13.05.2026

Capital Markets, AI, and Black-Box Trading: A Game Without Rules or a Referee

Capital markets, too, are undergoing major transformations driven by AI—even if this isn’t immediately apparent to the naked eye; complex data analysis and process automation are taking place here as well, with human intervention becoming increasingly less necessary. Both the labor market and intellectual property data (where patent applications in the AI field have grown exponentially in recent years) clearly signal this trend[1] .

Whether we are talking about optimizing investment decision-making processes or discovering new opportunities for capital market growth, the influence of AI is being felt, and the efficiency of financial markets can be dramatically enhanced in this way.

Although what AI brings to financial markets (increased productivity, greater precision in shaping investment frameworks and portfolios, improved forecasts of investment returns, more accurate quantification of risks) is, in principle, promising, optimism should nevertheless be matched by at least the same level of caution.

The most relevant example illustrating the need for caution regarding the use of algorithms is the 2010 flash crash, when,over the course of about 10 minutes, prices plummeted and the Dow Jones index fell by approximately 9%, only to recover a few minutes later. The market disruption apparently occurred due to the use of certain algorithms and the parameters they were set to, though the exact cause of the problem remains unclear to this day.

The 2010 incident is a textbook example of black-box trading, that is, a type of trading based on algorithms whose inner workings are unknown. These are automated trading systems that operate without human intervention and process enormous volumes of buy/sell orders at speeds measured in milliseconds. The most interesting aspect of these trading systems, however, stems from the fact that their operating methodology is generally secret; the logic behind the algorithm is not disclosed to the client (and sometimes cannot even be explained by the algorithm’s creator).

Why do these aspects matter, and not just for those with a particular interest in capital markets?

Such trading systems have not only been present in capital markets for some time, but currently dominate them in terms of trading volume compared to traditional trading mediated by human decisions.

Under such conditions, however, viewing black-box trading solely as an issue pertaining exclusively to hedge funds or investment firms would be imprudent, as the effects of such systems extend far beyond that. For example, pension funds or asset management firms (which are entities that are, so to speak, more "tangible” to the individual client, the average person) are institutions that, in turn, trade on the same markets as the systems mentioned above, competing, directly or indirectly, with the performance of black-box and high-frequency trading (HFT) algorithms, or perhaps even using them.

Practice has already shown that market manipulation through algorithmic strategies, particularly via HFT, works successfully (a fact that has, incidentally, also sparked litigation regarding the liability of those participants who employ such methods)[2] .

Traditionally, however, the concept of "market manipulation” was based, in turn, on the concept of "intent”: a market participant deliberately decides to distort prices.

But when it comes to executing trades through algorithms—which not only have capabilities far superior to those of humans (in terms of trading volume, speed, and the absence of emotion that might influence a trading decision), but also operate within a black-box trading system (i.e., completely non-transparent), how can we legally relate to the concept of "intent”? Implicitly, how can we verify the existence of market manipulation, with the consequence of being able to subsequently hold those responsible for the losses thus caused accountable?

Depending on the degree of autonomy with which it was designed, the algorithm executes orders or even makes autonomous decisions. Thus, with a "classic” program, the intent of the algorithm’s creator can be inferred based on the content of its code, but with a black-box AI system, such verification no longer works. It is true that we can know what the general objective was (for example, profit maximization), but how the AI reaches that objective—including whether it does so through market manipulation—may be impossible to understand, both before and after the transaction is executed (and this includes even by its creator). Consequently, legal tests based on intent become practically impossible to apply[3] .

Although real progress has recently been made in understanding black-box models—in the sense of researching them as "objects” observed from the outside (without addressing the method of direct interaction with them)—the applicability of these advances to the capital market is still far from becoming a reality. Since this new understanding is based on observing the model from the outside and identifying "triggers” that lead to a specific decision by the AI in question, the speed that characterizes the capital markets will make it extremely difficult to apply this method in this context as well. Reaction times on the order of milliseconds will not allow for the "external” observation of the black-box model’s behavior to be carried out in real time, which will effectively lead to a post-factum evaluation of the trading decisions made and the impossibility of observing, in this case, the "triggers” of the decision.  

Even in the absence of "intent” on the part of the trading algorithm, the effects of a transaction executed by AI in this system can be similar to those caused by market manipulation: artificial price drops or increases, abnormal trading volume, market shock (resulting in extreme, cascading volatility), as has already occurred in real-world cases.

In such a situation, the question of legal liability arises: will the algorithm’s programmer, the company using it, or the person who designed the trading strategy be held liable? Since the very process by which a black-box algorithm arrived at a certain result is completely opaque, determining the person on whom legal liability falls becomes a difficult goal to achieve. The EU AI Act, which appears to impose stricter regulatory requirements on providers of "high-risk” AI systems compared to MiFID II[4] , although it is moving in the right direction and attempting to establish comprehensive regulation in the field of AI, will nevertheless not be able to clearly answer the question "who is liable?” when an opaque algorithm carries out problematic transactions.

Given the extremely high reaction speed and the interconnectedness of markets—whether we’re talking about investment funds, publicly traded companies, or simply individual investors—it is absolutely essential that all these actors fully understand and accept exactly who they’re sharing the ring with. Although the game is already underway and the stakes are high, the referee is either late or completely absent. It remains to be seen, under these circumstances, who will be declared the winner.

An article by Ingrid-Amelia Apetrei, Managing Associate, STOICA & ASOCIAȚII -iapetrei@stoica-asociatii.ro.