Technology has underpinned financial trading since the emergence of electronic markets in the 1970s. Today, however, the integration of AI into hyperspeed markets has created an environment where rapid, autonomous decision-making threatens to outpace critical oversight.

Given that AI has become so pervasive in capital markets trading, this article examines the key ethical debates emerging today and the future outlook.

Introducing AI-powered algo trading

Traditional algorithmic trading (algo trading) relied on rules-based financial mathematics to execute strategies, such as high-frequency trading and arbitrage. As algo trading moves toward AI automation, this market is undergoing an explosive shift toward fully automated, agentic AI models. The automated algorithmic trading market is projected to grow from US$24 billion in 2025 to US$44.55 billion by 2030.

This rapid expansion has led to an escalation in potential ethical and regulatory risk as systems evolve from basic generative tools into “agentic AI”: fully autonomous solutions that can observe, contemplate and execute complex trading or hedging strategies, sometimes completely independent of human intervention. Utilizing advanced machine learning techniques such as reinforcement learning and deep neural networks, these next-generation systems operate at a velocity and complexity that demand serious re-examination of systemic market risks.

Risk mitigation efforts

As AI agents have started to dominate the trading markets, several efforts are ongoing to mitigate the ethical risks.

Regulatory oversight

The most effective way to minimize socio-economic impacts and maximize ethical safeguards for AI use in algo trading is to establish a robust regulatory framework. Recent regulatory advancements are aimed not only at mitigating market instability and systemic risks but also at ensuring enhanced accountability and ethical AI use. The European Union Artificial Intelligence Act classifies high-risk AI systems and details requirements, including establishing risk management, technical documentation and record-keeping. The U.S. Securities and Exchange Commission’s exam priorities for 2026 include AI advancements, and the U.K. Financial Conduct Authority emphasized in its AI update that it aims to promote the safe and responsible use of AI in U.K. financial markets while ensuring beneficial innovation.

However, as Dr. Zissis Poulos, assistant professor of financial technologies at York University and AI researcher, explains, “It is really hard to employ regulatory guidelines for phenomena such as algo collusion, since we only have lab evidence and no actual examples or signs that such an agentic collusion is occurring in real time and how it could impact the trading markets.”

It is really hard to employ regulatory guidelines for phenomena such as algo collusion, since we only have lab evidence and no actual examples or signs that such an agentic collusion is occurring in real time and how it could impact the trading markets.

Human-in-the-loop to human-on-the-loop

Ensuring effective human oversight is critical to the successful deployment of AI models and risk mitigation. However, as trading technology has grown more sophisticated, the financial industry has quietly redefined what “oversight” means as it moves from an active, defensive posture to a passive one.

Initially, risk management was based on a “human-in-the-loop” framework, where an experienced human trader reviewed and approved every single decision made by the algorithm. Lately, as technological sophistication has become multifold, the oversight mechanism has evolved from human-in-the-loop to “human-on-the-loop,” where a human only reviews AI decisions when there are breaches or other extreme events. Human-on-the-loop adoption is slowly giving rise to “human + AI agent” business model, where humans are the business strategists/managers while AI agents work as an assistant.

This operational shift has exposed a vulnerability that recent Wharton research explains as “cognitive surrender.” This research reveals that highly integrated AI systems have begun to function as a “third cognitive system,” an external framework that sits outside the human brain but actively shapes human reasoning.

Given the technical constraints with reviewing every decision, human-on-the-loop seems to be the only plausible solution, but it needs improvement. For example, Poulos explains, “Currently researchers are aiming to build a self-awareness system in AI which would trigger human intervention based on the confidence it has in its own decisions.”

Transparency vs. machine efficiency

As trading techniques advance to rely heavily on deep learning and neural networks, machine efficiency has been achieved at the direct expense of transparency. This trade-off has created a critical problem in capital markets: autonomous systems can execute multi-million-dollar trading and hedging strategies based on mathematical relationships that are opaque, leaving even their own creators unable to trace the exact logic behind a specific market action.

This trade-off has created a critical problem in capital markets: autonomous systems can execute multi-million-dollar trading and hedging strategies based on mathematical relationships that are opaque, leaving even their own creators unable to trace the exact logic behind a specific market action.

As a solution, the industry is turning to explainable artificial intelligence (XAI) techniques, deploying tools such as SHapley Additive exPlanations (SHAP), Local Interpretable Model-agnostic Explanations (LIME) and ELI5. While these frameworks are intended to improve interpretability and provide a clear audit trail for complex decisions, they have inherent limitations when applied to hyperspeed trading environments.

Relying only on these tools should be treated with caution. If a model’s underlying logic cannot be audited in real time, true transparency does not exist. Additionally, the emerging practice of deploying secondary “AI guardians” or supervisory agents to watch over primary trading models introduces another layer of ethical risk.

The new ethical frontier

When advanced autonomous systems operate without transparency or human oversight, potential moral risks can quickly become systemic threats to human society. True ethical risks extend far beyond individual bad actors and manifest as structural market failures, including algorithmic collusion, the threat of computational oligopolies, data-driven biases and deep-seated socio-economic instability.

The most subtle threat is algorithmic collusion. In a traditional market, collusion requires human intent, communication and explicit programming. However, in an ecosystem dominated by independent agentic AI, multiple reinforcement learning models can independently deduce that the most optimal path to maximizing profit is to inadvertently cooperate with one another to manipulate prices. Because this collusion could happen implicitly through autonomous trial-and-error at microsecond speeds, it wouldn’t leave a traditional paper trail. This may create a silent, artificial oligopoly where a firm’s success is defined mostly by computational power rather than market fundamentals.

In the absence of restricting regulations, these dynamics could mean firms achieve capitalist objectives at the risk of market failures and socio-economic suffering. Hence, the outlook will be shaped not only by technological advancements, but also by how regulatory frameworks and societal norms evolve to contain these ethical risks while optimizing AI usage to increase market efficiency and reliability.

Attika Raj, CFA, is a seasoned quant finance professional currently working at TD. She has wide experience in quant research and modelling, investment management and financial advisory. Attika also actively volunteers with CFA Institute as CFA grader and Editorial Committee member for The Analyst.