(Un)explainable AI: Should All AI Systems Be?


By Hamilton Mann

In the pursuit of harnessing the capabilities of Artificial Intelligence (AI), businesses and researchers grapple with paradoxes that emerge when aiming to achieve AI explainability. This article delves into six primary AI explainability paradoxes: Generalization vs. Particularization, Complexity vs. Simplicity, Overfitting vs. Adaptability, Engineering vs. Understandability, Computational Efficiency vs. Effectiveness, and oriented-learning vs. self-learning. These paradoxes highlight the inherent challenges of achieving both model accuracy and transparency, shedding light on the trade-offs between creating an accurate depiction of reality and providing a tool that is effective, understandable, and actionable. Using examples from healthcare, stock market predictions, natural language processing in customer service, autonomous vehicles, and e-commerce, the article elucidates the practical implications of these paradoxes from a value creation perspective. 

The article concludes by offering actionable recommendations for business leaders to navigate the complexities of AI transformation, emphasizing the significance of context, risk assessment, stakeholder education, ethical considerations, and the continuous evolution of AI techniques.

The capabilities of artificial intelligence are evolving at an unprecedented pace, simultaneously pushing the boundaries of our understanding. As AI systems become more sophisticated, the line between transparent, explainable processes and those concealed within a ‘black box’ becomes increasingly blurred. The call for “Explainable AI” (XAI) has grown louder, echoing through boardrooms, tech conferences, and research labs across the globe. 

Yet, as AI permeates various sectors, we must grapple with a complex, and perhaps even controversial, query: Should all AI systems be made explainable? This issue, though seemingly straightforward, is layered with nuance. As we navigate the intricate landscape of AI, it becomes evident that certain systems, due to their very distinct purpose, must be designed with a certain standard of explainability, while others might not necessitate such transparency.

Referring to methods and tools that make the decision-making process of AI systems clear and interpretable to human users, the explainable AI idea is simple: if an AI system “makes” a decision, humans should be able to understand how and why that decision was made.

In healthcare, some AI models used to detect skin cancers or lesions provide visual heatmaps alongside their diagnoses. These heatmaps highlight the specific areas of the skin image that the model found indicative of malignancy, allowing dermatologists to understand the AI’s focus and reasoning.

By providing a visual representation of areas of concern, the AI system allows healthcare professionals to “see” what the model is detecting. This not only adds a layer of trust but also enables a doctor to cross-reference the AI’s findings with their own expertise.

In homeland security, some Security agencies use AI to scan surveillance footage and identify potentially suspicious activities. Explainable systems in this domain will provide reasoning by tagging specific actions (like an unattended bag) or behaviors (a person frequently looking over their shoulder) as indicators, rather than just flagging an individual without context.

By tagging and detailing specific actions or behaviors that are considered suspicious, the AI system offers insights into its decision-making process. This not only aids security personnel in quick decision-making but also helps in refining and training the AI system further.

In the legal domain, AI systems have been developed to analyze and review legal contracts and documents. One such tool, ThoughtRiver, scans and interprets information from written contracts used in commercial risk assessments.

The call for “Explainable AI” (XAI) has grown louder, echoing through boardrooms, tech conferences, and research labs across the globe.

As it reviews documents, ThoughtRiver provides users with an explanation for its analyses. For example, if it flags a particular clause as potentially problematic, it will explain why, referencing the specific legal standards or precedents that are pertinent. This not only accelerates the document review process but also provides lawyers with a clear understanding of the potential risks identified by the AI. 

The fact that an AI system can be explainable allows society to have confidence in the decisions that the system helps make. It’s a guarantee of control over the influence that AI can have in our societies.

Conversely, when an AI system’s decision-making process is opaque or not easily interpretable by humans, it is often classified as “black-box” AI. Such systems, despite their efficacy, might not readily offer insights into their internal workings or the rationale behind their conclusions. 

In Healthcare, Deep learning models have been used in hospitals to predict sudden deteriorations in patient health, such as sepsis or heart failure. These models can analyze vast amounts of patient data—from vital signs to lab results—and alert doctors to potential problems.

These technological advancements truly have the potential to save lives. However, this magic has its secrets that might elude human understanding.

While these models have proven to be efficient, the exact pathways and combinations of data points they use to arrive at their conclusions are often complex and not immediately clear to clinicians. This “black-box” nature can make it challenging for doctors to fully trust the model’s predictions without understanding its reasoning, especially in life-or-death situations.

Advanced AI systems are deployed in surveillance cameras in airports, stadiums, and other large public venues to detect potential security threats based on facial recognition, behavioral patterns, and more.

While the value of such systems offers real benefits for the safety of individuals and critical infrastructures essential to a country, it must also be recognized that understanding the decisions issued by the system can appear complex for humans to justify.

These systems process vast amounts of data at incredible speeds to identify potential threats. While they can flag an individual or situation as suspicious, the intricate web of reasoning behind such a decision—combining facial recognition, movement patterns, and possibly even biometric data—can be difficult to fully articulate or understand.

Some jurisdictions, in the US and China in particular, have started using AI systems to aid in determining the risk associated with granting bail or parole to individuals. These models analyze numerous factors, including past behavior, family history, and more, to generate a risk score.

While the goal of protecting populations could make such systems a real asset, they remain dangerous because the reasoning leading to the decision cannot be reconstructed by humans.

The decision-making process of these systems is multifaceted, taking into account a wide variety of variables. While they provide a risk score, detailing the exact weightage or significance attributed to each factor, or how they interplay, can be elusive. This lack of clarity can be problematic, especially when dealing with individuals’ liberties and rights. 

So, the question arises: why not simply make sure that all AI systems are explainable?


The question of regulating artificial intelligence, particularly in terms of explainability, is gaining attention from policymakers worldwide. China’s Cyberspace Administration (CAC) has released its “Interim Measures for the Management of Generative Artificial Intelligence Services,” addressing issues like transparency and discrimination. In contrast, the United States has currently a less prescriptive approach U.S. The country’s regulatory framework is largely based on voluntary guidelines like the NIST AI Risk Management Framework and self-regulation by the industry. For instance, Federal Agencies like the Federal Trade Commission (FTC) are already regulating AI within their scope, enforcing statutes like the Fair Credit Reporting Act and the Equal Credit Opportunity Act. In Europe, the General Data Protection Regulation (GDPR) mandates a “right to explanation” for automated decisions, a principle further reinforced by the European Union’s recently proposed Artificial Intelligence Act (AIA), which aims to provide a comprehensive framework for the ethical and responsible use of AI. As it stands, although many regulations are still works in progress or newly implemented, a complex, patchwork regulatory landscape is emerging, with different countries focusing on elements like accountability, transparency, and fairness. 

The implications are twofold: on the one hand, organizations have and will have to navigate an increasingly complex set of rules, and on the other, these regulations might actually foster innovation in the field of explainable AI, as this is a ground of multifaceted constraints.

As a matter of fact, we are faced with a series of paradoxes, which are nothing: that of performance, exemplified here by application cases of predictions that challenge our predictive framework, opposing perfection, illustrated here by our need to understand and control how AI formulates and concludes to certain predictions or decisions.

This trade-off between model explainability and performance arises from the intrinsic characteristics of different machine learning models and the complexities inherent in data representation and decision-making.

In addressing the challenge of explainable AI, we can identify six core paradoxes:

First, there is the Complexity vs Simplicity paradox. 

More complex models, like deep neural networks, can capture intricate relationships and nuances in data that simpler models might miss. 

As a result, complex models can often achieve higher accuracy. 

However, their intricate nature makes them harder to interpret. On the other hand, simpler models like linear regression or decision trees are easier to understand but might not capture all the subtleties in the data. 

In the realm of medical diagnostics, the Complexity vs Simplicity Paradox manifests in a notable way. While complex deep learning models can predict diseases like cancer with high accuracy by identifying intricate patterns in MRI or X-ray images, traditional algorithms rely on simpler features such as tumor size or location. Though these complex models offer superior diagnostic capabilities, their “black box” nature poses a challenge. Healthcare providers find it difficult to understand the model’s decisions, a critical factor in medical treatments that often require clear human understanding and explanation.

Innovators and data scientists are at the forefront of creating value by developing sophisticated algorithms that harness the power of vast datasets, yielding potentially life-saving diagnostic capabilities.

Within this framework, value is created and destroyed at multiple junctures. Innovators and data scientists are at the forefront of creating value by developing sophisticated algorithms that harness the power of vast datasets, yielding potentially life-saving diagnostic capabilities. This innovation benefits patients by providing them with more accurate diagnoses, which can lead to more effective treatments. However, this value creation is balanced by the potential destruction or stifling of trust in the medical realm. When healthcare providers cannot comprehend or explain the decision-making process of a diagnostic tool, they might be hesitant to rely on it fully, depriving patients of the full benefits of technological advancements. Additionally, this lack of transparency can lead to skepticism from patients, who might find it difficult to trust a diagnosis derived from an enigmatic process. Thus, while data scientists create value through advanced model development, that value is simultaneously at risk of being diminished if these tools cannot be understood or explained by the medical community serving the patients. 

Second, there is the Generalization vs. Particularization paradox.

Models that are highly interpretable, such as linear regression or shallow decision trees, make decisions based on clear and general rules. But these general rules might not always capture specific or intricate patterns in data, leading to potentially lower performance. Complex models, on the other hand, can identify and use these intricate patterns but do so in ways that are harder to interpret. 

The Generalization vs. Particularization Paradox is vividly evident in the field of credit scoring. General models typically employ simple, overarching criteria such as income, age, and employment status to determine creditworthiness. On the other hand, particular models delve into more nuanced data, including spending habits and social connections. Although particular models may yield more accurate predictions, they introduce challenges for consumers who struggle to understand the rationale behind their credit scores. This opacity can raise serious concerns about fairness and transparency in credit assessments. 

In this dynamic, value is both generated and potentially compromised by the tug-of-war between general and particular modeling approaches. Financial institutions and lenders stand to gain immensely from particular models; these models’ refined accuracy enables them to better assess the risk associated with lending, potentially reducing financial losses and optimizing profits. For consumers, an accurate credit assessment based on intricate patterns could mean more tailored financial products and potentially lower interest rates for those who are deemed low risk. However, the value creation comes at a cost. The very nuance that grants these models their accuracy also shrouds them in a veil of mystery for the average consumer. When individuals can’t ascertain why their credit scores are affected in a certain way, it can erode their trust in the lending system. This mistrust can further alienate potential borrowers and diminish their engagement with financial institutions. Thus, while financial technologists and institutions might create value through precision, this can simultaneously be undercut if the end consumers feel disenfranchised or unfairly judged by incomprehensible algorithms.

Third, there is the Overfitting vs Adaptability paradox.

Highly complex models can sometimes “memorize” the training data (overfit), capturing noise rather than the underlying data distribution. While this can lead to high accuracy on training data, it often results in poor generalization to new, unseen data. Even though simpler, more interpretable models might not achieve as high accuracy on the training set, they can be more robust and generalizable.

The Overfitting vs Adaptability Paradox is particularly noticeable within the scope of stock market prediction. Complex models may excel at “memorizing” past market trends, but often falter when applied to new, unseen data. In contrast, simpler models are less prone to overfitting and tend to be more adaptable to market changes, although they might not capture more complex relationships in the data. However, overfit models can lead investors astray, causing them to make poor financial decisions based on predictions that don’t hold up over time. 

In the intricate world of stock market prediction, the creation and possible erosion of value intertwine at the nexus of this Overfitting vs Adaptability paradox. On the creation side, financial analysts and quantitative researchers work tirelessly to devise algorithms aiming to unearth market trends and anomalies, aspiring to provide investors an edge in their investment strategies. When these algorithms are aptly balanced, investors stand to gain significantly, reaping the benefits of well-informed decisions that translate to lucrative returns. However, the precarious terrain of overfitting, where models are seduced by the idiosyncrasies of past data, puts this value at risk. Overreliance on these overfit models can mislead even the most seasoned investors into making suboptimal investment choices, leading to substantial financial losses. In such scenarios, not only is monetary value destroyed for the investor, but the credibility of quantitative models and the researchers behind them risks being undermined. It’s a stark reminder that in the realm of financial predictions, the allure of complexity must be weighed carefully against the timeless virtues of simplicity and adaptability. 

Fourth, there is the Engineering vs Understandability paradox.

For simpler models to achieve high performance, substantial feature engineering might be necessary. This involves manually creating new features from the data based on domain knowledge. The engineered features can make the model perform better but can also make the model’s decisions harder to interpret if the transformations are not intuitive. 

In customer service applications using natural language processing, the Engineering vs Understandability Paradox comes into play. Feature engineering techniques can be applied to process text into numerous features like sentiment and context, which improves model performance. However, while this can enhance performance, it can also make the decision-making process opaquer. This can pose challenges for managers trying to understand how the model is categorizing customer complaints or inquiries.

In the nuanced arena of customer service applications powered by natural language processing, the balance between crafting high-performing models and maintaining their transparency becomes a delicate dance of value creation and potential erosion. Here, data scientists and NLP experts create immense value by leveraging their domain knowledge to engineer features, aiming to refine a model’s ability to discern customer sentiment, context, and intent. This refined discernment can lead to more tailored and effective responses, resulting in enhanced customer satisfaction and trust. But therein lies the double-edged sword: while businesses and their customers stand to benefit from more accurate and responsive AI-powered systems, the increasingly intricate engineering can obscure a model’s rationale. For team leaders and managers overseeing customer service, this murkiness complicates their ability to intervene, train, or even explain a model’s decisions. Such lack of clarity can lead to misalignments in strategy and potential missteps in customer interactions. Thus, while the technical prowess of data scientists lays the groundwork for enhanced customer experiences, the resulting complexity threatens to diminish the trust and actionable insights that teams require to function effectively.

Fifth, there is the Computational Efficiency vs Effectiveness paradox.

Simpler, interpretable models often require less computational power and memory, making them more efficient for deployment. In contrast, highly complex models might perform better but could be computationally expensive to train and deploy.

Complex models in autonomous vehicles enable better real-time decision-making but come at the cost of requiring significant computational power. On the other hand, simple models are easier to deploy but might struggle with handling road anomalies effectively. A balance must be struck between computational efficiency and the safety of the vehicle and its passengers. 

In the rapidly evolving world of autonomous vehicles, the interplay between computational demands and real-world effectiveness carves out a pathway for both profound value creation and potential risks. Passengers and road users stand to benefit from vehicles that can respond adeptly to a myriad of driving conditions, promising safer and more efficient journeys. Yet, this promise carries a price. The more intricate the model, the more it leans on computational resources, leading to challenges in real-time responsiveness and potentially higher vehicle costs. Moreover, the reliance on overly simplistic models to save on computational power can lead to oversights when the vehicle encounters unexpected road scenarios, risking the safety of passengers and other road users. As such, while the technological advancements in autonomous vehicles present a horizon filled with potential, the equilibrium between efficiency and effectiveness becomes pivotal, ensuring that value is neither compromised nor squandered in the quest for progress.

Sixth, there is the oriented-learning vs self-learning paradox.

Some techniques that make models more interpretable involve adding constraints or regularization to the learning process. For instance, “sparsity” constraints can make only a subset of features influential, making the model’s decision process clearer. However, this constraint can sometimes reduce the model’s capacity to learn from all available information, thus potentially reducing its performance.

Oriented-learning models in recommender systems often focus on specific rules or criteria such as user history, making them easier to understand but potentially less effective. Self-learning models, in contrast, adapt over time and consider a wider variety of data points, possibly surprising users with how well the system seems to “know” them. In eCommerce, the real-world implication suggests that while understanding why a recommendation was made might be less critical than in healthcare, there are still concerns around privacy and effectiveness.

If a system knows too much, it risks alienating users who feel their data is being overly exploited. 

In the intricate tapestry of eCommerce, the duality between oriented-learning and self-learning mechanisms delineates a realm where value and potential pitfalls intersect. eCommerce giants and data scientists invest heavily in developing sophisticated recommender systems, with the aim of tailoring user experiences and fostering customer loyalty. For the consumer, this can mean a more seamless shopping experience, where product recommendations align closely with their preferences and past behaviors. The immediate value here is twofold: businesses see higher sales and consumers enjoy more relevant content. However, the balance is delicate. Oriented-learning models, while easier to explain and understand, might at times feel too restrictive or predictable, possibly missing out on suggesting a wider variety of products that users might find appealing. On the flip side, the allure of self-learning models, with their uncanny knack for personalization, raises eyebrows on privacy concerns. If a system knows too much, it risks alienating users who feel their data is being overly exploited. 

Herein lies the paradox’s crux: in the endeavor to create a perfect shopping experience, the very tools designed to enhance user engagement could inadvertently erode trust and comfort, starving the relationship between consumer and platform of its inherent value. 

All these paradoxes, which means trade-offs to be made, do exist because the characteristics that make models interpretable (simplicity, clear decision boundaries, reliance on fewer features) can also limit their capacity to capture and utilize all available information in the data fully. On the other hand, models that utilize all data intricacies for decision-making do so in ways that are harder to articulate and understand.

The balance or tension between achieving a precise, accurate depiction of reality and having a practical, effective tool for understanding, prediction, and intervention is a recurring theme. 

Philosophers like Nancy Cartwright have discussed how scientific models work. Models are often idealized, simplified representations of reality, sacrificing precision for tractability and explanatory power. These models might not be fully “true” or precise, but they can be extremely effective in understanding and predicting phenomena.  

How should business leaders manage these paradoxes in their AI transformation?

Here are some recommendations for tackling the challenges posed by the six specific paradoxes outlined. 

  • Recognize the Importance of Context while acknowledging the audience (Generalization vs. Particularization): Understand that not all AI applications require the same degree of explainability, and not all explanations are equally interpretable depending on the audience. For example, AI used in healthcare diagnoses may demand higher transparency than AI used for movie recommendations. 
  • Risk vs. Reward (Complexity vs Simplicity): Analyze the potential risk associated with AI decision-making. If an incorrect decision could lead to significant harm or costs (e.g., in healthcare or legal decisions), prioritize explainability even if it sacrifices some performance. 
  • Embrace Appropriate Complexity (Complexity vs Simplicity): When developing or purchasing AI systems, make deliberate choices about complexity based on goals. If the goal is to capture intricate data patterns, a more complex model might be suitable. But always ensure that the decision-makers who use the AI outputs understand the model’s inherent limitations in terms of interpretability. 
  • Ensure Robustness over High Training Accuracy (Overfitting vs Adaptability): Always assess and monitor the AI model’s performance on unseen or new data. While complex models might achieve impressive results on training data, their adaptability to fresh data is paramount, guarding against overfitting. 
  • Feature Engineering with Interpretability in Mind, not as an afterthought. (Engineering vs Understandability): If your AI application requires feature engineering, ensure that those features are interpretable and meaningful in the domain context and don’t add unnecessary opacity, addressing the balance between enhancing performance and maintaining clarity. While these can enhance performance, they shouldn’t compromise understandability.
  • Efficient Deployment (Computational Efficiency vs Effectiveness): When deploying AI models, especially in real-time scenarios, weigh the benefits of model simplicity and computational ease against the potential performance gains of a more complex, computationally intensive model. Often, a simpler model might suffice, especially if computational resources are a constraint.
  • Steer Model Learning for Clarity (oriented-learning vs self-learning): When transparency is vital, for AI applications where interpretability is crucial, consider guiding the model’s learning through constraints or regularization. This may reduce performance slightly, but it’ll enhance the model’s clarity and decision-making process.
  • Educate Stakeholders on Model Nuances (Generalization vs. Particularization): Regularly train stakeholders who will interact with or rely on the AI system’s general rules and specific intricacies, ensuring they’re well-versed with its capabilities and limitations, and potential biases. The incorporation of expertise from disciplines such as psychology, sociology, and philosophy can provide novel perspectives on interpretability and ethical considerations. Human-centered design thinking can guide the development of AI systems that are both more interpretable and more acceptable.
  • Embrace a Hybrid Approach (Engineering vs Understandability): Merge machine and human decision-making. While AI can offer rapid data processing and nuanced insights due to feature engineering, human expertise can provide the necessary context and interpretability ensuring clarity where the AI might be less transparent.
  • Prioritize Feedback Loops (Overfitting vs Adaptability): Especially in critical domains, ensure that there are feedback mechanisms in place. If an AI system makes a recommendation or prediction, human experts should have the final say, and their decisions should be looped back to refine the AI model. 
  • Uphold Transparency and Documentation (Complexity vs Simplicity): Maintain clear documentation about the design choices, data sources, and potential biases of the AI system. This documentation will be crucial for both internal audits and external scrutiny. This practice aids in navigating the complexity of AI systems by providing a simpler, more transparent layer for review.
  • Protect Individual Rights (oriented-learning vs self-learning): Especially in sectors like law enforcement or any domain dealing with individual rights, ensure that the lack of full explainability does not infringe upon individuals’ rights, for instance due to the AI system leaning heavily towards certain data features or constraints, overlooking the bigger picture. Decisions should never be solely based on “black-box” AI outputs. 
  • Define an Ethical Framework (oriented-learning vs self-learning): Leaders should establish an ethical framework and governance model that set the parameters and ethical standards for the development and operation of AI systems. This should cover aspects like data privacy, fairness, accountability, and transparency. Data ethics committees can be useful in this regard. Businesses have to be cognizant of the evolving landscape of AI-related regulations. Being proactive in this aspect not only mitigates risk but also could serve as a competitive advantage. 
  • Stay Updated and Iterative (Computational Efficiency vs Effectiveness): The field of AI, especially XAI (Explainable AI), is rapidly evolving. Stay updated with the latest techniques, tools, and best practices. Regularly revisit and refine AI deployments to ensure they meet the evolving standards and needs while ensuring models remain computationally efficient. This includes re-evaluating and adjusting models as new data becomes available or as societal norms and regulations evolve.

In conclusion, the goal is not to swing entirely towards complete explainability at the expense of performance, or complete performance at the expense of explainability. It is about finding a balanced approach tailored to each AI application’s unique risks and rewards, taking into account the human and environmental implications that are inextricably intertwined with the purpose of building trust.

About the Author

Hamilton MannHamilton Mann is the Group VP of Digital Marketing and Digital Transformation at Thales. He is also the President of the Digital Transformation Club of INSEAD Alumni Association France (IAAF), a mentor at the MIT Priscilla King Gray (PKG) Center, and Senior Lecturer at INSEAD, HEC and EDHEC Business School.

The views expressed in this article are those of the authors and do not necessarily reflect the views or policies of The World Financial Review.