not clear whether the promises of increased privacy due to the application of (DLT-based) federated learning may actually strengthen data providers' trust in an AI-based system's ability to adequately protect their data and ultimately their willingness to contribute their data for purposes of training AI models.”

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

DLT can provide a ledger for incentivizing participation in federated learning networks.

“Moreover, and similar to the previously described token economy for AI training data, DLT can provide a ledger for incentivizing participation in federated learning networks.”

May combine two claims.

In federated learning, no input training data is directly shared; partial AI models are trained by nodes, with provenance and integrity preserved using a distributed ledger.

“In such a federated learning scenario, no input training data is directly shared.Instead, partial AI models are being trained by nodes participating in the federated learning network, while training data provenance and the integrity of the partial AI models are preserved using a distributed ledger (Pandl et al. 2020).”

• cites Pandl et al. 2020 sourceNot checked yet

  Not checked in the free run.

May combine two claims.

Approaches to quantify such uncertainties are often still in their infancy and not broadly available for all use cases.

“Although there exist first approaches in research and practice to quantify such uncertainties, these approaches are often still in their infancy and are not broadly available for all use cases (Begoli et al. 2019).”

• cites Begoli et al. 2019Not checked yet

  Not checked in the free run.

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

There remain several issues that warrant further research before DLT-based federated learning can be deployed in real-world use cases.

“However, there remain several issues that warrant further research before we will be able to deploy DLTbased federated learning in real-world use cases.”

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

The opacity of most current AI models is a popular topic of contemporary AI research.

“Lastly, the opacity of most current AI models is one of the most popular topics of contemporary AI research (Rudin 2019).”

• cites Rudin 2019 sourceNot checked yet

  Not checked in the free run.

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

Effective auditing of AI warrants creating an independent, tamper-resistant audit trail and continuous updating and assessment using continuous auditing procedures.

“Effective auditing of AI, therefore, does not only warrant creating an independent, tamper-resistant audit trail, but also the continuous updating and assessment of this audit trail using continuous auditing procedures (Lins et al. 2019).”

• cites Lins et al. 2019 sourceNot checked yet

  Not checked in the free run.

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

The potential negative consequences of a token economy that could conflict with TAI warrant further research.

“The potential negative consequences of a token economy that could conflict with the realization of TAI, therefore, warrant further research.”

May combine two claims.

This opacity impedes interpretability of output data and obstructs establishing accountability.

“Not only does this impede the interpretability of output data but also obstruct establishing accountability.”

May combine two claims.

Research has begun exploring DLT for auditing purposes in organizational contexts, and first results indicate feasibility of DLT for auditing AI.

“Owing to DLTs characteristics (e.g., decentralization, high tamper resistance), research has recently begun exploring the application of DLT for auditing purposes in organizational contexts (Hofman et al. 2019), while first research results also indicate the feasibility of DLT for the auditing of AI (Dillenberger et al. 2019).”

• cites Hofman et al. 2019 sourceNot checked yet

  Not checked in the free run.

• cites Dillenberger et al. 2019 sourceNot checked yet

  Not checked in the free run.

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

Several non-technical questions require further research.

“Despite technical questions, several non-technical questions require further research.”

May combine two claims.

DLT enables coordination of economic activity through secure, transparent, decentralized electronic markets and is best known as blockchain.

“It enables the coordination of economic activity through the creation of secure, transparent, and decentralized electronic markets (Berg et al. 2019;Kollmann et al. 2019;Subramanian 2017), and is probably best known under the name of blockchain, which is a specific type of DLT (Kannengiesser et al. 2020;Sunyaev 2020).”

• cites Berg et al. 2019 sourceNot checked yet

  Not checked in the free run.

• cites Kollmann et al. 2019Not checked yet

  Not checked in the free run.

• cites Subramanian 2017 sourceNot checked yet

  Not checked in the free run.

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

The design of effective token economies for democratizing access to AI models requires further research.

“Yet, analog to the question of how to design token economies capable of effectively democratizing access to input training data, the design of effective token economies for democratizing access to AI models requires further research.”

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

There is a vast space of TAI research opportunities for the IS and other research communities that is not limited to explainability.

“In doing so, we highlight a vast space of TAI research opportunities for the IS and other research communities that is not limited to the recent AI hype topic of explainability.”

May combine two claims.

AI-based systems have been shown to infer individuals' political views or sexual orientation from Facebook likes or predict mental state from posts.

“Toward this end, AI-based systems have, for example, been shown to be able to infer individuals' political views or sexual orientation based on the Facebook pages that they liked (Gibney 2018) or predicted the mental state of Facebook users based on an analysis of their posts (Goggin 2019).”

• cites Gibney 2018Not checked yet

  Not checked in the free run.

• cites Goggin 2019Not checked yet

  Not checked in the free run.

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

DLT-based data markets could serve as a means to address several tensions at the input and model stages.

“They are of particular interest for the electronic markets community and could serve as a means to address several tensions at the input and model stages.”

Classified as probably not a substantial claim.

A central facet of establishing AI accountability concerns our ability to independently audit AI, especially data provenance and model uncertainty.

“A central facet of establishing AI accountability concerns our ability to independently audit AI (i.e., AI's auditability), especially in terms of data provenance (i.e., addressing the malicious training data tension) and the degree of uncertainty with which AI models make their predictions (i.e., addressing the model uncertainty tension).”

Classified as probably not a substantial claim.

DaRe4TAI provides guidance to study technical and non-technical means in support of TAI.

“Further, we drew on a data-driven perspective toward AI to develop the research framework DaRe4TAI that provides guidance to those enticed to study technical and non-technical means in support of TAI, and demonstrated its feasibility on the example of fruitful avenues for future research on the DLT-based realization of TAI.”

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

The potential of a token economy for TAI extends to trading and licensing of other AI-related assets such as models or algorithms.

“Lastly, the potential of a token economy for TAI extends beyond the creation of data markets for input training data to the trading and licensing of other AI-related assets such as models or algorithms (i.e., addressing the model availability tension) (Sarpatwar et al. 2019).”

• cites Sarpatwar et al. 2019 sourceNot checked yet

  Not checked in the free run.

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

DLT ought to serve as a tamper-resistant trail for tracking the flow of data within AI-based systems, which may then be further analyzed to create explainable AI models.

“DLT is thereby ought to serve as a tamper resistant trail for tracking the flow of data within AI-based systems, which may then be further analyzed to create explainable AI models (Dinh and Thai 2018).”

• cites Dinh and Thai 2018 sourceNot checked yet

  Not checked in the free run.

May combine two claims.

Most research prototypes employ DLT-based federated learning to train relatively simple AI models, and extant research indicates a performance overhead of 5% to 15%.

“Most research prototypes, for example, employ DLT-based federated learning to train relatively simple AI models (Pandl et al. 2020;Preuveneers et al. 2018), while extant research indicates that DLT-based federated learning induces a performance overhead of 5% to 15% (Preuveneers et al. 2018).”

• cites Pandl et al. 2020 sourceNot checked yet

  Not checked in the free run.

• cites Preuveneers et al. 2018 sourceNot checked yet

  Not checked in the free run.

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

TAI provides several promising avenues of future research for the field of electronic markets, including and beyond its DLT-based realization.

“Especially for the field of electronic markets, TAI provides several promising avenues of future research, including and beyond its DLT-based realization.”

Classified as probably not a substantial claim.

The feasibility of DaRe4TAI was demonstrated on the example of fruitful avenues for future research on the DLT-based realization of TAI.

“Further, we drew on a data-driven perspective toward AI to develop the research framework DaRe4TAI that provides guidance to those enticed to study technical and non-technical means in support of TAI, and demonstrated its feasibility on the example of fruitful avenues for future research on the DLT-based realization of TAI.”

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

DaRe4TAI provides a good starting ground for exploring further tensions and revealing additional avenues for future research on technical and non-technical means in support of TAI.

“Nevertheless, we are convinced that DaRe4TAI provides a good starting ground for exploring further tensions and, thus, revealing additional avenues for future research on technical and non-technical means in support of TAI.”

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

The tensions between data at the different stages of the AI co-creation process and the five TAI principles outlined represent only a subset of tensions.

“The tensions between data at the different stages of the AI co-creation process and the five TAI principles that we outlined here represent only a subset of tensions.”