Ethics and Artificial Intelligence – Daniel Varona, Ph.D.

Navigating the Ethical Landscape of Artificial Intelligence: A Collective Responsibility

Artificial intelligence is a rapidly growing field with tremendous potential to improve our lives in countless ways. However, as we continue to develop more advanced algorithmic decision-maker systems, it’s important to consider the ethical implications of this technology. Researchers and practitioners in the field of AI ethics are working to ensure that AI is developed in a way that is fair, transparent, and responsible.

Some of the major current research interests around AI ethics include topics like bias, discrimination, and fairness. AI systems can inadvertently perpetuate societal biases and contribute to widespread discriminatory decisions if they are trained on data that is biased in some way. For example, a facial recognition system that is trained on a dataset of mostly white faces may not perform as well on faces of other ethnicities. Researchers are working to develop techniques to detect and mitigate bias in AI systems, which is usually understood as fairness.
Another area of interest is the issue of explainability. As AI systems become more complex, it can be difficult to understand how they make decisions. This can be a problem in situations where the consequences of an algorithmic decision-making system’s decisions are significant, such as in healthcare or justice administration. Researchers are working to develop methods for making AI systems more transparent and explainable so that their decision-making processes can be better understood.
A further ethical concern with AI is its impact on jobs and employment. Some researchers argue that automation powered by AI will disrupt many jobs, leading to job losses and social dislocation. Researchers are looking into ways to mitigate this negative impact.
Finally, autonomy is another key area of concern. As AI systems become more capable, they may be able to make decisions on their own, with little or no human supervision. This raises questions about accountability and responsibility. Researchers are working to determine accountability and remedies in scenarios where algorithmic decision-making is present.

Overall, AI has the potential to greatly benefit society, but it’s important to consider the ethical implications of this type of technology. By working to ensure that AI is developed in a fair, transparent, and responsible manner, we can maximize its benefits while minimizing its risks.

Risks of algorithmic decision-making

When we are unaware of the risks of blindly trusting algorithmic decision-making, we may be at risk of accepting decisions made by AI systems without fully understanding how or why they were made. This can lead to several problems:

Bias: As mentioned earlier above, AI systems may perpetuate societal biases if they are trained on data that is biased in some way. To use a different example from the one already mentioned, consider a job recruitment algorithm that is trained on resumes from mostly male candidates, it will surely be less likely to recommend female candidates.
Lack of accountability: If we don’t understand how an AI system is making decisions, it can be difficult to hold anyone accountable for its actions. This can be a problem in situations where the consequences of an AI system’s decisions are significant, such as in healthcare or criminal justice.
Lack of transparency: Without an understanding of how an AI system is making decisions, it can be difficult to ensure that it is operating in a fair and just manner. This can lead to mistrust in the technology and its results.
Lack of trust: If people do not trust the results of AI-driven decisions, they may be less likely to act on them. This can be a problem in cases where the decisions are important, such as in emergency situations.
Unintended consequences: Without understanding how an AI system is making decisions, it can be difficult to anticipate and address any unintended consequences of its decisions.

To mitigate these risks, it’s important to be aware of the limitations and potential biases in AI systems and to develop methods for making them more transparent, explainable, and accountable. Additionally, it’s important to have human oversight and decision-making in the loop to ensure that the AI system is being used in an ethical and responsible manner.

Responsible AI

The development of ethical AI is a collective responsibility that involves various stakeholders, including:

AI researchers and practitioners: These individuals are responsible for designing and building AI systems that should be ethical and fair. These actors should be familiar with the ethical implications of AI and take them into account during the development process.
Policymakers and regulators: These individuals are responsible for creating laws and regulations that govern the use of AI. They should ensure that AI is developed and used in a way that is consistent with societal values and protects individuals’ rights and interests. It can not be stressed enough the need and importance of Policymakers, regulators, and AI researchers and practitioners to work side by side.
Businesses and organizations: These entities are responsible for implementing and using AI in a way that is consistent with ethical principles. They should ensure that AI systems are transparent, fair, and accountable and that any negative impacts are minimized.
Civil society: This includes individuals, groups, and organizations that work to promote the public interest. They should help to raise awareness about the ethical implications of AI and advocate for policies that promote ethical AI development and use.
AI users: This includes individuals and organizations that use AI systems. They should be aware of the limitations and potential biases of AI systems and use them in an ethical and responsible manner.

It’s important to note that the responsibility for ethical AI is not just limited to the development stage, but it’s also important to ensure that the AI system is being used and maintained in an ethical way throughout its entire lifecycle.

Explainable AI: An Overview of Key Papers and Techniques

Artificial intelligence (AI) has the potential to revolutionize many areas of human life, from healthcare and education to transportation and finance. However, as AI systems become more complex and powerful, it becomes increasingly important to ensure that they are transparent, accountable, and interpretable so that their decisions and behaviors can be understood and trusted by humans.

In this blog post, I will explore some of – in my opinion- the key papers and techniques in the field of explainable AI, which aims to provide methods for interpreting and explaining the predictions and behaviors of AI systems. I will start with a brief overview of some of the main challenges and motivations for explainable AI and then delve into some of the most influential papers and techniques in the field.

Challenges and Motivations for Explainable AI

One of the main challenges of explainable AI is the “black box” nature of many machine learning models, which can be difficult or impossible to interpret by humans. These models often involve complex mathematical equations and processes that are beyond the understanding of most people, and the decisions and predictions made by these models can be difficult to explain in a meaningful way.

This lack of interpretability can be a major barrier to the adoption and deployment of AI systems, particularly in fields such as healthcare, finance, and law, where the consequences of incorrect or biased decisions can be significant. In addition, the lack of interpretability can make it difficult to debug or improve AI systems and can hinder the development of trust between humans and AI.

There are several motivations for explainable AI, including:

Accountability: Explainable AI can help ensure that AI systems are accountable for their decisions and actions, and can help identify and mitigate any potential biases or errors in the model.
Trust: Explainable AI can help build trust between humans and AI, by providing a means for humans to understand and verify the decisions and predictions made by the AI system.
Debugging: Explainable AI can help identify and debug errors or biases in the model, by providing insights into the factors that influenced the model’s predictions.
Model improvement: Explainable AI can help improve the performance and interpretability of the model, by providing feedback on the model’s behavior and highlighting areas for improvement.

Key Papers and Techniques in Explainable AI

There have been many influential papers and techniques in the field of explainable AI, which have contributed to the understanding of how to make machine learning models more interpretable and transparent. Here, I will highlight some of the papers and techniques that I find the most notable in this field:

“Why Should I Trust You? Explaining the Predictions of Any Classifier.” by Marco Tulio Ribeiro, Sameer Singh, and Carlos Guestrin (2016) Link

The paper introduces a method for generating post-hoc explanations for the predictions of any classifier, called LIME (Local Interpretable Model-Agnostic Explanations).

LIME is designed to be model-agnostic, meaning that it can be applied to any classifier, regardless of the specific model architecture or learning algorithm used.

LIME explains a classifier’s predictions by approximating the classifier’s decision boundary with a linear model, which is locally faithful to the classifier’s behavior around the point being explained.

The paper presents empirical results demonstrating the effectiveness of LIME in generating human-interpretable explanations for a variety of classifiers, including decision trees, logistic regression, and neural networks.

The paper also discusses the use of LIME for debugging classifiers and improving their interpretability, as well as for building trust between users and machine learning models.

The authors provide open-source code for implementing LIME, which has since been widely adopted and extended in the research community.

2. “Interpretable Machine Learning: A Guide for Making Black Box Models Explainable” by Christoph Molnar (2019) Link

The book provides a comprehensive overview of interpretable machine learning, including definitions, methods, and applications.

It covers a range of topics, including model-agnostic interpretability, post-hoc interpretability, and inherently interpretable models.

The book discusses different approaches to interpretability, including feature importance, decision trees, and partial dependence plots.

It also covers more advanced techniques such as rule-based models, local interpretable model-agnostic explanations (LIME), and sensitivity analysis.

The book provides detailed examples and case studies to illustrate the concepts and techniques discussed, as well as practical advice on how to apply interpretable machine learning in real-world scenarios.

It also addresses important considerations such as the ethical and societal impacts of interpretable machine learning, and the trade-offs between interpretability and accuracy.

3. “A Unified Approach to Interpreting Model Predictions” by Scott Lundberg and Su-In Lee (2017) Link

The paper introduces a method for generating global explanations of model predictions, called SHAP (SHapley Additive exPlanations).

SHAP is based on the idea of Shapley values from game theory, which provides a method for fairly distributing the contributions of different players to the overall value of a game.

SHAP assigns a unique importance value to each feature of a model’s input, based on its contribution to the model’s output.

The paper presents empirical results demonstrating the effectiveness of SHAP in generating accurate and consistent explanations for a variety of models, including linear models, decision trees, and neural networks.

The paper also discusses the use of SHAP for model debugging, model comparison, and feature selection.

The authors provide open-source code for implementing SHAP, which has since been widely adopted and extended in the research community.

4. “Human-Interpretable Machine Learning” by Gabriele Tolomei, Fabio Pinelli and Fabrizio Silvestri (2022) Link

The editorial presents a number of Frontiers Big Data introducing a framework for designing machine learning models that are both accurate and interpretable by humans.

The framework is based on the idea of decomposability, which refers to the ability to explain a model’s output as a combination of the contributions of its input features.

The number discusses different types of decomposable models, including linear models, decision trees, and additive models, and presents empirical results demonstrating their effectiveness in a range of applications.

The number also discusses the trade-offs between interpretability and accuracy and presents strategies for balancing these objectives.

The authors provide open-source code for implementing decomposable models, and discuss the potential applications of these models in areas such as healthcare and finance.

5. “Explainable Deep Learning for Speech Enhancement” by Sunit Sivasankaran, Emmanuel Vincent and Dominique Fohr (2021) Link

The paper presents a method for generating explanations of deep learning models for speech-processing tasks, based on the concept of attention.

The attention mechanism allows the model to focus on specific parts of the input during the prediction process, and the explanations generated by the attention mechanism can provide insight into the model’s decision-making process.

The authors apply their method to a number of speech processing tasks, including automatic speech recognition and speaker identification, and demonstrate the effectiveness of the generated explanations in improving the interpretability of the models.

The paper also discusses the limitations of the attention mechanism as a tool for explainability, and presents strategies for improving the interpretability of deep learning models in speech processing.

6. “Network dissection: Quantifying interpretability of deep visual representations” by David Bau, Bolei Zhou, Aditya Khosla, Aude Oliva, and Antonio Torralba (2017) Link

The paper introduces a framework for quantitatively evaluating the interpretability of machine learning models.

The framework is based on the idea of “information fidelity,” which measures the ability of a model’s explanations to accurately capture the information used by the model to make predictions.

The paper presents empirical results demonstrating the effectiveness of the proposed framework in evaluating the interpretability of a variety of models, including linear models, decision trees, and deep learning models.

The authors also discuss the potential applications of the proposed framework in model selection, model comparison, and model debugging.

The paper addresses important considerations such as the generalizability and robustness of the proposed framework and the potential biases in the evaluation process.

7. “Towards Robust Interpretability with Self-Explaining Neural Networks” by David Alvarez Melis, Tommi Jaakkola (2018) Link

The paper introduces a method for generating explanations of neural network models, called Self-Explaining Neural Networks (SENNs).

SENNs are designed to be inherently interpretable, by decomposing the model’s predictions into a combination of the contributions of its input features.

The paper presents empirical results demonstrating the effectiveness of SENNs in generating accurate and consistent explanations for a variety of tasks, including image classification and language translation.

The paper also discusses the limitations of SENNs and presents strategies for improving their interpretability and robustness.

The authors provide open-source code for implementing SENNs, and discuss the potential applications of these models in areas such as healthcare and finance.

8. “On the Quantitative Analysis of Decomposable Explainable Models” by Marco Ancona, Enea Ceolini, Cengiz Öztireli and Markus Gross (2018) Link

The paper presents a method for quantitatively evaluating the interpretability of decomposable machine learning models.

The method is based on the idea of “submodular pick,” which measures the amount of information gained by each feature in a decomposable model’s explanation.

The paper presents empirical results demonstrating the effectiveness of the proposed method in evaluating the interpretability of a variety of decomposable models, including linear models, decision trees, and rule-based models.

The paper also discusses the potential applications of the proposed method in model selection, model comparison, and model debugging.

The authors discuss the limitations of the proposed method and present strategies for improving its robustness and generalizability.

Applications of Explainable AI

In the previous section, I presented some of the key papers proposing techniques in the field of explainable AI, which aims to provide methods for interpreting and explaining the predictions and behaviors of AI systems. In this section, I will consider some of the potential applications and challenges of explainable AI in real-world scenarios.

Explainable AI has the potential to be applied in a wide range of fields and contexts, where the ability to understand and trust the decisions and predictions of AI systems is critical. Some potential applications of explainable AI include:

Healthcare: Explainable AI can be used to support healthcare professionals in decision-making, by providing insights into the factors that influenced the model’s predictions and recommendations. For example, an explainable AI system could help a doctor understand the specific symptoms or risk factors that contributed to a diagnosis or treatment recommendation, and could help identify and mitigate any potential biases or errors in the model.
Finance: Explainable AI can be used to support financial decision-making, by providing insights into the factors that influenced the model’s predictions and recommendations. For example, an explainable AI system could help a financial analyst understand the specific market conditions or company characteristics that contributed to a stock recommendation, and could help identify and mitigate any potential biases or errors in the model.
Law: Explainable AI can be used to support legal decision-making, by providing insights into the factors that influenced the model’s predictions and recommendations. For example, an explainable AI system could help a judge or lawyer understand the specific legal precedents or evidence that contributed to a decision, and could help identify and mitigate any potential biases or errors in the model.
Education: Explainable AI can be used to support educational decision-making, by providing insights into the factors that influenced the model’s predictions and recommendations. For example, an explainable AI system could help a teacher or student understand the specific learning needs or progress of a student, and could help identify and mitigate any potential biases or errors in the model.

Keyword “Explainable”

There are several hurdles to the development and implementation of the feature “explainable” in explainable AI, which can hinder the widespread adoption and impact of these AI systems. Some of the main hurdles include:

Trade-offs with accuracy: One of the main difficulties of explainable AI is that there is often a trade-off between interpretability and accuracy. Many of the techniques used to make machine learning models more interpretable, such as decision trees and linear models, are less powerful and accurate than more complex models such as deep neural networks. This can make it difficult to achieve both high accuracy and high interpretability in the same model.
Complexity and scalability: Some of the techniques used for explainable AI, such as local interpretable model-agnostic explanations (LIME) and global interpretable model-agnostic explanations (SHAP), can be computationally expensive and may not scale well to large datasets. This can make it difficult to apply these techniques to real-world scenarios with large amounts of data.
Model-specific explanations: Many of the existing techniques for explainable AI are model-specific, meaning that they can only be applied to specific types of models or architectures. This can make it difficult to use these techniques to explain the predictions of more complex or hybrid models, such as ensembles or transfer learning models.
Human biases: Explainable AI systems can be subject to human biases, either in the data used to train the model or in the way the explanations are generated or interpreted. It is important to carefully consider and address these biases in order to ensure the fairness and reliability of the explanations.
Lack of standardization: There is currently a lack of standardization in the field of explainable AI, with many different techniques and approaches being used. This can make it difficult to compare and evaluate the effectiveness of different approaches and can hinder the development of best practices and guidelines for explainable AI.

Artificial Intelligence design guiding principles: Review of “Ministerial Statement on Trade and Digital Economy of G20”

The document “Ministerial Statement on Trade and Digital Economy of G20” was authored by the Group of 20 (G20) and was published in Japan in June 2019. It aims to align efforts, among state members and some guests, in defining a set of consistent guidelines with their principles for the use and development of Artificial Intelligence solutions as an opportunity to deepen their understanding of the interface between trade and the digital economy in building a sustainable and innovative global society; considering their national needs, priorities, and circumstances.

Being joint authored, by members of a body that groups the efforts of the governments of several countries: I have classified the author type as “Intergovernmental Organization”. Also, in the light of the objective pursued by the recommendations and the type of principles proposed, I have classified the document as “Policies Principles”. Both classifications will allow me to make future contrasts between documents and authors of the same type; enriching the analysis that I aim to present in this series of posts.

The group adopts the principles stated by OECD, which are annexed in its entirety statement without any change as can be contrasted by viewing the following, and comparing them with t.ly/e2Wq:

Inclusive Growth, Sustainable Development, and Well-being: proactively engage in a responsible management of trustworthy artificial intelligence searching for benefits for people and the planet, increase human capacities and improve creativity, advance toward the inclusion of underrepresented populations by reducing economic, social, gender, and other inequalities; and in the protection of the natural environments,
Human-Centred Values, and Fairness: respect the rule of law, human rights and democratic values during all the life cycle of the artificial intelligence solution – freedom, dignity and autonomy, privacy and data protection, non-discrimination and equality, diversity, equity, social justice and work -; implementing mechanisms and safeguards, such as human capacity for self-determination, that are appropriate to context and consistent with the state of art,
Transparency and Explainability: provide relevant information, appropriate to context and consistent with state of the art: (a) to promote a general understanding of the operation of AI systems, (b) to enable stakeholders to be aware of their interactions with AI systems, (c) to allow those affected by an artificial intelligence system to understand the outcome, and (d) to allow those adversely affected by an AI system to challenge their outcome based on easy-to-understand information about the factors and logic that served as the basis for the prediction, or recommendation,
Robustness, Security, and Safety: develop robust and safe AI systems, and protect them throughout their life cycle so that – under normal use, foreseeable use or misuse, or other adverse conditions – the keep functioning properly without becoming a security risks; ensuring traceability towards data sets, processes and decisions made, and applying a systemic approach to risk management at every stage of the AI system lifecycle that includes factors such as: privacy, digital security, security and bias; and,
Accountability: make the AI actors accountable for the proper functioning of AI systems and its correspondence with the proposed principles, according to their roles, the context and consistent state of the art.

Similarly, they echo OECD’s cooperation policies in defining its national policies and international cooperation among the acceding countries in favor of reliable Artificial Intelligence:

Investing in AI research and development,
Fostering a digital ecosystem for AI,
Shaping an enabling policy environment for AI,
Building human capacity and preparing for labor market transformation, and
International co-operation for trustworthy AI.

Consequently, I emphasize the analysis of the principles proposed by OECD and which can be read in t.ly/e2Wq. and that I pointed as necessary intermediate layers for the adoption of these principles as a methodological reference in the design of artificial intelligence solutions.

After an analysis of the language used in the document, in which I used the NLTK library and the development environment for Python, extracting the 50 most frequent n-grams in statement turns out that:

The uni-grams with relative frequencies greater than 1.00 units described the objective intended with the principles and recommendation proposal, or the variables in which they are expressed: digital (2.99), ai (1.89), economy (1.58), trade (1.31), and development (1.19). While the uni-grams relative frequencies between .05 and 1.00 units represent the action environment: society (.76), international (.73), growth/ investment (.70), policy (.64), global/ sustainable/ use (.58), and innovation/ technologies (.55).
The bi-grams, from their part, delimited the document´s field of action in the context described by the unigrams, exhibiting with higher relative frequencies the terms: digital economy (1.40), trade investment (.37), trustworthy ai (.34), and human centered (.24). Although, other variables such as responsible stewardship and stewardship trustworthy exhibit minimum relative frequency values with 0.12 units each.
The tri-grams, similarly, exhibit a greater representation of the terms linked to the macro-objective of the document with terms like: trade and security digital economy (.24); while the pursued objectives are less represented: Sustainable development goals/ human-centred future society/ competitive non discriminatory/ countries underrepresented populations/ human centred future/ and centred future society/ con .06 unit each.

I would like to conclude by saying that the recommendations of the council on AI addressed in this post, along with other documents I am including in this series; constitutes an effort to solve some of the ethical problems rooted in the design and use of artificial intelligence solutions. In this case, specifically in the context of public policy; the remaining documents will include other scenarios. Also, I would like to add that, with this reading exercise I seek to draw attention to the opportunity of public policy designers and designers of artificial intelligence solutions to collaborate in the achievement of a common goal: what is the responsible design of artificial intelligence.

If you are interested about this topic and have any idea that complements this review of the Recommendations of the Council on Artificial Intelligence let me know with a comment.

Artificial Intelligence design guiding principles: Review of “Recommendation of the council on Artificial Intelligence”

The document ” Recommendations of the Council on Artificial Intelligence” was authored by the Organization for Economic Cooperation and Development (OECD), and was published in France in May 2019. It aims to foster innovation and trust in AI by promoting the responsible stewardship of trustworthy AI while ensuring respect for human rights and democratic values; whose adherence would be ratified by the organization’s member countries, and some non-members such as: Argentina, Brazil, Colombia, Costa Rica, Malta, Peru, Romania, and Ukraine.

Being joint authored, by members of a body that groups the efforts of the governments of several countries in the Americas, Europe, the Middle East, and Australia: I have classified the author type as “Intergovernmental Organization”. Also, in the light of the objective pursued by the recommendations and the type of principles proposed, I have classified the document as “Policies Principles”. Both classifications will allow me to make future contrasts between documents and authors of the same type; enriching the analysis that I aim to present in this series of posts.

The Recommendation identifies five complementary values-based principles for the responsible

stewardship of trustworthy AI and calls on AI actors to promote and implement them:

Inclusive Growth, Sustainable Development, and Well-being: proactively engage in responsible management of trustworthy artificial intelligence searching for benefits for people and the planet, increase human capacities and improve creativity, advance toward the inclusion of underrepresented populations by reducing economic, social, gender, and other inequalities; and in the protection of the natural environments,
Human-Centred Values, and Fairness: respect the rule of law, human rights and democratic values during all the life cycle of the artificial intelligence solution – freedom, dignity and autonomy, privacy and data protection, non-discrimination and equality, diversity, equity, social justice and work -; implementing mechanisms and safeguards, such as a human capacity for self-determination, that are appropriate to the context and consistent with the state of art,
Transparency and Explainability: provide relevant information, appropriate to context and consistent with state of the art: (a) to promote a general understanding of the operation of AI systems, (b) to enable stakeholders to be aware of their interactions with AI systems, (c) to allow those affected by an artificial intelligence system to understand the outcome, and (d) to allow those adversely affected by an AI system to challenge their outcome based on easy-to-understand information about the factors and logic that served as the basis for the prediction, or recommendation,
Robustness, Security, and Safety: develop robust and safe AI systems, and protect them throughout their life cycle so that – under normal use, foreseeable use or misuse, or other adverse conditions – the keep functioning properly without becoming a security risks; ensuring traceability towards data sets, processes and decisions made, and applying a systemic approach to risk management at every stage of the AI system lifecycle that includes factors such as: privacy, digital security, security and bias; and,
Accountability: make the AI actors accountable for the proper functioning of AI systems and its correspondence with the proposed principles, according to their roles, the context and consistent state of the art.

Additionally, a set of recommendations are made, as can be seen below, for the definition of national policies and international cooperation between the adherent countries in favors of trustworthy Artificial Intelligence:

Investing in AI research and development,
Fostering a digital ecosystem for AI,
Shaping an enabling policy environment for AI,
Building human capacity and preparing for labor market transformation, and
International co-operation for trustworthy AI.

On this occasion will limit myself to only comment on the principles. Since the recommendations are aimed at public policymakers; and that is a context which I do not have enough experience in.

From my computer science background with hands-on experience in software project management, I find it difficult to adopt these principles as a methodological reference without them being subject to additional layers of interpretation, and integration into tools such as standards or checklists, to name a few examples. As I have already mentioned in other posts; on the one hand, standards would support the assurance of expected outcomes of the artificial intelligence solutions since early development stages in accordance with the framework scope delimited by the proposed principles; and, on the other hand, checklists are an effective tool in the verification stages, used to whereas the designed solution complied with the proposed principles – using the same examples -.

In that same line of thoughts, from my experience defining checklists, and as a member of international software development standards designing working groups, I can highlight the following elements:

The definition of the conceptual neighborhood for variables related with concepts such as: discrimination, bias, justice, and equity, in the context of artificial intelligence; that can serve as a frame of reference for the software developer at every stage – including maintenance – of the development process,
The operationalization of the concept of well-being as a dependent variable on the discriminatory or non-discriminatory nature of decisions based on decisions, predictions, and/or recommendations proposed by AI systems,
The operationalization of the concept of “natural environment friendly” as a dependent variable on the aggressive or non-aggressive nature of made decisions based on decisions, predictions, and/or recommendations proposed by AI systems,
The formalization of metrics aimed at evaluating how discriminatory or aggressive with the natural environment is a decision, prediction, and/or recommendation proposed by AI systems,
The definition of checklists to guide the developer of AI systems during the verification and measurement of these variables at every stage of the development process,
The demarcation of which values resulting from the measurements and what factors within the checklists triggers a formal review of the architecture baseline of the current version of the AI system being developed,
The demarcation of which values resulting from the measurements and of which factors within the checklists triggers a formal change request in the case of medium and large projects, with medium and high complexity,
The creation of a competent authority that continuously assesses the adequacy of the formalization of measurements to the corresponding social context, including that the causal elements of discrimination and other related terms are variable over time,
The operationalization of the variables on which human rights and the democratic values are based on which artificial intelligence solutions are expected to be in correspondence with,
The operationalization of “transparency” and “understanding” as dependent variables within the understanding of the methods used for data processing, which can be used in defining a metric that assesses the levels at which the understanding of the methods and the results of the AI system by potential stakeholders and auditors, can be expressed with, and
The definition of information management flows associated with the use of AI systems including the necessary elements (access policy to which piece of information, period of time the information will be available, for example) and moderating the communication between the stakeholder and the decision maker (regardless of the latter) to be incorporated – ex officio – in Report modules; helping those adversely affected by an AI system supported decision, to obtain relevant information and details of the decision.

As necessary intermediate layers towards the principle’s adoption as a methodological reference for the design of artificial intelligence solutions.

After an analysis of the language used in the document, in which I used the NLTK library and Python´s development environment for extracting the 50 most frequent n-grams from the charter´s body text it turned out that:

The uni-grams with relative frequencies greater than .50 units described the objective intended with the principles and recommendation proposal, or the variables in which they are expressed: policy (1.10), international (1.02), legal/ trustworthy/ development (.98), council/ principles/ work (.78), digital (.69), human (.65), operations/ stakeholders/ systems (.61), and responsible/ implementation/ systems (.57); in contrast with: stewardship/ rights/ inclusive/ sustainable/ recommendations (.33) that also being elements among the objectives pursued by the document are less represented along the text body.
The bi-grams, from their part, delimited the document´s field of action in the context described by the unigrams, exhibiting with higher relative frequencies the terms: trustworthy ai (.90), ai actors (.53), legal instruments/ international co-operation (.45), and responsible stewardship/ stewardship trustworthy (.33). Although, other variables such as risk management/ growth sustainable/ security safety/ digital ecosystem/ privacy data/ and ai government exhibit minimum relative frequency values with 0.16 units each.
The tri-grams, similarly, exhibit a greater representation of the terms linked to the macro-objective of the document with: international co-operation instruments (.45), responsible stewardship trustworthy/ ai systems lifecycle (.33); while the pursued objectives are less represented: human centred values/centred values fairness/ robustness security safety/ investing ai research/ fostering digital ecosystem/ building human capacity, preparing labor market/ practical guidance recommendations (.12) and, artificial intelligence first/ first intergovernmental standard (.08).

I would like to conclude by saying that the recommendations of the council on AI addressed in this post, along with other documents I am including in this series; constitute an effort to solve some of the ethical problems rooted in the design and use of artificial intelligence solutions. In this case, specifically in the context of public policy; the remaining documents will include other scenarios. Also, I would like to add that, with this reading exercise I seek to draw attention to the opportunity of public policy designers and designers of artificial intelligence solutions to collaborate in the achievement of a common goal: what is the responsible design of artificial intelligence.

If you are interested about this topic and have any idea that complements this review of the Recommendations of the Council on Artificial Intelligence let me know with a comment.

Artificial Intelligence design guiding principles: Review of “European Ethical Charter on the Use of AI in Judicial Systems and their environment”

The European ethics charter on the use of artificial intelligence in Judicial Systems and their environment was authored by the Council of Europe´s European Commission for the Efficiency of Justice CEPEJ and was published in France in December 2018. The charter aims to align regional efforts by defining a set of principles governing the design of artificial intelligence solutions, and their use in the context of the judicial system; based on the International Law of Human Rights.

As the charter was produced by joint authorship, gathering members of several government bodies from several European countries I have classified the author type as “Intergovernmental Organization”. Also, in the light of the objective pursued by the charter, and the nature of the principles it proposes, I have classified the document type as “Policies for Use”. Both classifications will allow future contrasts between documents and authors of the same type; enriching the analysis that I aim to present in this series of posts.

The principles proposed within the charter are listed below:

1. 1. Principle of respect for fundamental rights: ensure that the design and implementation of artificial intelligence tools and services are compatible with fundamental rights,
  2. Principle of non-discrimination: specifically prevent the development or intensification of any discrimination between individuals or groups of individuals,
  3. Principle of quality and security: with regard to the processing of judicial decisions and data, use certified sources and intangible data with models elaborated in a multi-disciplinary manner, in a secure technological environment,
  4. Principle of transparency, impartiality, and fairness: make data processing methods accessible and understandable, authorize external audits, and
  5. Principle “under user control”: preclude a prescriptive approach and ensure that users are informed actors and in control of the choices made.

From my computer science background, I find it difficult to adopt these principles as a methodological reference without them being subject to additional layers of interpretation, and integration into tools such as standards or checklists, to name a few examples. On the one hand, standards would support the assurance of expected outcomes of the artificial intelligence solutions since early development stages in accordance with the framework scope delimited by the proposed principles; and, on the other hand, checklists are an effective tool in the verification stages, used to whereas the designed solution complied with the proposed principles – using the same examples -.

- - The operationalization of the variables fundamental rights are based on, and artificial intelligence solutions are expected to comply with within the environment described by the charter being reviewed,
  - The definition of the environment framing the possible discriminations to which an individual or groups of individuals may be exposed to given the attributes included on each decision; consisting of a finite number given their typology according to the environment enclosed within the charter,
  - The definition of the current causal discrimination variables´ neighbor environments an individual or groups of individuals may be subject to according to the attributes included on each decision; given that discrimination is a variable phenomenon with – nonexclusive – temporal, geographical, and cultural dimensions.
  - The definition of variables that can be integrated into metrics to assess the different intensity levels in possibly discriminatory decisions which individuals or groups of individuals may be exposed to,
  - The creation of a certifying authority refereeing the adequacy of data sources used in support of the decision-making process when delivering justice,
  - The creation of a certifying authority evaluating the suitability of team members, and the completeness of the multidisciplinary team, designers of the models for the processing of data to be used in support of the decision-making process while delivering justice,
  - The operationalization of “accessibility” and “understanding” as dependent variables for understanding the methods used on data processing, which can be used in the definition of a metric that assesses the levels at which understanding of methods can be expressed, by potential stakeholders and auditors,
  - The creation of a competent auditing authority certifying the adherence of the design team – of artificial intelligence solutions for the use of the justice administrators within the environment delimited in the charter – with the proposed principles,
  - The definition of parameters that can be integrated into constraints within the artificial intelligence solution´s reasoning model, while in design stages, to avoid recommending decisions describing a prescriptive approach as per the context delimited by the charter, and
  - The definition of parameters that can be integrated into metrics evaluating the prescriptive nature of the approach described by the artificial intelligence solutions´ recommended decisions within the charter delimited environment.

As necessary intermediate layers towards the principle’s adoption as a methodological reference for the design of artificial intelligence solutions in the administration of justice.

- - The uni-grams with relative frequencies greater than .50 units described the environment delimited in the letter and not the objective intended with the principles proposal, or the variables in which they are expressed: Judicial (.87), Decisions (.82), Law (.72), Processing (.63), Legal (.62), Case (.56), Public/ Tools/ Judges/ Use (.53), and Justice (.52),
  - The bi-grams, however, begin to delimit the charter´s scope in the context described by the uni-grams, displaying, with higher relative frequencies the terms: Machine Learning (.31), Judicial Decisions (.28), Artificial Intelligence/ Open Data (.27), Judicial Systems (.20) and Personal Data (.19). Although, other variables like Data Protection and Fair Trial exhibit lower values, .6 and .5 units of relative frequency, respectively.
  - The tri-grams, on the other hand, connect both the environment and scope using the following text compositions: Protection Personal Data (.09), Artificial Intelligence Tools/ Processing Judicial Decisions/ Judicial Decisions Data (.06), and
  - It is interesting how, through the identified trigrams: Use Artificial Intelligence/ Intelligence Tool Services/ Predictive Justice Tools and Checklist Evaluating Processing/ Evaluating Processing Methods, all with a relative frequency of .05; the letter itself points to the need for tools like the ones mentioned earlier in this post.

I would like to conclude by saying that the charter addressed in this post along with other documents I will further include in this series constitutes an effort to solve some of the ethical problems rooted in the design and use of artificial intelligence solutions. In this case, specifically in the context of the administration of justice; the remaining documents will include other scenarios. Also, I would like to add that, with this reading exercise I seek to draw attention to the opportunity of public policy designers and designers of artificial intelligence solutions to collaborate in the achievement of a common goal: what is the responsible design of artificial intelligence.

If you are interested in this topic and have any idea that complements this review of the European Ethical Letter on the Use of Artificial Intelligence in Judicial Systems and its environment let me know with a comment.

Image is taken from pixabay.

Expendable deaths

Recently, I had a conversation regarding driverless cars’ decision core and the parameters they currently do not take into consideration when analyzing possible outcomes while trying to avoid a collision. We were arguing upon the ethical differences around the globe regarding which lives values the most among the people involved in the accident. On the one hand, we had a pedestrian and on the other hand, we had a passenger. In our discussion, we referred to several studies from different countries were factors such as the age of both pedestrian and passenger will determine who´s death will be more ethically correct.

Some studies point the surveys inclined to favor the youngest in contrast with other surveys exhibiting the opposite idea. The different surveyed regions showed results too dissimile and the available research it is too scarce to adopt an inflexible posture. We finally agreed it was like the tale of a father, a son, and a donkey. In which every other individual they crossed in their path will criticize them no matter of who was riding the donkey, or if any of them was doing it at all. Unfortunately, in the case we were discussing the focus point was not who is more comfortable but who lives and who dies.

I was extremely surprised when my interlocutor said: “Anyhow, driverless cars, once introduced into society, will drop the volume of deaths because of traffic accidents; at the end, no matter who lives and who dies but the number of deaths are being saved from the current stats”.

I was pretty worried about my friend not being a regular person, with a regular job. This person is someone highly educated in business, sciences, and public relations; whose job allows him to influence the actual implementation of developing and research policies, to budget a portfolio of research and innovation projects. A person it’s dedicated to connecting people around the world to facilitate ideas flow and project executions.

Then I wondered if I was the one not seeing an imperfect decision core of an artificial intelligence system was actually capable of lowering the number of deaths by traffic accidents regardless of who dies and who doesn’t. After all, these systems are superior in their imitation of our own decision making in the same situation; by being able to calculate hundreds if not thousands of scenarios while in the moment of distress. And the factors those systems take into consideration are just a detail in the equation.

Nevertheless, I thought about the previous technological and scientific advances we had experienced as humanity; thought about their methodological frameworks and concluded: then why we wasted money and effort with ethical approvals?, If vaccines are intended to lower the number of deaths by a particular virus, or emergent medical treatments; why do we need extense trial periods?, some of them even before experimenting with humans. Why do we need to identify and communicate limitations and secondary effects on our drugs?. Why pilots and airlines have had documented and turned into endless checklists every single parameter influencing a decision?.

So I understood, at least for the pilot example, that accidents should be reproducible in order to assess all outcomes – including the actual one that took place-; and I comprehended that responsibility transference its the real issue, accountability. Who it’s responsible for the deaths at accidents where driverless cars are involved? Like doctors are in medicine as prescriptions expenders, or like the pilots/airlines when a plane crash occurs – a closer example to the one I am referring to in this post -.

Are we truly that comfortable with the idea of twenty deaths instead of hundreds when those twenty, depending on the factors parametrized in the driverless car´s decision-making core, might be typified as murder?. This is not an inaccurate thought. It can manifest by prioritizing with a given weight the value of the passenger (the one who bought the driverless car, paid the dealer, pays the insurance…), and a given – just smaller enough – value to the pedestrian or another driver involved in a sinister; to balance the outcome in favour of one side or the other. And yet radio advertisements are currently highlighting the possibility of sleeping while commuting to work when workers live in a city diffrerent from their workplace using a driverless car.

Please do not get me wrong, I am not promoting any conspiracy theory or promoting freezing fear upon technological progress. On the contrary; my intention goes more attuned with arising the need for a technical framework for artificial intelligence designers based on ethics and moral; highlighting the need for speed up our laws and public policies to meet our current realities, to address possible negative outcome given the introduction of new tech in anticipation to the actual events. I am tuned with responsible design as a way of accomplishing justice from the early stages of design.

How to reference: Varona, Daniel. 2019. Expendable deaths. www.danielvarona.ca [Year of Creation:2019] http://www.danielvarona.ca/2019/06/14/expendable-deaths/ [Consulted on: DDMMYY].

Right from Wrong

Right and Wrong — (Image taken from pixabay)

I often read or listen to people´s ideas on handling accountability in the artificial intelligence field of study and solution development. I could spare the main trends in the two: the first being the need for formal modeling of human morality and ethics, and its further coding into algorithms; and the second, the need to build a machine learning solution who follows (the idea of how is still a black box) human moral and ethics rules. Both, in my opinion, are opposite poles just as the sciences and backgrounds of their promoters. A practical example of the eternal dichotomy between hard-sciences and soft-sciences practitioners and academics.

One can find several instruments supporting the latter trend like the “Statement on Artificial Intelligence, Robotics and ‘Autonomous’ Systems” of the European Group on Ethics in Science and New Technologies; the “The Toronto Declaration: Protecting the rights to equality and non-discrimination in machine learning systems”; or the multinational research project aiming to arise with a number of universal moral rules showcasing a common understanding of every society on the planet. All of them with strong fundamentals but lacking on technical specifications to be followed by computer scientists and software developers.

On the other hand, the first trend finds some mechanisms to evaluate responsibility degree, or to accountability recognition after bad decisions are made. Something similar to David Gunkel´s “responsibility gap” theory. One example of this is the calibration checks algorithms; which will tell whereas an algorithm is being biased at a given point comparing the results for a target object in different datasets (the object included). However, yet exist a long path before we are able to formalize ethical and moral frameworks for artificial intelligence solutions to work with on their learning process.

What brings me to request the attention upon artificial intelligence´s dependence on machine learning development as per the decision making processes, consequently the learning core its a capstone; then machine learning methods and techniques aim to simulate human behavioral systems on its procedures. What makes the very object we are criticizing the reflection of our own – let´s not simplify it by saying poor – but complex learning and decision-making processes.

What could be an accurate answer to how effectively we teach right from wrong to children or other individuals? We can all agree prisons are full of people who were found accountable after a bad decision was made. I like this particular statement cause also includes the number of individuals that have been erroneously convicted when judged. To verify that one only need to consult the innocence project records in the United States of America just to put an example. Where, as a result of bad decisions, some individuals have spent a mean time of 32 years in prison before their innocence was proved. An issue that is currently costing millions to the American government to amend.

Yet a large number of critics believe AI Systems has to be impeccable when learning from our biases (present in datasets), misconceptions (introduced by supervised learning to mention one example), and practices (limiting the machine capability to our human system context due we are usually afraid of what we cannot controlled/explain). Therefore, our way of teaching, learning, experiencing the way of distinguishing right from wrong as humans does not have to be the same for AI systems. Paralelly AI Systems are not the ones should be accountable for when a bad decision is made. We will be incurring in responsibility transference otherwise. Our perceptions and laws need to catch up with AI development.

An approach to tackle this problem could be to ideate how to work the data AI solutions will further use to learn and train a topic left aside most of the time. As in regard to data, the frequently tackled issues are related to duplicity, noise, absence, normalization, access, privacy, and governance; not bias. Complementarily, we also need to develop new methods tunned with the conception of AI solutions being machines and not human imitations; acknowledging we: developers, consumers… have the entire responsibility upon any outcome produced by an AI solution – except when such AI solution was designed and created by other AI solution 🙂 -. Otherwise, ideals such as fairness or justice will become even more subjective issues in a data-driven society.

Distinguishing right from wrong, we are the ones called to do that.

How to reference: Varona, Daniel. 2019. Right from wrong. www.danielvarona.ca [Year of Creation:2019] http://www.danielvarona.ca/2019/06/07/right-from-wrong/ [Consulted on: DDMMYY].