Corresponding author: Aditya Vasan Srinivasan ( aditya.vasan22@gmail.com ) Academic editor: Peter Hartog
© 2020 Aditya Vasan Srinivasan, Mona de Boer.
This is an open access article distributed under the terms of the Creative Commons Attribution License (CC BY-NC-ND 4.0), which permits to copy and distribute the article for non-commercial purposes, provided that the article is not altered or modified and the original author and source are credited.
Citation:
Vasan Srinivasan A, de Boer M (2020) Improving trust in data and algorithms in the medium of AI. Maandblad Voor Accountancy en Bedrijfseconomie 94(3/4): 147-160. https://doi.org/10.5117/mab.94.49425
|
Artificial Intelligence (AI) has great potential to solve a wide spectrum of real-world business problems, but the lack of trust from the perspective of potential users, investors, and other stakeholders towards AI is preventing them from adoption. To build and strengthen trust in AI, technology creators should ensure that the data which is acquired, processed and being fed into the algorithm is accurate, reliable, consistent, relevant, bias-free, and complete. Similarly, the algorithm that is selected, trained, and tested should be explainable, interpretable, transparent, bias-free, reliable, and useful. Most importantly, the algorithm and its outcomes should be auditable and properly governed.
Artificial Intelligence, Data, Data Quality, Algorithm, Trust
The identification of essential trust factors of data and algorithms in the medium of AI, and the presentation of a trusted AI model incorporating such factors with detailed indicators, would be one of the prime contributions from this research. It would aid technology developers in assessing those trust factors upfront and thereby providing a seal of trust to potential users, stakeholders over the resulting AI solutions.
Perhaps, the most transformative technology that is available today is Artificial Intelligence (AI). The resurrection of AI has been driven in recent times by the availability of more data, increased computing power and advances in algorithms. AI is claimed to ultimately transform every business in every industry (
The adoption of AI has, however, been met with skepticism from various stakeholders, as the technology is relatively new, untried and has posed several complications. One of the major impediments in adopting AI-based solutions is how to trust the particular technology. Imagine when two people meet for the first time, the first thing that would strike them is, questioning themselves: Can the other person be trusted? Similarly, in the context of AI, where investors, consumers or end-users encounter an AI-solution for the first time, what that typically strikes them is whether the AI-solution can be trusted. This question on trust has become a major concern in recent times. People are worried that AI won’t behave as intended, that it makes mistakes, produces unethical results, shows bias, and through that harms society (
In the study of human-robot interaction, trust can be defined as “willingness of the people to accept robot-produced information and follow robots suggestions, share tasks, contribute information, and provide support to the robot” (
The problem statement of this research can, therefore, be described as:
“AI has the capability to transform the business model of every firm and industry, but the lack of trust from the investors, consumers and society as a whole is one of the major concerns and challenges that are preventing the adoption of such technology”.
As can be seen in Figure
Siau and Wang (2017) cited that building trust is a dynamic process that involves movement from initial trust to continuous trust development. Likewise, in the context of AI, initial trust primarily depends on the quality of data and ensuring that data that is being fed to algorithms is completely trustworthy, while continuous trust depends on the performance and the purpose of algorithms.
In line with the above problem statement, the following section (2) defines the objective and scope of the research. Section (3) provides detailed analysis on the theoretical concepts of trust in terms of data and algorithms and it follows a sequential pattern. First, conscious of the value of data in AI, an overview of data quality and its dimensions is presented, followed by the identification of potential trust factors of data. Next, in the context of algorithms, potential trust factors of algorithms are identified based on studies conducted on trust factors in the field of automation and important themes laid by academic institutions and companies. Section (4) describes the interview process conducted with actors. One of the key aspects of this section is the analysis and findings from the interviews. Using the findings from interviews as a reference along with the analysis done on section (2), a trusted AI model is developed by recognizing the relevant phases in an AI development along with the associated trust factors with detailed indicators. The results therefore are presented in section (5). Additionally, section (5) discusses the value of the trusted AI model. In the final section (6), the research determines the essential trust factors of data and algorithms in order to realize the primary contribution to academic research, along with limitations and recommendations.
AI – Artificial Intelligence, ML – Machine Learning, DQ – Data Quality, IQ – Information Quality, GDPR – General Data Protection Regulation, SMEs – Small and Medium-sized Enterprises.
When it comes to autonomous systems, the level of human trust in such systems can be determined based on human characteristics, environment characteristics, and technology characteristics (
The remainder of this study focuses on the category ‘technology characteristics’. In the context of AI, technology characteristics relate to both the data used in the development of AI-solutions and that data’s quality (DQ), and the algorithms which are the end result of the development process. The study is an attempt to gain an initial understanding of the essential trust factors of data and algorithms in the medium of AI and a first move towards the development of a trusted AI model. Such model is aimed at guiding organizations and their internal auditors in the assessment of AI-solutions, with the overall objective to build trust in them. The objective of this research is, therefore, to identify the essential trust factors of data and algorithms in the medium of AI and develop a trusted AI model with detailed indicators for each of the identified factors.
The research is exploratory in nature in terms of identifying the essential trust factors of data and algorithms in the medium of AI and uses a qualitative study with an inductive approach in order to generate valuable theory. The research consists of three phases: the first phase focuses on the identification of potential trust factors through a literature review, the second phase is aimed at validating the trust factors, identified in the first phase, with different actors involved in the development of AI-solutions, the third phase attempts to capture the views from the previous two phases in a trusted AI model with detailed indicators for each of the identified factors.
In the second phase, a clear distinction is made between the various types of actors involved in AI-development. Most of the research conducted on AI till date hasn’t necessarily focused on the actors or stakeholders involved in the AI development process. Phase two of this research starts with identifying the actors typically involved in these development processes. This overview is not intended to and does not cover all relevant actors in all relevant situations of AI-development, as when AI becomes more complicated and expansive, we may need to identify more actors than done in phase two of this research. Based on the engagement with AI experts and the examination of several white papers, the research has identified a number of probable actors, and these were the actors who were interviewed in order to find the essential trust factors of data and algorithms. The probable actors were AI experts, data scientists, data engineers, risk advisors, and auditors. These actors would be a representation for most of the large sized technology and auditing firms.
A literature study was performed to identify potential trust factors of data and algorithms in the context of AI. The study examined literature on concepts of trust in terms of data quality dimensions, and factors influencing trust in automation. Although extensive studies have been conducted on the concept of trust and its factors from multiple dimensional aspects in scientific academia, no substantial research has been done to determine the factors that could possibly influence trust in emerging technologies, and AI in particular. For this study we have presumed that some of the trust factors that influence trust in automation would also hold key for influencing trust in AI, as automation is one of the dimensions of AI. Factors that could influence trust towards the data were recognized in terms of DQ dimensions. There is no single agreement on what dimensions constitutes DQ, as multiple dimensions of DQ could be found in the academic literature and they are subject to variation, depending on the context of the environment. As such, for this research, the chosen factors were ensured to: (1) be the most frequently recommended dimensions for DQ, (2) be relevant to the context of AI, and (3) be able to highly influence the trust in AI as a whole.
Data is considered to be good only when it conforms to relevant use and meets the basic requirements, and the best way to assess it is by DQ dimensions. DQ dimensions are the most commonly accepted and widely used method in the assessment of data, and they have been gaining a great deal of attention from researchers and data experts in recent years due to the abundance of (big) data and its influence in AI. It is commonly believed that most enterprises are moving towards data-driven decision making, using these technologies, and it is essential to have data that meets the required quality. That means, amongst other things, that all fields in the data should be present, that there are no duplicates in the data, that the data doesn’t contain any special characters, etc. Data which is of low quality, outdated, incomplete or incorrect at different stages of data processing, would lead to poor prediction and assessment and, in turn, to biases, possibly resulting in the breach of fundamental rights of the individual, or purely incorrect conclusions and bad outcomes (
One of the central aspects of this research is the quest for dimensions that constitute data quality in the context of AI:
Accuracy – refers to the degree to which the data is correct, reliable, certified and free of error (
Consistency – refers to the degree to which the data is presented in a format that is the same and compatible with previous data (
Completeness – refers to the degree to which a given dataset contains all the relevant data and mandatory attributes which shouldn’t be null (
Security – refers to the extent to which data received from third parties, client sources and other external sources is highly secured, especially when the data holds sensitive information.
Privacy – When the data holds user/customer information, assurances must be made that such data is not being used unlawfully and usage is in compliance with necessary data protection laws and EU General Data Protection Regulation (GDPR).
Relevancy – refers to the extent to which data obtained is relevant to the specific domain and purpose of interest in a given context (
Interpretability – refers to the extent to which the data obtained is in proper language and metrics, meaningful, and the symbols used are understandable to the users working on the data (
Reliability – refers to the extent to which data obtained from client sources or any external party is trustable and can be relied on to convey the right information (
Bias-Free or Objectivity – refers to the extent to which the data obtained and processed is unbiased, unprejudiced, and impartial (
Usefulness – refers to the extent to which the information (processed data) is applicable to the actual context of the goal and useable for the task at hand (
Since not many models of trust were proposed in academia with respect to the development of trust in AI, several researchers have identified multiple trust factors and proposed models in developing trust in automation. It is assumed that some of those factors could also influence trust in algorithms and its resulting outcomes in the context of AI, as automation is one of the dimensions of AI. One of the most widely accepted definitions of automation in the academic literature, which summarizes the whole process of automation, was proposed by
“Automation is a technology that actively selects data, transforms information, makes decision, or control process” (
The same researchers proposed a basic definition of trust, which is in line with their evaluation of trust as an attitude in automation: the attitude that an agent will achieve an individual’s goal in a situation characterized by uncertainty and vulnerability and in this case, an agent could be an automated machine driven by AI.
The issue of trust in AI is top of mind for several academic researchers, technology developers and consumers, as they can exhibit vulnerabilities such as privacy and security concerns, lack of explainability and transparency, exposure to bias, etc. Academic researchers and experts are confronting such issues by laying down some foundational elements of trust as themes of AI in order to trust the technology. There has been no consensus or universally accepted elements on what contributes to the themes of AI, as several tech giants and research institutions have their own defined principles and themes of AI. A detailed summarization of these themes has been presented in Table
Overview of the AI themes laid by leading technology giants, EU commission, and research institutions.
Themes | Description | Source |
---|---|---|
Fairness | Using training data and models that are free of bias in order to avoid unfair treatment of certain groups | Pillars of AI proposed by IBM – (Mojsilovic, 2018) |
Robustness | AI systems should be safe and secure and not vulnerable to tampering or comprising that data that is being trained on. | |
Explainability | Providing decision and suggestions that can be understood by their users or developers | |
Lineage | Including the details of the development, deployment, and maintenance in order to be audited throughout the life cycle. | |
Fairness | Making sure that AI systems are treating all people fairly. | Microsoft principles on AI – (“Microsoft AI principles,” 2019) |
Reliability & Safety | Making sure that AI systems are performing reliably and safely. | |
Privacy & Security | Making sure that AI systems are secured & are respecting privacy. | |
Transparency | Making sure that AI systems are understandable. | |
Fairness | 1.Using representative datasets to train and test the model. | Responsible AI practices by Google – (“Responsible AI Practices,” 2019) |
2.Checking the system for unfair biases. | ||
3.Analyzing the performance of the system. | ||
Interpretability | 1.Designing a model that is interpretable. | |
2.Understanding the trained model. | ||
Privacy | 1.Collecting and handling data responsibly | |
2.Safeguarding the privacy of AI/ML models. | ||
Security | 1.Identifying the potential threats to the system | |
Fairness | Whether the bias in the data and AI model is minimized and Are the bias been addressed when AI is being used? | Dimensions of Responsible AI by PwC – (Building Trust in AI and Data Analytics, 2018) |
Interpretability | Can we explain how an AI model is making decisions and ensuring that the decisions made by AI models are accurate? | |
Robustness & Security | Can we rely on the performance of the AI system and whether AI systems are vulnerable to attack? | |
Governance | Who is accountable for the AI system and ensuring whether proper controls are in place? | |
Technical robustness and Safety | Including resilience to attack and security, fall back plan and general safety, accuracy, reliability, and reproducibility | Some of the key requirements for building a trustworthy AI – (EUROPEAN COMMISSION, 2019) |
Privacy and data governance | Including respect for privacy, quality, and integrity of data, and access to data. | |
Transparency | Including traceability, explainability, and communication | |
Bias evaluation | Understanding, monitoring, and documenting the biases in development and production | Some of the key machine learning principles by Institute of AI & Machine learning – (“The Institute for Ethical AI & Machine Learning,” n.d.) |
Explainability | Developing tools and process to improve transparency and explainability | |
Trust by Privacy | Protect and handle data with stakeholders that interact with the system directly or indirectly | |
Security risks | Develop and improve process to ensure data and model security are taken into consideration. |
Based on the studies conducted on the concept of trust in relation to automation and themes laid down by research institutions and tech giants, potential trust factors of the algorithm in the context of AI have been identified and detailed below:
Accuracy – refers to the ability of algorithms to make accurate judgments, predictions, and recommendations. Users and business leaders would expect a high level of accuracy from the AI based solutions, especially in critical environments that directly affect human lives (healthcare, financial sectors, etc.). This mainly depends on the choice of the algorithm used and the data that is being fed into the algorithm for training.
Consistency – refers to the ability of algorithms to exhibit the same behaviour without ambiguities, when repeated under same conditions.
Interpretability – refers to the ability of algorithms to be interpretable and to explain its actions (
Bias-Free or Objectivity – refers to the ability of algorithms to exclude biases from the resulting AI solution. This by ensuring the solution is not being trained with biased data.
Reliability – refers to algorithms’ ability to work appropriately with a range of inputs and a range of situations in order to prevent accidental harmful outcomes (European Commssion 2019). It is critical that the outcomes produced by algorithms are reliable, and performance is as desired, in order to place trust in the resulting AI solution.
Explainability – When AI gets more complicated, we speak of decisions being made by the so called ‘algorithmic black box’. In order to place confidence in the outcomes produced by such technology and to cement the trust of the stakeholder, as they want to know why specific outcomes occur, it becomes necessary to know the rationale of how the algorithm arrived at its decision or recommendation.
Transparency – To trust AI solutions, one would expect AI-based solutions to be programmed and functioning according to the goal defined and containing evaluation metrics by which the likely behaviour of the algorithm can be inferred.
Auditability – refers to the ability for the system to undergo assessment, not only on the algorithm but also on the data and process encompassing it (
Governance – The term governance in the context of AI aims to close the gap that exists between accountability and ethics in technological development. Governance in AI involves identifying answers to the questions surrounding the safety of AI, identifying what legal and institutional sectors to involve, identifying who has the control and access to personal data, identifying what the roles are of moral and ethical institutions when interacting with AI, identifying control, and monitoring of the algorithms and the data (
Usefulness – An algorithm is said to be only useful when it produces accurate results, has an explainable interface, has consistent results that should be reliable, and has transparency that helps the auditors assess the algorithm’s behaviour.
The list of potential factors identified from the academic literature was presented to the actors to make them familiar with the context of the issue. Furthermore, some potential use case examples were explained upfront to accentuate the importance of trust and factors influencing trust towards AI. The actors could then identify the factors from the visual aids that were presented beforehand, in order to identify the important factors from their perspective. The actors were expected to give a clear explanation when they identified the important factors from their perspective, as there could be factors which could be tied to both the data and algorithm. For example, accuracy can be considered as one of the main dimensions of DQ and it can also be tied to the algorithm as a trust factor in terms of the accuracy of results the algorithm produces. Figure
Actors such as data scientist, data engineer and stewards were interviewed to understand the value of data, its challenges, and its influence towards the algorithm.
Actors proficient in dealing with and mitigating the risk of emerging technologies were involved (risk advisors & auditors).
AI experts and specialists were also involved in this study.
Upon analysing interview transcripts using coding, it was observed that accuracy, auditability, bias-free, consistency, governance, explainability, privacy, reliability, security, transparency, and usefulness were identified to be prime trust factors from the perspective of the actors interviewed. Nevertheless, these factors must be clearly attributed towards the data and the algorithm, as there could be possibilities where the factors could mean crucial only to data or algorithms or both. For example, factors like consistency, accuracy, reliability, usefulness, interpretability, etc. can be tied to both data and algorithm. Based on further analysis and distinguishing the factors, it was clear that most of the actors with data backgrounds paid high attention to the data. To be more specific, actors like data scientists, data engineers and data analytics consultants recognized consistency, accuracy, completeness, bias-free, reliability, etc. as the important dimensions of DQ and governance as the most critical trust factor in order to be able to trust the entire data process and its lifecycle. Though these actors didn’t outline many of the trust factors towards the algorithm, it was believed that an algorithm which has accurate, reliable, consistent, bias-free data and is being trained with such data would also be reliable, consistent, and bias-free. Actors like risk advisors, auditors and managers recognized auditability, security, privacy, reliability, bias-free, transparency, explainability, governance, and usefulness as crucial factors in trusting the algorithm and its resulting outcomes. It was also believed that most of these actors would be involved either in the initial phase or at the final phase of developing an AI solution and they would mainly see the above factors as important in order to trust the algorithm and its results.
So, it can be inferred that DQ dimensions like accuracy, reliability, consistency, completeness, bias-free, etc. were highly important and indeed, many actors with a data background emphasized the same factors. Furthermore, risk advisors and auditors highlighted the same set of factors towards the data. Figure
Using our literature research and the findings collected from the interviews with relevant actors as the main source, a trusted AI model was developed. We have not found any existing or concrete trustworthy AI models in academic research that have explicitly defined the prime trust factors towards AI. As such a new model needed to be developed, encompassing the phases involved in AI development and in those respective phases, relevant trust factors with indicators had to be identified.
It should be noted that there are no specific standard phases defined in scientific research with respect to AI development. This is possibly due to the variation’s dependent upon the context of every organization’s workflow. This research has used
(1) Problem/Improvement Exploration, (2) Human-Centered Design, (3) Data Acquisition, (4) Data Preparation and Validation, (5) Feature Selection, (6) Model/Algorithm Selection, (7) Model/Algorithm Training and Testing, (8) Model/Algorithm Validation and Deployment, and (9) Model/Algorithm Monitoring.
It is important to recognize that phases like Problem/Improvement Exploration (1) and Human-Centered Design (2) won’t be elaborated on in this research. The purpose of outlining those phases is to make the model complete and provide a clear view for the reader to help them recognize the actual phases involved in the development of AI. The research places its emphasis from the data acquisition phase on, as the research mainly presents the trust factors concerning data and algorithms. The associated factors for every phase were identified based on the engagement with the actors, along with the analysis done on literature and themes of AI laid down by research institutions and companies.
Data Acquisition (3) – is one of the prime phases to establish initial trust. This is the phase where the developers of data such as data scientists, data engineers and stewards get involved and start gathering the data. There are several ways to acquire the data, which are dependent on the context of the client case. There can be instances where data can be obtained directly from the client or gotten from an external party or data provider or from public data sources. The importance of trust towards the data initiates from this phase where the dimensions of DQ serve as a base in assessing the quality of data and placing the initial trust over such data. Dimensions like reliability, completeness and consistency would be the prime trust factors that need to be assessed regardless of the context of the case. For instance, to assess completeness of the data, the developers working on the data must see whether all the possible states relevant to the user population are being represented in the stored data. Assessing the accuracy of the data at the first instance would be very hard, as the data received would be raw and unstructured most of the times.
Dimension of DQ in Data Acquisition phase.
Factors | Indicators |
Reliability | • The contents present in the data set are credible enough to process. |
• The data has been generated from a trusted source. | |
Completeness | • The mandatory and main attributes in the data are not null. |
• The data has all the possible states relevant to the user population. | |
• The data is completely representative of the real-world state. | |
Consistency | • The source data is consistent even after a certain point in time. |
• The physical instance of the data is in accordance with some standard format. | |
• The value and entries in the data are the same in all the case. |
Data Preparation and Validation (4) – Once the initial trust is established through the data acquisition phase, the same level trust needs to be maintained in the phase of data preparation and validation. This phase involves a series of steps for processing the data. This is also the phase where the processed data can be split into two parts, where the first part is the training data that is used for training the model/algorithm, and the second part is the testing data that is used to evaluate the performance of the model/algorithm once the model is trained (
Dimension of DQ in Data Preparation and Validation phase.
Factors | Indicators |
---|---|
Relevancy | • The processed data is relevant to the specific domain and purpose of interest in a given context. |
• Processed data contains the required variable in the right form and a representative of population interest. | |
Consistency | • The processed data and actual data source are consistent and representative, have no contradiction and are compatible with the previous data. |
• The attribute values in the data have no ambiguities after data processing. | |
Bias-Free | • The biased data identified during the data acquisition and processing phase has been removed. |
• Data contains all the possible representation of the subject, object and has no inclinations towards a specific object or a thing, i.e. no traces of biases, prejudiced, and partial data were found. | |
• Determining the cause if any variance is detected so as to avoid biases. | |
Accuracy | • The processed data hasn’t lost the structure during the data processing stages. |
• The processed data is certified and free of error. | |
• The processed data is accurate, objective and is corresponding to a real-world context. |
Feature Selection (5) – At this point, there need to be checks to ensure that the most valuable information has been derived out of the data, as that can help the algorithm to produce the best results. So, selecting the right features from the data is paramount, as it reduces over-fitting of the algorithm, improves the accuracy of the results by discarding the data that is misleading, and reduces the training time by having only important data points that make the algorithm train faster (
Dimension of DQ in Feature Selection phase.
Factors | Indicators |
---|---|
Relevancy | • The chosen features are applicable and addressing the actual context of the goal. |
Completeness | • The chosen features are representative of the entire population. |
Model/Algorithm Selection (6) – Once trust is established towards the data by mainly assessing the dimensions of DQ as trust factors in the last three phases where data has been the core driver, a relevant model/algorithm should be selected. Choosing an algorithm that can be easily interpretable with respect to the data and can be explainable for every certain output, should be the primary driver to establish continuous trust at this phase. In addition, it needs to be assessed whether the chosen algorithm can produce accurate outcomes if trained and it is not pre-trained to show any biases.
Trust factors in Model/Algorithm Selection.
Factors | Indicators |
Interpretability | • The result that an algorithm can produce is interpretable with respect to the data. |
Accuracy | • The chosen algorithm can produce correct predictions, decisions, and recommendations results if trained properly. |
Bias-Free | • The chosen algorithm is not pre-trained to show any biases and can work well if trained properly. |
Explainability | • The algorithm can provide some explanation for every certain output. |
Model/Algorithm Training and Testing (7) – Having selected the relevant algorithm based on the data analysis, the algorithm requires training and it needs to be trained based on the training data that was kept aside during the earlier phase of data preparation and validation. Once the algorithm has been trained, it needs to be tested. Test data at this stage would come into the picture and this test data should be different from the training data, so as to not give misleading results on the model’s performance. One of the key trust factors at this stage is ensuring that the algorithm that is being trained and tested doesn’t show any biases or variances and, if found, they have to be addressed. Second, to evaluate the accuracy of the algorithm, it needs to produce the same set of desired outcomes when trained and reflect the same outcomes when tested with new data that are bias-free. Also, it requires careful review and evaluation of training and test data with data subject expertise, and in order to improve the feasibility of auditing, it should be made transparent to the (internal) auditors.
Trust factors in Model/Algorithm Training and Testing.
Factors | Indicators |
---|---|
Transparency | • The training and testing data is open to be reviewed by domain experts. |
Usefulness | • Tuning has been made to improve the overall usefulness of the model. |
Bias-Free | • The potential skews found during the training have been addressed. |
• Any skews observed during the testing were identified and addressed. | |
• The training and testing data are free of biases. | |
Accuracy | • Testing data and training data has samples that represent all the targeted subjects, things, objects, etc. |
• The model can produce desired results based on the training. | |
• The model can produce the same set of desired results when tested with new data (test data). |
Model/Algorithm Validation and Deployment (8) – The algorithm that has been trained and tested needs to be validated and deployed. The main purpose for validation is to ensure that the algorithm has taken into consideration of the important themes like explainability, reliability, and usefulness, which serve as trust factors. An algorithm that has an explainable interface is one of the powerful tools that can help users detect major flaws, interpret the model behavior, ideally maximize the model performance, and make alterations based on the outcomes and rationales behind the outcome at this stage. It is also the phase where the actors have to make sure that the algorithm is actually working as intended, addressing the actual challenge and providing value in terms of reliability and usefulness of the outcomes.
Model/Algorithm Monitoring (9) – In this final stage, the algorithm needs to be closely monitored, especially when it is deployed in critical/high risk environments. The key to monitoring the algorithm is to understand the potential weaknesses as it would give a better indication of how the algorithm is performing, how it is interpreting the inputs and where it is failing, etc. when it is deployed in a real-time environment.
On top of these factors come the other core trust factors like auditability and governance, which need to play a central role through the development of AI. In terms of auditability, every change and assumption made with respect to the data and algorithm has to be documented in order to assess and verify the data process and handling of algorithms. This could improve the overall integrity and trust towards the data and algorithms. A detailed indicator for auditability has been presented in Table
The development of a trusted AI model requires visualizing trust factors of the data in the form of DQ dimensions and algorithms and projecting those factors clearly in the respective phases of the AI development. This is one of the key merits to developing a trusted AI model. The presentation of the trusted AI model itself serves as a main and added value to the organization, next to the identification of DQ dimensions and trust factors of the algorithm. Firstly, the trusted AI model can assist technology developers in assessing the technology critically using the trust factors identified in the respective phases of the AI development. Second, the top- and mid-management levels of the organization comprising directors and managers can use this model as a guide or an assessment tool to validate the AI solution at the end or at every phase of the development. The model not only portrays the trust factors present in every phase of the AI development, but also provides detailed indicators for every trust factor. Having these indicators available can guide the management to validate the technology and see what factors would require more attention when the indicators of respective trust factors don’t meet the required expectations. Third, looking at scenario’s where there would be a need to audit an AI solution, internal auditors – in such a situation – can use this model as a guide to critically assess the solution. At this stand, it must be observed that the model as a whole won’t be efficient when it is being used by small and medium-sized enterprises (SMEs) or any AI startups. It is assumed that most of the startups and SMEs have a flexible workflow and they don’t necessarily follow the same phase identified in the model. Nevertheless, the trust factor which is one of the prime aspects of the model would serve as major value provider to those enterprises. As can be seen in Figure
Trust factors in Model/Algorithm Validation and Deployment.
Factors | Indicators |
---|---|
Explainability | • The algorithm can explain every decision, recommendation, and the prediction made. |
• The decision made by the system can be understood and traced by developers. | |
Reliability | • The algorithm is performing as intended. |
• The algorithm is working properly when a new range of inputs is being applied. | |
Usefulness | • The resulting AI solution had addressed the challenges and is providing value to the clients, society and the stakeholders involved. |
Trust factors in Model/Algorithm Validation and Deployment.
Factors | Indicators |
Auditability | • Every change made with respect to datasets that have been documented is verified. |
• A process on how data was received, analyzed, cleansed, filtered, and labeled are documented and verified. | |
• Assumptions made on the data to improve the usefulness of the data has been documented and reviewed. | |
• Possible risks and opportunities that were encountered and how they were addressed is documented. | |
• The results produced by AI systems are compared with the actual results. | |
• Reviewing the output of the AI model and the meaning derived from those outputs. | |
• Verify by interviewing the ones responsible for developing AI models that they can understand and explain the AI black box data. | |
• Assessing the already existed data from the client or other third-party vendors to validate the usefulness of the data for requirements. | |
Governance | • A clear line of accountability, responsibility standards is clearly established. |
• Know the responsibility of the data & model owners at each and every phase involved in the creation of AI. | |
• Ensuring that human is involved in every phase of the model as well as to oversee the overall activity of the AI application. | |
• Data protocols outlining | |
1. Who can access the data and under which circumstances? | |
2. How the data are being handled and how is it being protected. | |
3. Whether the data is in compliance with GDPR & data protection policies. | |
4. What kind of biases were identified, how were they solved? | |
5. Were there any measures been taken to inform the relevant stakeholders when major bases were notified? etc. should be put in place. | |
Under the governance, special attention must be paid to factors like privacy and security | |
Privacy and Security | • Identify and protect the core strategic data assets. |
• Access to final processed data is restricted & secured from unauthorized access. | |
• Processed data has been encrypted & anonymized if any personal or confidential information is present. | |
• Processed data is used fairly and respects the privacy of the user’s data. | |
• The processed data is in compliance with GDPR and other data protection policies. | |
• The chosen model can be resistant to malicious training. | |
• The model is making fair predictions with the data being trained | |
• The model is neither using the test data maliciously nor being fed with malicious data and is making fair predictions with the test data. | |
• The data used by AI algorithm or model is not used unlawfully or unfairly against the users. | |
• The AI algorithm are safe and secured and are not vulnerable to tampering or compromising the data that they have been trained on. | |
• AI models are protected from being exposed to unexpected situations. | |
• The AI models developed are limited to the context of the goal. | |
• The AI algorithm respects the user privacy of information |
AI has been gaining more traction over the recent years, especially in business. Several leading tech companies and start-ups have already invested in the development of AI by delivering data-driven AI-based solutions. These solutions are currently more focussed on performing a specific task i.e. more towards narrow AI, but when AI starts getting more complex and moves into the stage of enhanced and cognitive automation, the value of trust becomes increasingly important. Leading research institutions, scientific researchers, and tech companies have already started addressing the importance of trust towards AI. Some of the questions being asked upfront about AI are: Does AI explain the results, how is AI using my data, is AI being governed, etc. and all these questions lead to one main question – Whether AI can be trusted.
One of the major challenges during the development of AI are the trust issues that are encircled not only towards the outcome of an AI solution, but in the process of developing those solutions. Data has been the seed to the success of AI and as such the quality of data is an essential property that determines the value and trust towards the data. To trust the data, one needs to look for all relevant dimensions of DQ and needs to validate recursively at every phase of handling and processing the data.
Bad DQ can lead to bad information quality (IQ) and good IQ can imply good DQ. Good DQ however might not lead to good IQ.
Assessing the quality of data using the dimensions is certainly not a once-off validation as there are multiple levels of dealing with the data in the light of AI. So, dimensions of DQ need to be assessed right from gathering the data to the phase where the data is processed and ready as information. Once such processed data is loaded to the algorithm, the trust needs to be shifted towards the algorithm’s behaviour and its outcome.
The assumption is that realizing the trust factors associated with AI and assessing them upfront using a model can improve the overall trust of the technology and thereby provide a stamp of trust to the required stakeholders.
The identified factors, though they seem to be comprehensive in this paper, can be further boiled down to determine the essential set of trust factors of the data and algorithm in the medium of AI. Deriving the essential factors of the data and algorithm would be one of the prime opportunities for academic research. With respect to the data, it can be concluded that factors such as reliability, accuracy, completeness, consistency, bias-free, relevancy are the key factors that were identified to be indispensable in the context of AI, based on findings from the interviews conducted with relevant actors and from the studies done on the DQ dimensions, factors influencing trust in automation and themes of AI. So, the data that is received and processed needs to be consistent, complete, relevant, bias-free, accurate and reliable and these dimensions can’t be ignored and must be vital regardless of the context of the case and the target environment. In the context of the algorithms, the most vital trust factors identified from this research were bias-free, explainability, interpretability, reliability, transparency, and usefulness of algorithms. Factors like bias-free, explainability and interpretability were highly stressed during the interviews and it was even emphasized by several AI research institutions, as these factors would ideally help to place stronger trust in the technology and help develop AI for good.
Furthermore, governance was identified as the most essential trust factor that needs to be in place during the whole process (mainly from data gathering to monitoring the final AI solution). Perhaps, governance would serve as an umbrella for major influential trust factors like privacy, data security and other unexplored trust factors like accountability and ethicality, which might play a prime role in the resulting AI solution. Most interestingly, one of the trust factors, auditability happens to be highly important from the interview findings, as it was believed that examining the AI solution thoroughly can positively influence the trust in AI. The consensus was found to be that the AI solution developed has to be explainable, interpretable and transparent, as that would improve the feasibility of auditing such solutions. In such a scenario, factors like explainability, interpretability, and transparency of the algorithms would stand as the main prerequisites for internal auditors while assessing the resulting AI solutions.
Though there might seem to be an array of trust factors in terms of data and algorithms (as presented in Figure
Mr. A.V. Srinivasan MSc is graduate student at Delft University of Technology.
Drs. M. de Boer RE RA is director data analytics at PwC and PhD researcher at the University of Amsterdam.
■ Bias – Inclination of prejudice towards or against a particular person, subject, or object.
■ Blackbox Algorithm – Algorithm that makes decisions without explanation or without traces on how the algorithm has arrived at a decision.
■ Data Quality Dimension – Set of data quality attributes that represent a single aspect or construct of data quality.
■ Feature Selection – The process of selecting a subset of relevant features from a large pool of features present in the data.
■ Information Quality – Quality of the processed data.
■ Machine Learning – Process of enabling a machine to learn from the data without explicitly programming it with rules.
■ Narrow AI – AI that is designed to perform a repetitive and narrow task. Also, referred as Weak AI or Artificial Narrow Intelligence (ANI).
■ Test Data – Set of new-data (data never seen by algorithm) that is separate from the training data which is used to evaluate the algorithm post training.
■ Training Data – Set of data that is used to train the algorithm.