Home Columns The Hamilton Mann Conversation: Digital For Good Artificial Integrity

Artificial Integrity

December 16, 2024

By Hamilton Mann

Much has been said about how to instill integrity into AI-generated content. So far, the focus has been on policing training data or filtering output, but wouldn’t it be better if the AI system itself had its own integrity at its core? Hamilton Mann puts the case.

Warren Buffet famously said, “In looking for people to hire, look for three qualities: integrity, intelligence, and energy. And if they don’t have the first, the other two will kill you.”

This principle equally applies to AI systems—because AI systems are not just tools.

So-called “tools” are deterministic; they follow a clear path from creation to usage, through degradation and, ultimately, obsolescence. In contrast, AI, especially machine learning systems, is fundamentally different. It does not follow this trajectory, because it is not static; it learns over time through interaction with data. Systems that use techniques like reinforcement learning or deep learning continuously refine themselves based on new input, making them more akin to dynamic entities that continuously evolve.

No two AI systems function identically if they are exposed to different data streams or used in varying contexts. This sets AI apart from traditional tools, which have deterministic functions that do not change from within. This indeterministic quality of AI makes it essential to not just be developed, but continuously led effectively and responsibly.

As AI systems increasingly take on critical roles across healthcare, education, transportation, finance, and public safety, having these systems only capable of mimicking a form of intelligence, relying on incredible computational energy and without any form of embedded integrity into their design, represents a major failure.

While AI can quickly process data, many don’t inherently consider whether their actions are ingrained with the perspective of doing the right thing.

They are like the engine and GPS of a car; the engine provides the power to get you anywhere quickly and efficiently, while the GPS intelligently calculates the fastest or most efficient route to your destination. The car can analyze road conditions, traffic, and distances, making real-time decisions to optimize the journey. However, intelligence and energy alone don’t consider whether the chosen path is safe, legal, ethical, moral, or socially acceptable; it just focuses on getting there efficiently, whatever “getting there” means.

Some AI systems have taken steps to reduce harmful biases in their responses by training them on diverse datasets and continuously fine-tuning them to avoid producing unethical outputs.

However, this is still an ongoing challenge. Even among the best image generation applications powered by GenAI, biases still persist, such as when these tools suggest image modifications that reflect stereotypical or sexist cultural clichés, which can offend certain populations and perpetuate discriminatory biases.

Some AI systems have taken steps to reduce harmful biases in their responses by training them on diverse datasets and continuously fine-tuning them to avoid producing unethical outputs.

Not to mention the near-perfect execution of imitating a person’s identity traits and characteristics made possible by certain systems, without any verification, prevention, or restriction, leading to what we call deepfakes, and the severe consequences such productions can have on individuals’ reputations, privacy, or safety, as well as broader societal harms such as misinformation, manipulation in politics, and fraud. This is no exception. The Global market for AI-generated deepfakes is expected to reach a value of US$79.1 million by the end of 2024, and it is further anticipated to reach a market value of $1,395.9 million by 2033 at a CAGR of 37.6 per cent.

We can assume that an AI system that we use in our daily life has been designed to align with broadly accepted ethical values. However, as its value system is shaped by its training data, it does not necessarily reflect cultural ethical norms. It does not “learn” ethical norms dynamically after deployment in the way a system with integrity might. It is updated periodically by its developers to improve its alignment with ethical values, but it does not adapt autonomously to changing ethical contexts. It lacks the autonomous reinforcement learning system where it could continuously learn and improve its behavior based on ethical intelligence, moral reasoning, and social intelligence without human intervention.

While some AI systems can explain certain processes or decisions, many AI systems cannot fully explain the decision-generating process (i.e., how they generate specific responses). Those based on machine learning, and even more so those based on more complex models like deep learning, are often opaque to users and operate as “black boxes.” While these systems may produce accurate results, users, those affected by the systems, and even developers often cannot fully explain how specific decisions or predictions are made. This lack of transparency can lead to several critical issues, particularly when they are used in sensitive areas such as healthcare, criminal justice, or finance.

Many AI systems used in daily life often lack the ability to autonomously adapt to evolving ethical contexts. These systems do not necessarily reflect the cultural or ethical norms of the diverse societies in which they operate. As a result, this creates potential misalignment with local values and societal expectations, leading to consequences such as rendering cultural aspects invisible or making ethically questionable decisions. While these AI systems may be periodically updated by developers to improve their alignment with ethical values, they still lack the capability to dynamically learn and adapt to new ethical standards in real time. This static approach to ethical adaptation leaves AI systems vulnerable to ethical lapses in fast-changing environments, especially when they are used in global and culturally diverse settings.

Some GenAI systems, such as ChatGPT, are designed to provide useful information, but true Artificial Integrity would involve a higher degree of consistency in ensuring that all information provided is reliable, verifiable with sources, and fully respects copyright of any kind, so as not to infringe on anyone’s intellectual property.

Responsibility in AI means nothing less than ensuring that AI systems operate with integrity over intelligence, prioritizing fairness, safeguarding human values, and upholding societal imperatives over raw intelligence.

Can AI demonstrate Artificial Integrity?

This goes beyond ethical guidelines. It represents a self-regulating quality embedded within the AI system itself. Artificial Integrity is about incorporating ethical principles into AI design to guide its functioning and outcomes, much like how human integrity guides behavior and impact even without external oversight, to mobilize intelligence for good.

It fills the critical gap that ethical guidelines alone cannot address by enabling several important shifts:

Shifting from inputs to outcomes:

AI ethical guidelines are typically rules, codes, or frameworks established by external entities such as governments, organizations, or oversight bodies. They are often imposed on AI systems from the outside as an input, requiring compliance without being an integral part of the system’s core functioning.
Artificial Integrity is an inherent, self-regulating quality embedded within the AI system itself. Rather than merely following externally imposed rules, an AI with integrity “understands” and automatically incorporates ethical principles into its decision-making processes. This internal compass ensures that the AI acts in line with ethical values even when external oversight is minimal or absent, maximizing the delivery of integrity-led outcomes.

Shifting from compliance to core functioning:

AI ethical guidelines focus on compliance and adherence. AI systems might meet these guidelines by following a checklist or performing certain actions when prompted. However, this compliance is often reactive and surface-level, requiring monitoring and enforcement.
Artificial Integrity represents a built-in core function within the AI. It operates proactively and continuously, guiding decisions based on ethical principles without needing to refer to a rule book. It’s similar to how human integrity guides someone to do the right thing even when no one is watching.

Shifting from fixed stances to contextual sensitivity:

AI ethical guidelines are often rigid and can struggle to account for nuanced or rapidly changing situations. They are typically designed for broad applicability and might not adapt well to every context an AI system encounters.
Artificial Integrity is adaptable and context-sensitive, allowing AI to apply ethical reasoning dynamically in real-time scenarios. An AI with integrity would weigh the ethical implications of different options in context, making decisions that align with core values rather than rigidly applying rules that may not fully address the situation’s complexity.

Shifting from reactive to proactive decision-making:

AI ethical guidelines are often applied reactively, after a potential issue or ethical violation is identified. They are used to correct behavior or prevent repeated errors. However, by the time these guidelines come into play, harm may have already occurred.
Artificial Integrity operates proactively, assessing potential risks and ethical dilemmas before they arise. Instead of merely avoiding punishable actions, an AI with integrity seeks to align every decision with ethical principles from the outset, minimizing the likelihood of harmful outcomes.

Shifting from enforcement to autonomy:

AI ethical guidelines require enforcement mechanisms, like audits, regulations, or penalties, to ensure that AI systems adhere to them. The AI doesn’t inherently prioritize these rules.
Artificial Integrity autonomously enforces its ethical standards. It doesn’t require external policing, because its ethical considerations are intrinsic to its decision-making architecture. This kind of system would, for example, refuse to act on commands that violate fundamental ethical principles, even without explicit human intervention.

This goes beyond AI guardrails.

If we continue our analogy with the car, the role of integrity systems does not just rely on the rules set by humans that others must comply with, such as the traffic code or the law.

In the context of a car, internal systems play a role in ensuring safe and responsible operation. Components such as the steering, braking, and stability control systems are designed to maintain the vehicle’s functionality and safety, even when human judgment or conditions falter. These systems don’t operate ethically in a human sense but are built to adhere to predetermined safety principles, ensuring that the car stays within its intended operational boundaries and minimizes risk.

In the context of AI, the mechanisms designed to ensure ethical, safe, and trustworthy AI outputs are commonly referred to as guardrails. These mechanisms, while foundational, exhibit limitations that highlight the need for a transformative shift towards an approach grounded in Artificial Integrity.

Current guardrails such as content filters, output optimizers, process orchestrators, and governance layers aim to identify, correct, and manage issues in AI outputs while ensuring compliance with ethical standards.

Content filters function by detecting offensive, biased, or harmful language, but they often rely on static, predefined rules that fail to adapt to complex or evolving contexts. Output optimizers address errors identified by filters, refining AI-generated responses, yet their reactive nature limits their ability to anticipate problems before they arise. Process orchestrators coordinate iterative interactions between filters and optimizers, ensuring that outputs meet thresholds, but these systems are resource-intensive and prone to delivering suboptimal results if corrections are capped. Governance layers provide oversight and logging, enabling accountability, but they depend heavily on initial ethical frameworks, which can be rigid and prone to bias, particularly in unanticipated scenarios.

By focusing on contextual understanding, AI systems with Artificial Integrity can make nuanced decisions that balance ethical considerations with operational goals, avoiding the pitfalls of rigid compliance models.

Despite their contributions, these guardrails expose critical gaps in the broader mission to create ethical AI systems. Their reactive design means they address problems only after they occur, rather than preventing them. They lack the contextual awareness necessary to navigate nuanced or situational ethics, which often leads to outputs that are ethically sound in isolation but problematic in context. They rely heavily on static, human-defined standards, which risks perpetuating systemic biases rather than challenging or correcting them. Furthermore, their iterative processes are computationally intensive, raising concerns about energy inefficiency and scalability in real-world applications.

The limitations of these mechanisms point to the need for a new paradigm that embeds integrity-led reasoning into the core of AI systems.

Artificial Integrity represents this shift by moving beyond the rule-based constraints of guardrails or the static-based constraints of ethical guidelines to systems capable of proactive ethical reasoning, contextual awareness, and dynamic adaptation to evolving societal norms.

Unlike existing AI systems, Artificial Integrity allows AI to anticipate ethical dilemmas and adapt its outputs to align with human values, even in complex or unforeseen situations. By focusing on contextual understanding, AI systems with Artificial Integrity can make nuanced decisions that balance ethical considerations with operational goals, avoiding the pitfalls of rigid compliance models.

Artificial Integrity also addresses the pervasive issue of bias by enabling systems to self-evaluate and refine their ethical frameworks based on continuous learning. This adaptability ensures that AI systems remain aligned with diverse user needs and societal expectations rather than reinforcing pre-existing inequalities.

By embedding these safeguards into the AI’s core logic, Artificial Integrity eliminates the inefficiencies of iterative guardrail processes, delivering outputs that are ethically sound and resource-efficient in real time.

The transition from guardrails and ethical guidelines to Artificial Integrity is not just an operational enhancement but a new AI frontier.

While current guardrails and ethical guidelines approaches provide essential protections, they fall short in addressing the complexities of AI’s societal impact. Artificial Integrity bridges this gap, creating systems that are not only intelligent but also inherently integrity-driven in alignment with human values.

This evolution is crucial to ensuring that AI systems contribute positively to society, reflecting the principles of fairness, accountability, and long-term ethical, moral, and social responsibility.

Without integrity embedded at the core, the risks and externalities posed by unchecked machine intelligence make them unsustainable, and render society even more vulnerable, even though they also bring positive aspects that coexist.

Integrity in AI is like the steering and braking systems of a car, which ensure that the vehicle, no matter its power, stays on the right path and avoids harmful, dangerous, or illegal situations.

While computational intelligence might suggest taking a shortcut down a one-way street to save time, integrity ensures that the AI follows the rules, just as a car must follow the rules of the road, and prioritizes safety over efficiency.

Integrity would keep the car from speeding through red lights or driving recklessly, even if it’s the fastest way. Integrity would ensure that the system makes ethical decisions, even if they are less efficient or less profitable, prioritizing fairness, safety, and the well-being of those affected.

The question is not how intelligent AI can become, whether it involves calls for super artificial intelligence or artificial general intelligence. No amount of intelligence can replace integrity.

The question is how we can ensure that AI exhibits Artificial Integrity—a built-in capacity to function with ethical intelligence, moral intelligence, and social intelligence, aligned with human values and guided by principles that prioritize fairness, safety, and societal considerations. In so doing, it exhibits a context-sensitive reasoning, both ex-ante (proactively) and ex-post (reflectively) as it learns from real-world interactions, ensuring that its outputs and outcomes are integrity-led first, and intelligent second.

This means that integrity-led steering mechanisms should be part of the code, the training processes, and the overall architecture of the AI, not just ethical guidelines on paper, websites, or in committee discussions. In this way, they become intrinsic to the functioning of the AI, rather than being applied separately or retroactively.

Without the capability to exhibit a form of integrity, AI would become a force whose evolution is inversely proportional to its necessary adherence to values and its crucial regard for human agency and well-being.

Just as it is not sheer engine power that grants autonomy to a car, nor to a plane, so it is not the mere increase of artificial intelligence that will guide the progress of AI that we need in order to foster a better future in society.

Why should organizations care?

Companies have long recognized that brand reputation and customer loyalty depend on an uncompromising integrity-driven social proof as a do-or-die imperative.

The entire history of business is filled with examples of integrity lapses that led “Achilles-type” companies to collapse, such as Enron, Lehman Brothers, WorldCom, Arthur Andersen, and, more recently, WeWork, Theranos, and FTX.

Yet, as businesses integrate AI into their operations, from customer service to marketing and decision-making, all eyes are fixed on the promise of productivity and efficiency gains, and many overlook a critical factor: the integrity of their AI systems’ outcomes.

What could be more irresponsible? Without this, companies face considerable risks, from regulatory scrutiny to legal repercussions, to brand reputation erosion, to potential collapse.

The rule in business has always been performance, but performance achieved at the cost of amoral behavior is neither profitable nor sustainable.

The excitement and rush toward AI is no excuse or tolerance for irresponsibility; it’s quite the opposite.

Relying on AI only makes responsible sense if the system is built with Artificial Integrity, ensuring it delivers performance while being fundamentally guided by integrity first—especially in outcomes that may, more often than we think, be life-altering.

To systematically address the challenges of Artificial Integrity, organizations can adopt a framework structured around three pillars: the Society Values Model, the AI Core Model, and the Human and AI Co-Intelligence Model.

Each of these pillars reinforces each other and focuses on different aspects of integrity, from AI conception to real-world application.

The Society Values Model revolves around the core values and integrity-led standards that an AI system is expected to uphold. This model demands that organizations start to consider doing the following:

Clearly define integrity principles that align with human rights, societal values, and sector-specific regulations to ensure that the AI’s operation is always responsible, fair, and sustainable.
Consider broader societal impacts, such as energy consumption and environmental sustainability, ensuring that AI systems are designed to operate efficiently and with minimal environmental footprint, while still maintaining integrity-led standards.
Embed these values into AI design by incorporating integrity principles into the AI’s objectives and decision-making logic, ensuring that the system reflects and upholds these values in all its operations while optimizing its behavior in prioritizing value alignment over performance.
Integrate autonomous auditing and self-monitoring mechanisms directly into the AI system, enabling real-time evaluation against integrity-led standards and automated generation of transparent reports that stakeholders can access to assess compliance, integrity, and sustainability.

This is about building the “Outer“ perspective of the AI systems.

The AI Core Model addresses the design of built-in mechanisms that ensure safety, explicability, and transparency, upholding the accountability of the systems and improving their ability to safeguard against misuse over time. Key components may include:

Implementing robust data governance frameworks that not only ensure data quality but also actively mitigate biases and ensure fairness across all training and operational phases of the AI system.
Designing explainable and interpretable AI models that allow stakeholders, both technical and non-technical, to understand the AI’s decision-making process, increasing trust and transparency.
Establishing built-in safety mechanisms that actively prevent harmful use or misuse, such as the generation of unsafe content, unethical decisions, or bias amplification. These mechanisms should operate autonomously, detecting potential risks and blocking harmful outputs in real time.
Creating adaptive learning frameworks where the AI is regularly retrained and updated to accommodate new data, address emerging integrity concerns, and continuously correct any biases or errors with regard to the value model that may occur over time.

This is about building the “Inner“ perspective of the AI systems.

The Human and AI Co-Intelligence Model emphasizes the symbiotic relationship between humans and AI, highlighting the need of AI systems to function considering the balance between “Human Value Added” and “AI Value Added”, where the synergy between human and technology redefines the core design of our society, while preserving societal integrity.

They would be able to function considering four distinct operating modes:

Marginal Mode

In the context of Artificial Integrity, Marginal Mode refers to situations where neither human input nor AI involvement adds meaningful value. These are tasks or processes that have become obsolete, overly routine, or inefficient to the point where they no longer contribute positively to an organization’s or society’s goals. In this mode, the priority is not about using AI to enhance human capabilities, but about identifying areas where both human and AI involvement has become useless.

One of the key roles of Artificial Integrity in Marginal Mode is the proactive detection of signals indicating when a process or task no longer contributes to the organization. For example, if a customer support system’s workload drastically decreases due to automation or improved self-service options, AI could recognize the diminishing need for human involvement in that area, helping the organization to take action to prepare the workforce for more value-driven work.

AI-First Mode

Here, AI’s strength in processing vast amounts of data with speed and accuracy takes precedence to the human contribution. Artificial Integrity would ensure that, even in these AI-dominated processes, integrity-led standards like fairness and cultural context are embedded.

When Artificial Integrity prevails, an AI system that analyzes patient data to identify health trends would be able to explain how it arrives at its conclusions (e.g., a recommendation for early cancer screening), ensuring transparency. The system would also be designed to avoid bias—for example, by ensuring that the model considers diverse populations, ensuring that conclusions drawn from predominantly one demographic group don’t lead to biased or unreliable medical advice.

Human-First Mode

This mode prioritizes human cognitive and emotional intelligence, with AI serving in a supportive role to assist human decision-making. Artificial Integrity ensures that AI systems here are designed to complement human judgment without overriding it, protecting humans from any form of interference with the healthy functioning of their cognition, such as avoiding influences that exploit vulnerabilities in our brain’s reward system, which can lead to addiction.

In legal settings, AI can assist judges by analyzing previous case law, but should not replace a judge’s moral and ethical reasoning. The AI system would need to ensure explainability, by showing how it arrived at its conclusions while adhering to cultural context and values that apply differently across regions or legal systems, while ensuring that human agency is not compromised regarding the decisions being made.

Fusion Mode

This is the mode where Artificial Integrity involves a synergy between human intelligence and AI capabilities, combining the best of both worlds.

In autonomous vehicles operating in Fusion Mode, AI would manage a vehicle’s operations, such as speed, navigation, and obstacle avoidance, while human oversight, potentially through emerging technologies like brain-computer interfaces (BCIs) would offer real-time input on complex ethical dilemmas. For instance, in unavoidable crash situations, a BCI could enable direct communication between the human brain and AI, allowing ethical decision-making to occur in real time, blending AI’s precision with human moral reasoning. These kinds of advanced integrations between human and machine will require Artificial Integrity at its highest level of maturity. Artificial Integrity would ensure not only technical excellence but also ethical, moral, and social soundness, guarding against the potential exploitation or manipulation of neural data and prioritizing the preservation of human safety, autonomy, and agency.

Finally, Artificial Integrity systems would be able to perform in each mode, while transitioning from one mode to another, depending on the situation, the need, and the context in which they operate.

Considering the Marginal Mode (where limited AI contribution and human intelligence is required—think of it as “less is more”), AI-First Mode (where AI takes precedence over human intelligence), Human-First Mode (where human intelligence takes precedence over AI), and Fusion Mode (where a synergy between human intelligence and AI is required), the model Human and AI Co-Intelligence ensures that:

Human oversight remains central in all critical decision-making processes, with AI serving to complement human intelligence rather than replace it, especially in areas where ethical judgment and accountability are paramount.
AI usage promotes responsible and integrity-driven behavior, ensuring that its deployment is aligned with both organizational and societal values, fostering an environment where AI systems contribute positively without causing harm.
AI usage establishes continuous feedback loops between human insights and AI learning, where these inform each other’s development. Human feedback enhances AI’s integrity-driven intelligence, while AI’s data-driven insights help refine human decision-making, leading to mutual improvement in performance and integrity-led outcomes.
AI systems are able to perform in each mode, while transitioning from one mode to another, depending on the situation, the need, and the context in which they operate.

Reinforced by the cohesive functioning of the two previous models, the Human and AI Co-Intelligence Model reflects the “Inter“ relations, dependencies, mediation, and connectedness between humans and AI systems.

This is the aim of Artificial Integrity.

Systems designed with this purpose will embody Artificial Integrity, emphasizing AI’s alignment with human-centered values.

This necessitates a holistic approach to AI development and deployment, considering not just AI’s capabilities but its impact on human and societal values. It’s about building AI systems that are not only intelligent but also understand the broader implications of their actions. Such a question is not just a technological one. With the interdisciplinary dimensions it implies, it is one of the most crucial leadership challenges.

Ultimately, the difference between intelligent-led and integrity-led machines is simple: the former are designed because we could, while the latter are designed because we should.

Concrete applications include:

Hiring and recruitment

Case: AI-powered hiring tools risk replicating biases if they are purely data-driven without considering fairness and inclusivity.
Artificial Integrity systems would proactively address potential biases (ex-ante) and evaluate the fairness of its outcomes (ex-post), making fair, inclusive hiring recommendations that respect diversity and equal opportunity values.

Ethical product recommendations and consumer protection

Case: E-commerce AI systems often push products based on profit margins or user profiles, potentially promoting unnecessary or harmful items to vulnerable customers.
Artificial Integrity systems would assess the suitability and ethics of recommended products, avoiding manipulation, and considering consumer well-being, particularly for vulnerable demographics.

Insurance claims processing risk assessment

Case: AI systems in insurance might prioritize cost-saving measures, potentially denying fair claims or overcharging based on demographic assumptions.
Artificial Integrity systems would consider the fairness of its risk assessments and claims decisions, adjusting for ethical standards and treating clients equitably, with ongoing ex-post analysis of claims outcomes to refine future assessments.

Supply chain ethical sourcing and sustainability

Case: AI systems in supply chain management may optimize costs but overlook ethical concerns around sourcing, labor practices, and environmental impact.
Artificial Integrity systems would prioritize suppliers that meet ethical labor standards and environmental sustainability criteria, even if they are not the lowest-cost option. It would conduct ex-ante ethical evaluations of sourcing decisions and track outcomes ex-post to assess long-term sustainability.

Content moderation and recommendation algorithms

Case: AI systems on social platforms often prioritize engagement, which can lead to the spread of misinformation or harmful content.
Artificial Integrity systems would prioritize user well-being and community safety over engagement metrics. They would preemptively filter content that could be harmful or misleading (ex-ante) and continually learn from flagged or removed content to improve their ethical filtering (ex-post).

Self-harm detection and prevention

Case: AI systems may encounter users expressing signs of distress or crisis, where insensitive or poorly chosen responses could exacerbate the situation. Some users may express thoughts or plans of self-harm in interactions with AI, where a standard system might lack the ability to recognize or appropriately escalate these red flags.
Artificial Integrity systems would be equipped to recognize such red-flag reactions, taking proactive steps to alert human supervisors or direct the user to crisis intervention resources, such as helplines or mental health professionals. Ex-post data reviews would be critical to improve the AI’s sensitivity in recognizing distress cues and responding safely.

Intelligence alone can easily stray off course, risking harm or unintended consequences.

Artificial Integrity over intelligence has become crucial from the moment AI systems, like Google DeepMind’s AlphaGo, demonstrated capabilities that far exceed human prediction or control.

As we stand on the verge of robotic intelligence (RI), the advance of AI systems capable of exhibiting integrity over intelligence is a critical question that will shape the course of human history.

Artificial Integrity represents the new AI frontier and a critical path to creating a better future for all.

Go to top