Markets have always been more than machines for setting prices. They are also systems for deciding what to pay attention to, whose claims to trust, and how quickly uncertainty can be reduced. That is the core warning behind recent commentary on how artificial intelligence could make markets worse: not because AI is inherently malicious, but because it changes the information economics of trading, investing, and even everyday market-like behavior.
Information economics starts from a simple but uncomfortable premise. In many real markets, participants do not merely “observe” the world and react. They interpret signals that are incomplete, delayed, noisy, or strategically produced. Prices then become only one part of the story. Markets also allocate attention, coordinate beliefs, and discipline misinformation—sometimes imperfectly, but with real consequences for welfare. When AI alters how information is generated, filtered, packaged, and acted upon, it can shift those mechanics in ways that are not automatically efficiency-enhancing.
The debate often frames AI as a productivity upgrade: faster computation, better pattern recognition, more data processed per second. But the more consequential question is different. What happens when AI becomes the intermediary between reality and the market’s collective understanding? If AI changes the credibility of signals, the speed of feedback, or the incentives around producing information, then the market may become less accurate, less stable, or less fair—even if it becomes more “efficient” in a narrow technical sense.
To see why, it helps to think about markets as information-processing networks. In such networks, three things matter: signal quality, the structure of incentives, and the dynamics of learning and reaction. AI can improve each of these—or it can distort them. The risk is that we will notice the benefits first (better forecasts, tighter spreads, quicker execution) while underestimating the longer-run costs (misallocation, fragility, and a widening gap between what is visible and what is true).
A faster signal pipeline can still deliver worse information
One of the most intuitive ways AI could worsen markets is by accelerating the production of signals. Modern AI systems can generate recommendations, predictions, and classifications at scale. In principle, that should help. In practice, it can also increase the volume of low-quality or misleading information.
In information economics, noise is not just an annoyance; it changes behavior. When participants face more signals, they must decide which ones to trust. If AI-driven signals are abundant and cheap, the market may respond by lowering its standards for verification. That can create a subtle degradation loop: more output leads to more consumption, which leads to less scrutiny, which leads to more output that is not grounded in durable information.
This is especially relevant in environments where the “signal” is not a direct measurement of fundamentals but an interpretation of them. Consider earnings narratives, creditworthiness assessments, fraud risk scores, or even sentiment indicators. If AI systems produce these interpretations faster than humans can evaluate them, the market may treat them as if they were facts. Yet the underlying model may be wrong in systematic ways—wrong not randomly, but in ways that correlate across firms, sectors, or time periods.
The result can be a market that reacts quickly to what looks like information but is actually model artifacts. Prices may move, but the movement may not reflect improved knowledge. It reflects improved signaling throughput.
Speed is not neutral when feedback loops exist
AI also changes the tempo of trading and decision-making. Higher speed can be beneficial when it reduces latency between new information and price discovery. But speed becomes dangerous when it amplifies short-lived signals and creates feedback loops.
Feedback loops occur when market participants’ actions influence the very signals they later observe. In a world without AI, feedback loops exist too—liquidity dries up, momentum strategies reinforce trends, and narratives spread. AI can intensify these dynamics by making reactions more immediate and more synchronized.
Imagine a scenario where an AI system detects a pattern in order flow or social media activity and triggers trades. Those trades then alter order flow and social media visibility, which the next round of AI systems interprets as confirmation. Even if each individual model is “reasonable,” the combined system can behave like a self-fulfilling prophecy. The market becomes a closed loop: models react to outputs that were partly created by earlier model reactions.
This is where information economics offers a particularly sharp lens. Markets are not only about discovering truth; they are about coordinating around beliefs under uncertainty. When coordination becomes too fast, it can lock in incorrect beliefs before enough counter-evidence arrives. The market may then overshoot—moving too far, too quickly—because the information processing system has reduced the time available for skepticism.
The danger is not that AI makes everyone irrational. The danger is that AI makes everyone similar. When many participants deploy comparable architectures, training regimes, or data sources, their errors can become correlated. Correlated errors are the enemy of stability.
Asymmetric advantages can distort the allocation of attention
Another way AI could make markets worse is through asymmetric data advantages. Information economics emphasizes that markets are shaped by who knows what, when they know it, and how costly it is to verify. If some participants deploy superior AI systems—better at extracting signals from messy data, better at predicting outcomes, better at detecting manipulation—then the distribution of informational advantage widens.
That might sound like a standard competitive story: better tools win. But the welfare implications depend on how the advantage affects the rest of the market. If AI-enabled players can extract value by front-running or outpacing others, the market may become less attractive to slower participants. Liquidity can thin. Bid-ask spreads can widen. Or, more subtly, the market may shift toward strategies that exploit informational asymmetries rather than strategies that improve price discovery.
There is also an attention dimension. Markets do not just allocate capital; they allocate attention. AI systems can dominate what gets surfaced—what headlines appear, what recommendations are boosted, what risks are flagged. If the attention layer becomes concentrated among a few AI-driven intermediaries, then the market’s “information” becomes partially determined by the incentives of those intermediaries.
In other words, the market may not just be reacting to information. It may be reacting to the curation of information.
When what’s visible diverges from what’s accurate
Perhaps the most distinctive risk in the AI era is the divergence between visibility and accuracy. Markets rely on proxies: rankings, feeds, headlines, and aggregated metrics. These proxies are useful because they compress complex information into something tradable. But proxies can also become targets.
AI can turn proxies into self-reinforcing objects. If models learn that certain visible signals correlate with profits, then participants have incentives to produce those signals—sometimes by improving fundamentals, but sometimes by gaming the proxy itself. The market then becomes a contest over what is measurable and model-friendly, not necessarily what is true.
This is not a new problem, but AI can make it more intense. With AI, it becomes easier to generate content that matches the statistical patterns models reward. It becomes easier to simulate credibility. It becomes easier to automate persuasion. And it becomes easier to scale both legitimate and illegitimate information production.
The information economics lesson is that markets can fail when the cost of producing misleading information falls faster than the cost of verifying it. AI lowers production costs dramatically. Verification costs may not fall at the same rate, especially when verification requires human judgment, slow audits, or access to proprietary data.
So the market can end up with a thick layer of “information” that is highly visible but not reliably informative.
Model-driven markets and the problem of correlated belief
A deeper concern is that AI can change not only the content of information but the structure of belief formation. Many AI systems are trained to optimize predictive accuracy on historical data. But markets are adaptive. When participants use models to trade, their actions change the environment that future models learn from.
This creates a moving target. If many models are trained on similar data and optimized for similar objectives, they may converge on similar strategies. That convergence can reduce diversity in belief. Diversity matters because it provides friction against consensus errors. Without it, the market can become brittle.
Information economics often highlights that uncertainty is not just a lack of data; it is a strategic condition. Participants may have reasons to reveal or conceal information. They may have reasons to invest in verification. When AI changes the verification landscape—by automating detection, by generating explanations, by producing confidence scores—it can also change strategic behavior.
Confidence scores are particularly tricky. A model that outputs a probability can make uncertainty look quantified and therefore actionable. But calibrated probabilities are hard to guarantee, especially under distribution shift. If market participants treat model confidence as a substitute for verification, they may overweight signals that are systematically miscalibrated.
The market then coordinates around a false sense of precision.
Fraud detection and compliance: better tools, new attack surfaces
AI is already used in fraud detection, compliance monitoring, and risk scoring. These systems can improve detection rates and reduce manual workload. Yet they also create new attack surfaces.
When detection becomes automated, adversaries can adapt. They can probe model boundaries, exploit blind spots, and generate synthetic behavior designed to pass filters. This is a cat-and-mouse dynamic, but AI accelerates both sides. The result can be a market environment where compliance systems become reactive rather than preventive.
From an information economics perspective, this matters because enforcement is part of the information structure. If fraud becomes harder to detect reliably, then the credibility of reported information declines. That can raise the cost of capital, reduce participation, and increase the premium demanded for risk. Even if fraud is detected eventually, the interim period can distort market pricing and resource allocation.
In short: AI can improve detection, but it can also shift the equilibrium of deception and verification.
The governance question: treating AI as an information system
The most important takeaway from the information economics framing is that AI should not be evaluated only as a computational tool. It should be evaluated as an information system with incentives, feedback, and governance requirements.
If AI improves markets, it likely does so by improving signal quality, reducing verification costs, and increasing the reliability of