Artificial intelligence (AI) in its current form relies heavily on large quantities of data (Yavuz, 2019), and data-driven Deep Neural Networks (DNNs) have prompted fast-paced development of AI (Greene, 2020). Currently, the research community is under great strain to keep up with the potential ethical concerns which arise as a result of this (Naughton, 2019). Within the AI community such ethical concerns can require quite some disentanglement (Allen et al., 2006), and it is not until recently that AI-based research groups have begun to provide public manifestos concerning the ethics of AI, e.g., Google's DeepMind, and the Partnership AI.¹

The Ethics of AI (Boddington, 2017) is now an essential topic for researchers, both internal and external, to core-machine learning and differs from Machine Ethics (Baum et al., 2018). The latter refers to giving conscious ethical based decision-making power to machines. The Ethics of AI, although somewhat informing Machine Ethics, refers more broadly to decisions made by researchers and covers issues of diversity and representation, e.g., to avoid discrimination (Zliobaite, 2015) or inherent latent biases (van Otterlo, 2018). Herein, our discussion focuses on topics relating to the Ethics of AI unless otherwise stated.

There has been recent research which shows promise for improved data learning from smaller quantities (“merely a few minutes”) of data (Chen et al., 2018). However, machine learning algorithms developed for AI commonly require substantial quantities of data (Schneider, 2020). In this regard, Big Data ethics for AI algorithms are an expanding discussion point (Berendt et al., 2015; Mittelstadt and Floridi, 2016). Crowdsourcing (i.e., data gathered from large amounts of paid or unpaid individuals via the internet), is one approach to collect such quantities of data. However, ethical concerns including worker exploitation (Schlagwein et al., 2019), may have implications on the validity of the data. Additionally researchers utilize in-the-wild internet sources, e.g., YouTube (Abu-El-Haija et al., 2016) or Twitter (Beach, 2019), and apply unsupervised labeling methods (Jan, 2020). However, in Parikh et al. (2019), the authors describe how approaches for automated collection and labeling can result in the propagation of historical and social biases (Osoba and Welser IV, 2017). In the health domain, such bias could have serious consequences, leading to misdiagnosis or incorrect treatment plans (Mehrabi et al., 2019).

One method to avoid bias in AI is through the acquisition of diverse data sources (Demchenko et al., 2013). With Veracity (i.e., habitual truthfulness) being one of the 5 Vs (e.g., Velocity, Volume, Value, Variety and Veracity) for defining truly Big Data (Khan et al., 2019). However, big data is commonly, stored in centralized infrastructures which limit transparency, and democratic, decentralized (i.e., peer-to-peer blockchain-based) approaches are becoming prevalent (Luo et al., 2019).

Centralized data storage can be efficient and beneficial to the “central” body to which the infrastructure belongs. However, it is precisely this factor amongst others (i.e., proprietary modeling of underrepresented data) that are problematic (Ferrer et al., 2019).

Furthermore, centralized platforms limit the access and knowledge that data providers receive. The General Data Protection Regulation (GDPR) was established within the European Union (The-European-Commission, 2019) to partly tackle this. GDPR is a set of regulations of which the core goal is to protect the data of individuals that are utilized by third parties. In its current form, GDPR promotes a centralized approach, supporting what are known as commercial governance platforms. These platforms control restrictions to employees based on a data providers request but primarily function as a centralized repository. In essence, GDPR meant that companies needed to re-ask for data-consent more transparently. However, the “terms of agreement” certificate remains the basis, and 90 % of users are known to ignore its detail (Deloitte, 2016).

As a counter approach to the centralized storage of data, for some time researchers have proposed the need for a decentralized (cf. Figure 1) networking in which individual data is more easily protected (i.e., there is no “single point” of failure). In this infrastructure, individuals have more agency concerning the use of their data (Kahani and Beadle, 1997). Primarily, individuals choose to access parts of a network rather than its entirety. On a large scale, this paradigm would remove the known biases of centralized networks, as targeted collection, for example, would be less accessible by companies and sources of the data more complex to identify. In this way, various encryption algorithms, including homomorphic encryption (a method which allows for data processing while encrypted), or data masking, are being integrated within decentralized networks, allowing for identity preservation (Setia et al., 2019). Federated Learning (FL) (Hu et al., 2019), is one approach which can be applied to decentralized networks to improve privacy (Marnau, 2019). In FL, weights are passed from the host device and updated locally, instead of raw data leaving a device (Yang et al., 2019).

With these topics in mind, in this contribution, we aim to outline core ethical considerations, which relate to data and the ethics of AI. Our focus remains on the ethics of data representation and data infrastructure, particularly selection-bias and decentralization. We chose these topics due to their common pairing in the literature. A regular talking-point in machine learning is selection-bias and a networking infrastructure which may help to more transparently observe this is decentralization (Swan, 2015; Montes and Goertzel, 2019).

Our contribution is structured as follows; firstly we shortly define key terminology used throughout the manuscript in section 2, followed by a brief background and overview of the core themes as they pertain to AI in section 3. We then introduce our ethical data considerations in section 4 providing specific definitions and general ethical concerns. Following this in section 5, we connect these ethical considerations more closely with data representation and infrastructure, and in turn, outline technical approaches which help reduce the aforementioned ethical concerns. Finally, we offer concluding remarks in section 7.

There are a variety of core terms which are used throughout this manuscript which may have a dual meaning in the machine learning community. For this reason, we first define here three core terms, ethics, bias, and decentralization used within our discussion.

As mentioned previously, we focus on the Ethics of AI rather than Machine Ethics. However, further to this, we use the term ethics based on guidelines within applied ethics, particularly in relation to machine understanding. In Döring et al. (2011), the principles of beneficence, non-maleficence, autonomy, and justice are set out as being fundamental considerations for those working in AI. Although this is particular to emotionally aware systems, we consider that such principles are relevant across AI research. Of particular relevance to this contribution, is autonomy, i.e., a duty for systems to avoid interference, and respect an individual's capacity for decision-making. This principle impacts upon both data representation and infrastructure choices (e.g., centralized or decentralized).

We consistently refer to the term bias throughout our contribution. First introduced to machine learning by Mitchell (1980), we typically discuss statistical biases, unless otherwise stated, which may include absolute or relative biases. To be more specific, we focus closely on data in this contribution, and therefore dominantly refer to selection-bias. Selection-bias stems in part from prejudice-based biases (Stark, 2015). However, selection-bias falls within statistical biases as it is a consequence of conscious (hence prejudice) or unconscious data selection. Selection-bias is particularly relevant to AI given that real randomization (or diverse representation) of data is not always possible.

As a critical aspect of our contribution, relating to the mitigation of bias, through a more ethical approach to data infrustructure, we consistently refer to decentralized AI. A broad definition of decentralization is the distribution of power moving away from central authorities. In the context of AI, when discussing decentralization, we refer to decentralized architectures which allow for this type of distribution, in regards to data sourcing, management and analysis. We do touch on literature relating to blockchain, which is a well-known decentralized approach. However, the term is utilized here more generally and is not exclusive to the blockchain.

Funding and global research efforts in the field of AI have increased in the last decade, particularly in the areas of health, transportation, and communication (Mou, 2019). Along with this increase has come a rise in ethical demands related to Big Data (Herschel and Miori, 2017). Although true Big Data is said to need Veracity, the reality of this is sometimes different, with large-scale data often showing particular biases toward clustered demographics (Price and Ball, 2014). As a result, terms, such as Machine Learning Fairness—promoted initially by Google Inc.²—is now regularly referred to in an endeavor to build trust and show ethical sensitivity (Mehrabi et al., 2019). In this regard, IBM released their AI Explainability 360 Toolkit³ in which the overarching goal appears to be improving trust in AI, through more deeply researching machine learning biases, as it pertains to the research areas of fairness, robustness and explainability.

Three common forms of bias are discussed concerning AI, i.e., interaction-bias, latent-bias, and selection-bias. Selection-bias occurs when the data used within a paradigm is selected with bias, leading to misrepresentation rather than generalization. In particular, researchers are repeatedly finding bias in regards to gender (Gao and Ai, 2009). Wang et al. (2019a) found for example that models tend to have a bias toward a particular gender even when a dataset is balanced—which could point to lower level architecture-based biases (Koene, 2017). Selection-bias is essential to combat when referring to models developed for human interaction. Based on data decision making, a bias can propagate through system architectures, leading to lower accuracy on a generalized population. Lack of generalization is particularly problematic for domains, such as health, where this may result in a breach of patient safety (Challen et al., 2019).

Furthermore, the evaluation of fairness in machine learning is another prominent topic, highlighted as a machine learning consideration in Hutchinson and Mitchell (2019). Additionally, researchers propose fairness metrics for evaluating the bias which is inherent to a model (Friedler et al., 2019), including the Disparate Impact or Demographic Parity Constraint (DPC). DPC groups underprivileged classes and compares them to privileged classes as a single group. Similarly, there are novel architectures which mitigate bias through prioritization of minority samples, and the authors of this approach suggest that there is an improvement in generalized fairness (Lohia et al., 2019).

A core contributing factor to bias in AI is the management of data. Current AI networking is based on centralized infrastructure (cf. Figure 1), where individuals present a unified data source to a central server. This centralization approach not only limits privacy but also creates a homogeneous representation, which is less characteristic of the individual interacting (Sueur et al., 2012).

Decentralization in AI was initially coined as a term to describe “autonomous agents in a multi-agents world” (Miiller, 1990), and researchers have proposed decentralization for large AI architectures e.g., integrating machine learning with a Peer-to-peer style blockchain approach Zheng et al., 2018] to improve fairness and bias (Barclay et al., 2018). In this architecture, collaborative incentives are offered to the network users and approaches allow for improved identity-representation, as well as more control in regards to data-usage, resulting in more freedom and higher privacy. Furthermore, a decentralized network may inherently be more ethical as more individuals are interacting with and refining the network with agency (Montes and Goertzel, 2019).

For individuals interfacing with AI, privacy is a concern (Montes and Goertzel, 2019). Improving privacy is a core advantage of decentralized data approaches (Daneshgar et al., 2019). In a centralized approach, anonymization processes exist (e.g., that which are enforced by GDPR), although it is unclear how this is consistently applied. To this end, identification of a participant in the data source may not be needed, yet, unique aspects of their character (e.g., how they pronounce a particular word), are still easily identified (Regan and Jesse, 2019).

There are multiple organizations and corporations which focus on the benefits of decentralization, including Effect.AI and SingularityNET⁴ Such organizations promote benefits including “diverse ecosystems” and “knowledge sharing.” The Decentralized AI Alliance⁵ is another organization which integrates AI and blockchain, promoting collaborative problem-solving. In general, the term decentralization comes not only from technical network logistic but from philosophical “transhuman” ideologies (Smith, 2019). In regards to the latter, decentralization promotes the improvement of human-wellbeing through democratical interfacing with technology Goertzel (2007). This democratic view is one aspect of decentralization that aids in the reduction of AI bias (Singh, 2018).

Similarly, there are organizations which focus primarily on the challenge of bias in AI, from many viewpoints including race, gender, age, and disability⁶, most of which implement responsible research and innovation (RRI). When applying RRI to the AI community, the aim is to encourage researchers to anticipate and analyse potential risks of their network, and ensure that the development of AI is socially acceptable, needed, and sustainable (Stahl and Wright, 2018). Biases are an essential aspect of AI RRI (Fussel, 2017), as poor identity-representation has dire consequences for real-world models (Zliobaite, 2015).

There are an array of concerns relating to the ethics of AI, including, joblessness, inequality, security, and prejudices (Hagendorff, 2019). With this in mind, academic and industry-based research groups are providing tools to tackle these ethical concerns (cf. Table 1), mainly based on four key areas. In this section, we introduce and conceptually discuss these four ethical considerations—auditing, benchmarking, confidence and trust and explainability and interpretability—chosen, due to their prominence within the AI community. As well as this, these four aspects, each have a pivotal impact on data representation, and an inherent relation to data infrastructures. An overview of a typical machine learning workflow with these four considerations highlighted based on their position in time is given in Figure 2. To this end, herein, we first define our four considerations more concretely, followed by a description of specific ethical concerns ([±]) which relate to them.

Having defined our four key consideration more concretely, we now discuss them more closely with representation (w.r.t., bias) and infrastructure of AI data in mind. Where meaningful, we highlight technical approaches which are implemented to reduce the aforementioned ethical concerns.

Due in part to the ethics-based commitments by some of the larger AI companies, we see from this review that, there is momentum toward a more ethical AI future. However, interdisciplinarity in AI research is one aspect which requires more attention. To the best of the authors' knowledge, most public forums (particularly those based on a centralized infrastructure) come from a mono-domain viewpoint (e.g., engineering). Incorporating multiple disciplines in the discussion appears to be more prominent with those promoting decentralized AI.

Interdisciplinary will not only improve implementation of the four ethical consideration described herein, but has been shown to be a necessary step forward for the next AI phase of Artificial General Intelligence (AGI), proposed by the decentralized community (Goertzel and Pennachin, 2007). Interdisiplinarity is particularly of value as infrastructures developed in this way more easily tackle ethical concerns relating to; (i) integration, (ii) selection-bias, and (iii) trust.

Seamless integration of AI is necessary for its success and adoption by the general public. Aspects including cultural and environmental impact need to be considered, and various experts should provide knowledge on the target area. For example, the synthesized voice of bus announcements not representing the community to which it speaks may have a negative impact on those communities, and a closer analysis of the voice that best represents that community would be more ethically considerate. In this way, working alongside linguists and sociologists may aid development.

Similarly, from our literature overview, we observe that knowledge of selection-bias often requires contributions from experts with non-technical backgrounds, and an approach for facilitating discussion between fields of research would be a valuable next step. For example, within the machine learning community, techniques, such as few-shot learning are receiving more attention in recent years (Wang and Yao, 2019), however, perceptual-based biases pose difficulties for such approaches (Azad et al., 2020), and discussion from experts of the targeted domains may help understand the bias at an earlier stage. Despite this, communication between fields speaking different “languages” (i.e., anthropology and engineering), is a challenge in itself, which should be addressed by the community. Furthermore, due to historical stereotypes, AI continues to lack in trust by the general user. Users who without an understanding of the vocabulary of the field, may not be able to grasp the concept of such networks. Through a better collaboration with various academic researchers, communicating AI to the general public may also see an improvement, which in turn will help to build trust and improve wellbeing of the user during AI interaction.

The themes of data representation and infrastructure as they pertain to selection-bias and decentralization in AI algorithms have been discussed throughout this contribution. Within these discussion points, we have highlighted four key consideration; auditing, benchmarking, confidence and trust, and explainability and interpretability to be taken into account when handling AI data more ethically.

From our observation, we conclude that for all of the four considerations, issues which may stem from multimodal approaches should be treated cautiously. In other words, relating to auditing, there should be standards for each modality monitored, as this follows through into the ability for accurate benchmarking. In this same way, although the literature may argue this, confidence and trust come from diverse representations of human data, which in turn are more explainable to the general public due to its inherent human-like attributes.

With this in mind, we see that efforts are being made, for fully audited, benchmarkable, confident, trustworthy, explainable and interpretable machine learning approaches. However, standardization for the inclusion of all of these aspects is still needed. Furthermore, with the inclusion of multiple members who take equal responsibility, decentralization may enable the ethical aspects highlighted herein. We see that through social-media (which is in some sense a decentralized network for communication) group morality is developed. Opinions of a political nature, for example, are highlighted, and any prejudices or general wrongdoing is often shunned and which can have enormous impact on business (Radzik et al., 2020). In this way, a more transparent and open platform makes masking potential network biases a challenge.

AB: literature analysis, manuscript preparation, editing, and drafting manuscript. BS: drafting manuscript and manuscript editing. All authors revised, developed, read, and approved the final manuscript.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.