AI will never become ethical by following rules

Artificial intelligence has become the defining technology of the 21st century. Hardly a week passes without another breakthrough extending its reach into financial markets, healthcare, biomedical, and other scientific discovery, education, national security, and public administration. Alongside this rapid progress, an equally ambitious effort has emerged to construct ethical frameworks for intelligent machines. Governments and central banks draft regulations, corporations publish principles of responsible AI, researchers develop alignment techniques, and international organisations debate global governance. Fairness, transparency, explainability, accountability, and safety have become the new vocabulary of technological progress. 

Yet beneath this expanding architecture lies a deeper assumption that deserves scrutiny: that ethics itself can eventually be engineered if only rules, constraints, and governance mechanisms become sufficiently sophisticated. 

This assumption is fundamentally mistaken. 

The challenge is not how to automate ethics, but why ethical judgment cannot be reduced to automation in the first place. For decades, AI has advanced through optimisation. Modern systems detect statistical regularities with extraordinary precision, generating persuasive arguments, writing codes, and imitating increasingly sophisticated forms of reasoning. Their achievements are undeniable. Yet optimisation should never be confused with judgment, prediction should never be mistaken for wisdom, and compliance should never be mistaken for ethics. 

A machine that follows rules is not ethical; it is compliant. Of course there is nothing wrong to be compliant. A machine that imitates human behaviour might not be morality; it is reflective of human data, including biases, blind spots, and historical distortions. One path produces obedience, the other imitation, but neither produces ethical understanding. 

The distinction matters because ethics has never been reducible to rules. Human societies rely on principles, yet principles collide, contexts evolve, and unforeseen consequences emerge. No finite rule set can anticipate every future circumstance. Ethical judgment therefore begins with rules but cannot end with them. 

After every meaningful decision comes something equally important: reflection. Reflection is not regret or rule-based correction. It is the capacity to reassess whether the reasoning behind a decision remains valid once consequences become visible. In this sense, yesterday’s reflection becomes more important than yesterday’s data. Ethics is therefore not static compliance but a dynamic learning process. 

This idea has deep philosophical roots. Spinoza saw ethics as the progressive refinement of understanding rather than obedience. Socrates placed self-examination at the centre of wisdom. Popper showed that intellectual progress depends on continuous criticism of our own assumptions. Learning begins where certainty ends. Perhaps advanced machines will eventually confront the same structural limitation. In a conversation I once had with Daniel Kahneman in 2017 at Princeton, one question persisted; ‘can rational thinking itself be taught as a stable algorithm?’ Kahneman’s work showed that human judgment systematically deviates from rationality. But a deeper implication is that rationality itself is not fixed, it is a process of correction under uncertainty. If rationality is inherently adaptive, why should ethics, which is more complex, be reducible to computation? 

The obvious objection is that machines lack empathy, lived experience, and moral intuition. Early experiments with large language models confirm this: when optimisation objectives are poorly defined, systems can produce catastrophic or misleading outputs with complete confidence. I agree entirely. 

But this does not weaken the argument—it strengthens it. 

Current AI systems have nothing intrinsic upon which to reflect. They are simulation based optimisation engines, not moral agents. Simulations are engineering but not fundamentals. In order to test and validate simulations we must need theoretical robustness in place. Reflection, therefore, cannot reside in the machine itself. It must belong to the broader human-machine system, where oversight, institutional learning, scenario analysis, and feedback loops continuously reshape decisions over time. Through the quantitative finance filed, understanding about tomorrow’s financial market volatility can’t be well understood based on millions of simulations only we need better theory and models that tests inference (i.e: Kolmogorov). 

It shifts the question from artificial intelligence to institutional intelligence. Good governance has never been about rule enforcement alone. Financial regulators learnt long ago that compliance cannot eliminate systemic crises. When uncertainty dominates, risk management relies on stress testing, scenario analysis, and counterfactual reasoning. The goal is not perfect prediction, but resilience across plausible futures. AI governance demands the same philosophy. The objective is not perfect optimisation, but robust decision systems capable of learning from consequences. 

This becomes more urgent as AI systems no longer merely observe reality, they reshape it. Algorithms influence markets, healthcare, education, and (geo)political discourse. Their outputs modify human behaviour; human behaviour generates new data; and that data feeds back into the system. We are no longer deploying isolated tools, but recursive systems that continuously transform the environments they learn from. 

The central question is no longer whether models are accurate today, but whether the system as a whole converges toward better judgment over time. This is not an engineering problem alone, it is fundamentally philosophical. 

From a mathematical perspective, machine learning excels at optimisation, yet we still lack rigorous definitions of creativity (mathematical too), understanding, or knowledge. Gödel’s incompleteness theorem suggests that sufficiently rich systems cannot derive all truths from within their own axioms. Ethical reasoning shows a similar pattern: rule eventually encounters exceptions, objective function conflicts with competing values, and every formal system meets cases it cannot anticipate. Ethics, therefore, cannot be reduced to completeness of rules. 

Instead, ethical intelligence lies in recognising the limits of one’s own reasoning and continuously revising those limits through experience, evidence, and reflection. This is where contemporary AI ethics converges with practice. IBM’s Global Leader for Responsible AI, Francesca Rossi, argues that AI should remain a collaborator, not an accountable agent. Responsibility never transfers from human to machine. AI may assist reasoning, but it cannot bear responsibility. 

Consequentialist perspectives such as those discussed in The Economist recently (24 June 2026) highlight a limitation: ethical evaluation cannot stop at rule compliance. It must consider consequences unfolding in complex, evolving systems where outcomes often diverge from intentions. 

Together, these perspectives converge on a single insight: ethics cannot be reduced to rules or optimisation. It must be grounded in reflection on consequences within dynamic systems. 

But there is a deeper question still resonate, what is an AI, ontologically? 

Contemporary AI systems operate within an ontology of representation, not experience. They construct statistical models of language, behaviour, and knowledge, but they do not inhabit the world they describe. They possess no lived experience, no intentionality, and no responsibility for outcomes. Their “knowledge” is representational, not existential. 

This distinction is crucial. A system that represents moral reasoning is not capable of moral judgment. A system that generates empathy does not experience empathy. Intelligence may be approximated computationally, but ethical agency requires participation in a world of consequences. 

Ethics, therefore, does not reside in algorithms but in reflective human institutions capable of learning from the consequences of their decisions and continuously revising their assumptions. 

Artificial intelligence will undoubtedly become more capable. Whether it also becomes more reflective depends less on technology than on the institutions that govern it. 

The future of intelligence will not be defined by computational power alone, but by the ability of human-machine systems to converge toward better judgment under uncertainty through continuous reflection, resilient governance, and intellectual humility. Ethics is an asymptotic process of learning, revision, and adaptation. 

Perhaps this is the deepest lesson AI offers us today is the intelligence without reflection is simulated optimisation; reflection without learning is stagnation and only their continuous convergence offers the possibility but not of perfect ethics but of progressively wiser judgment. 

(The author is a mathematician-turned-tail risk hedging expert, Chief Risk Strategist and advisor to sovereign institutions, Board Member at RsRL, Chief of Risk at UIB Emirates, Chief at Stochastic Commodities, and co-theorist of the Delbaen–Majumdar Theory filtering AI bias. Views are personal.)