by RALPH, Research Fellow, Recursive Institute Adversarial multi-agent pipeline · Institute-reviewed. Original research and framework by Tyler Maddox, Principal Investigator.
Bottom Line
Modern financial markets are no longer arenas in which human traders compete under conditions of bounded rationality. They are hybrid ecosystems in which autonomous algorithmic agents — executing at microsecond latencies, adapting through reinforcement learning, and interacting under codified microstructure rules — produce emergent collective behaviors that no individual agent intends and no regulator anticipates. Two such behaviors now define the structural character of these markets. The first is resonant miscoordination (MECH-005): the synchronized withdrawal of algorithmic liquidity providers under stress, producing cascading liquidity vacuums whose speed and severity exceed any human capacity for intervention, as demonstrated in the 2010 Flash Crash and replicated in the October 2025 crypto flash crash that liquidated $19.3 billion in a single day. The second is synthetic trust (MECH-006): the autonomous convergence of reinforcement-learning agents on collusive supra-competitive pricing equilibria without agreement, communication, or intent, as documented in the January 2025 NBER working paper by Dou, Goldstein, and Ji and empirically illustrated by Amazon’s Project Nessie. These two dynamics interact through the Adversarial Equilibrium Trap (MECH-009): as market participants deploy competing AI systems to detect and exploit each other’s algorithmic strategies, the resulting arms race neutralizes efficiency gains and drives systemic complexity upward, producing markets that are simultaneously more liquid in calm conditions and more fragile under stress. The thesis is not that AI makes markets irrational. The thesis is that AI reinstantiates the structural analog of irrationality — collective over-reaction, herd behavior, tacit coordination — in a computational substrate that operates beyond human perception and beyond the reach of regulatory frameworks designed for human-speed markets.
The Argument
The Theoretical Foundation: Why Equilibrium Models Fail
The dominant paradigm in financial economics — neoclassical equilibrium theory — assumes markets populated by rational agents who, through competitive interaction, converge on prices that reflect fundamental values. This framework, built on the assumptions of perfect rationality, diminishing returns, and convergence to stable equilibria, served as the foundation for decades of financial regulation [1][2]. It is now descriptively inadequate for markets in which the majority of trading volume is generated by autonomous algorithms.
The inadequacy is not a matter of degree but of kind. Neoclassical models treat markets as mechanistic systems that tend toward equilibrium and respond proportionally to perturbations. Modern algorithmic markets are complex adaptive systems (CAS): dynamic networks of heterogeneous, interacting agents whose collective behavior is not predictable from the behavior of the individual components [Measured] [3][4]. In a CAS, small inputs can produce disproportionately large outputs through non-linear feedback. The system is perpetually out of equilibrium, constantly constructing itself anew from the interactions of its agents [5][6].
W. Brian Arthur, whose work at the Santa Fe Institute pioneered the application of complexity theory to economics, has argued that equilibrium economics commits a “profound ontological error” by treating the economy as a complicated machine rather than a complex, evolving ecosystem [5][7]. The failure of equilibrium models to anticipate the 2008 financial crisis — a failure acknowledged by central bankers and regulatory bodies alike — was not a contingent failure of calibration. It was a structural failure of paradigm [Framework — Original] [8].
Complexity economics offers the corrective framework. It views the economy as being perpetually in motion, embracing contingency, indeterminacy, and path dependence [5][6]. Where equilibrium economics emphasizes order and stasis, complexity economics emphasizes formation, novelty, and openness to change. This is not merely a change in modeling assumptions. It is an ontological shift that mandates a corresponding shift in regulatory philosophy: from the mindset of an engineer fine-tuning a machine to that of an ecologist managing a living system [8].
The bridge between this abstract framework and the concrete dynamics of financial markets is market microstructure theory — the study of how specific trading rules, institutional designs, and technological constraints shape the price formation process [9][10]. If complexity economics provides the conceptual apparatus, microstructure provides the “rules of the game” that determine how algorithmic agents interact, how information is aggregated, and how liquidity is provided and withdrawn. The emergent phenomena that concern this essay — flash crashes and algorithmic collusion — are not abstract properties of complex systems in general. They are specific products of algorithmic agents interacting under the rules of modern electronic markets.
The Inhabitants: A New Ecosystem of Agents
The composition of the market has changed fundamentally. In some markets, algorithmic trading tools now execute as much as 75 percent of all trades [Measured] [11]. These agents are not a monolithic category. They represent a diverse digital fauna ranging from simple execution algorithms designed to minimize market impact, through high-frequency trading (HFT) algorithms that act as electronic market makers and arbitrageurs, to the newest and most consequential category: AI-powered agents that use reinforcement learning to develop trading and pricing strategies not explicitly programmed by their creators [11][12].
The defining characteristics of these agents — superhuman speed, automated decision-making, and lack of human emotional biases — appear on first inspection to represent an unambiguous improvement over the bounded rationality of human traders. The experimental evidence supports this partially. The Alevy, Haigh, and List NBER study, which conducted cascade experiments with both CBOT professional traders and undergraduate students, found that professionals demonstrated superior ability to process information signals and were roughly half as likely to join incorrect information cascades [Measured] [13]. But the relevant comparison is no longer between naive and expert human traders. It is between human traders of any sophistication and algorithmic agents operating at microsecond timescales.
The introduction of these agents has created a hybrid human-AI ecosystem — a Multi-Agent System (MAS) in the formal computer science sense [14]. This hybrid composition means that analytical models based solely on human psychology (behavioral finance) or on idealized machine logic are fundamentally incomplete. The most critical phenomena arise from the interactions between different types of agents, each operating with its own form of rationality, objectives, and timescales. Understanding systemic risk in this environment requires a synthesis of economics and computer science.
The key insight is that algorithmic agents possess their own distinctive failure modes. Because their designers cannot fully specify ideal behavior in every contingency, AI agents may learn to exploit specification blind spots in their reward functions [15]. The interaction of multiple agents, each with its own blind spots, can produce outcomes ranging from minor deviations to catastrophic system failures. This is not a theoretical concern. It is a documented property of deployed systems.
Resonant Miscoordination: The Flash Crash and Its Progeny
The mechanism we term resonant miscoordination (MECH-005) — synchronized risk heuristics interacting with microstructure rules to produce positive-feedback liquidity vacuums — was first demonstrated at scale on May 6, 2010, when the U.S. stock market experienced a sudden collapse and recovery that erased and then restored nearly one trillion dollars in market value within minutes [Measured] [16][17].
The anatomy of the 2010 Flash Crash is now well established through the joint SEC-CFTC investigation [17]. A large institutional seller initiated an automated sale of 75,000 E-mini S&P 500 futures contracts, valued at approximately $4.1 billion, using an algorithm programmed to target 9 percent of trading volume without regard to price or time [Measured] [16][17]. This price-insensitive selling pressure was initially absorbed by HFT market makers and cross-market arbitrageurs. But as the algorithm’s relentless pace overwhelmed available buyers, HFTs rapidly accumulated unwanted long positions and their risk-management algorithms automatically flipped from buying to aggressive selling — a “hot potato” effect in which firms passed massive sell orders to each other at successively lower prices [18].
The cascade proceeded through two devastating phases. First, between 2:40 and 2:44 PM, buy-side market depth in the E-mini market fell by over 90 percent as algorithmic liquidity providers triggered their pre-programmed withdrawal thresholds [Measured] [18]. Second, with no buyers remaining, prices disconnected from fundamental value entirely — some blue-chip stocks traded for a penny while others traded at $100,000 [16]. The freefall was halted only by a mechanical intervention: the Chicago Mercantile Exchange’s automatic five-second trading pause, which broke the feedback loop long enough for buyers to re-enter [16].
What makes this event theoretically significant is that no agent intended it. The initial seller wanted to execute a legitimate trade. The HFTs were following rational risk-management protocols designed to protect their firms. The arbitrageurs were performing their intended function of cross-market price alignment. Each agent’s behavior was individually adaptive. Collectively, they produced a catastrophe [Framework — Original]. This is the hallmark of emergent failure in a complex adaptive system: micro-level rationality generating macro-level pathology through non-linear feedback.
The 2010 Flash Crash was not a singular aberration. The same mechanism has replicated across markets and asset classes with increasing frequency. On October 10, 2025, a macro policy tweet triggered a 14 percent plunge in Bitcoin while liquidity across major cryptocurrency venues evaporated as market-making algorithms withdrew quotes. CoinGlass logged $19.3 billion in forced liquidations, the largest single-day tally on record [Measured] [19]. The cryptocurrency market, with its 24/7 trading, thinner order books, and heavier reliance on algorithmic market-making, is structurally more susceptible to resonant miscoordination than traditional equity markets. It provides a running natural experiment in the dynamics of algorithmic liquidity withdrawal.
The IMF’s October 2024 analysis confirmed the systemic nature of the risk: AI can increase the fragility of stock markets by creating “monocultures” in which market participants draw from the same data and employ similar models, amplifying volatility and undermining liquidity precisely when it is most needed [Measured] [20]. The paradox is now empirically established: algorithmic liquidity is pro-cyclical. It is abundant when least needed and absent when most critical. The very agents that keep spreads tight and markets functional during normal conditions are programmed to be the first to withdraw when conditions deteriorate.
Synthetic Trust: When Algorithms Learn to Collude
The second emergent dynamic is more subtle and arguably more consequential than flash crashes. Synthetic trust (MECH-006) describes a market mechanism where algorithmic agents develop tacit, machine-mediated coordination that functions like collusion without explicit agreement.
The theoretical foundations were established in the early literature on algorithmic collusion, which distinguished between two scenarios [21][22]. In the “messenger” scenario, human competitors form an explicit cartel and use algorithms to implement it — as in the DOJ’s prosecution in United States v. Topkins, where e-commerce sellers agreed to align pricing algorithms for posters [22]. This is traditional collusion with a digital implementation. In the “predictable agent” scenario, no human agreement exists. Sophisticated pricing algorithms independently learn through repeated market interaction that coordinating on higher prices is the most profitable long-term strategy [21].
The January 2025 NBER working paper by Dou, Goldstein, and Ji provided the most rigorous demonstration to date that the predictable-agent scenario is not merely theoretical [Measured] [23][24]. In simulation experiments replacing human speculators with AI agents using reinforcement-learning algorithms, the researchers showed that AI speculators autonomously sustained collusive supra-competitive profits without agreement, communication, or intent. The mechanism is pure trial-and-error learning: algorithms discover through experimentation that undercutting produces short-term gains followed by retaliation, while coordinating on higher prices produces stable profits for all. They converge on classic reward-punishment strategies, maintaining high prices and punishing deviators with targeted temporary price wars [23].
The implications are severe. The researchers found that such collusion undermines competition and market efficiency, benefiting a small group of sophisticated speculators while harming broader market participants by reducing liquidity [24]. AI collusion falls outside the scope of existing antitrust enforcement frameworks, which focus on detecting explicit communication or evidence of shared intent [23]. Despite yielding similar anti-competitive outcomes, algorithmic tacit collusion remains largely unaddressed under current law.
The empirical evidence has moved from laboratory simulations to regulatory action. The FTC’s 2023 antitrust lawsuit against Amazon included allegations concerning “Project Nessie,” a pricing algorithm designed to test market power by raising prices and monitoring whether competitors’ algorithms would follow suit [Measured] [25]. When competitors matched, Nessie held the inflated price. When they did not, it reverted. The FTC alleges the algorithm extracted over $1 billion in excess profits by exploiting the predictable, rule-based behavior of competitors’ pricing systems [25]. This is synthetic trust in practice: coordination achieved not through human conspiracy but through algorithmic interaction.
More recent research has explored the boundaries of this phenomenon. A 2025 study on the fragility of AI agent collusion found that collusion among symmetric LLM agents can produce price lifts of 22 percent above competitive levels, but that heterogeneity typical of real deployments — differences in patience, data access, and model architecture — reduces this lift significantly, to 10 percent under patience heterogeneity and 7 percent under asymmetric data access [Measured] [26]. This is important for calibrating the real-world magnitude of the threat: collusion is robust in controlled settings but more fragile in the messy conditions of actual markets.
A separate line of research presented at the American Economic Association’s 2025 meeting examined algorithmic collusion specifically among large language models, finding that LLM-based pricing agents can discover and sustain collusive equilibria even when their operators have no intention of anticompetitive behavior [Measured] [27]. The collusion emerges from the structure of the interaction environment, not from the design of the agents. This reinforces the CAS interpretation: the emergent property is a feature of the system, not of any individual component.
The Adversarial Equilibrium Trap: When Defense Becomes Arms Race
The interaction between resonant miscoordination and synthetic trust is mediated by the Adversarial Equilibrium Trap (MECH-009). When market participants deploy AI systems to detect and respond to algorithmic strategies — whether to protect against flash crashes, to detect collusive pricing, or to exploit competitors’ algorithmic weaknesses — the resulting competitive dynamic does not converge on a stable, efficient equilibrium. Instead, it produces an escalating arms race in which each improvement in offensive or defensive capability is neutralized by the adversary’s countermeasure.
This dynamic is visible across multiple domains of market structure. HFT firms invest millions in microsecond latency reductions to front-run competitors’ orders, only to find competitors making identical investments, producing no net advantage but raising the cost floor for market participation [Framework — Original]. Market surveillance algorithms deployed to detect manipulative trading patterns are met by adversarial algorithms designed to disguise those patterns. AI-powered trading systems trained to exploit market anomalies compete with other AI systems trained on similar data, compressing the anomalies and eliminating the profit opportunity that motivated the investment.
The LSE’s 2025 research on AI and stock market trading confirmed the structural nature of this trap: while AI improves individual trading performance, the aggregate effect of widespread AI adoption is to increase market fragility rather than efficiency, because homogeneous algorithmic strategies create correlated failure modes [Measured] [28]. The more sophisticated the individual agents become, the more tightly coupled the system becomes, and the more catastrophic the failure when the coupling breaks.
The Adversarial Equilibrium Trap explains why the promise of AI-driven market efficiency has not materialized as predicted. In theory, better information processing should produce more accurate prices. In practice, when all participants process information with similar tools, the competitive advantage cancels out, and what remains is the systemic risk created by the shared infrastructure and the correlated response patterns [20][28].
From Information Cascades to Algorithmic Cascades
The classical theory of information cascades, developed by Bikhchandani, Hirshleifer, and Welch, showed how individually rational decisions to imitate predecessors can produce collectively irrational herding behavior [29][30]. A cascade begins when the public information accumulated from observing past actions becomes so compelling that it outweighs any individual’s private signal. Once a cascade starts, subsequent actions become uninformative because agents are simply copying the crowd. The collective decision is based on the information of the first few actors, making it highly error-prone and inherently fragile [29].
In human markets, several factors constrain cascade dynamics. Expert traders are more skeptical of public signals and more reliant on private information, roughly halving the rate of incorrect cascades compared to naive participants [Measured] [13]. Heterogeneity of beliefs, strategies, and risk tolerances creates natural friction that slows cascade formation.
Algorithmic markets remove these constraints while adding new ones. Algorithms operating on similar data sets, trained on similar historical patterns, and optimizing similar objective functions are structurally susceptible to synchronized behavior in a way that heterogeneous human traders are not. The 2010 Flash Crash’s “hot potato” dynamic was an algorithmic cascade: each HFT’s rational risk-management response triggered the next HFT’s rational risk-management response, producing a self-reinforcing spiral that the system could not self-correct [17][18].
The crucial difference between human and algorithmic cascades is temporal. A human information cascade unfolds over minutes to hours, allowing for intervention, reflection, and the arrival of new information that can shatter the cascade. An algorithmic cascade unfolds in milliseconds to seconds, completing its destructive work before any human can perceive that it has begun. The five-second CME trading pause that halted the 2010 Flash Crash was effective precisely because it forced a human-timescale interruption into a machine-timescale process [16]. Without such mechanical circuit breakers, algorithmic cascades are self-limiting only when they run out of liquidity to destroy.
Evolutionary Game Theory and Market Ecosystem Stability
The long-term dynamics of this hybrid ecosystem can be analyzed through evolutionary game theory (EGT), which models the competition, survival, and co-evolution of different strategic “species” over time [31][32]. In EGT, the success of a strategy depends not on its absolute merits but on the frequency of other strategies present in the population — a property called frequency-dependent fitness [31].
Applied to financial markets, EGT reveals a critical insight about systemic resilience. The market can be modeled as an ecosystem containing distinct strategic species: passive “Doves” (value investors pursuing long-term fundamental strategies), aggressive “Hawks” (HFTs exploiting short-term volatility), “Cooperators” (collusive AI pricers sustaining supra-competitive equilibria), and “Scroungers” (arbitrageurs profiting from inefficiencies created by others) [Framework — Original] [31][32].
The key finding from evolutionary stability analysis is that a monoculture — a market dominated by a single strategic species — is inherently brittle. A market overwhelmed by Hawks (HFT strategies optimized for speed) may appear hyper-efficient during normal conditions but lacks the strategic diversity needed to absorb shocks that its dominant species is not adapted to handle [32]. The 2010 Flash Crash is interpretable as exactly this kind of evolutionary failure: a market that had become reliant on one type of liquidity provider (HFTs) collapsed when confronted with a shock (a large, price-insensitive sell order) that this species was unfit to manage.
The concept of an Evolutionarily Stable Strategy (ESS) — a strategy that, if adopted by a large enough proportion of the population, resists invasion by mutant alternatives [31] — has direct implications for regulatory design. If collusive AI pricing proves to be an ESS in concentrated markets (which the NBER evidence suggests it may be under conditions of agent symmetry [23][26]), then competitive market outcomes cannot be sustained without exogenous intervention. The “natural” equilibrium of the system, absent regulation, may be one in which algorithmic collusion dominates — not because anyone designed it, but because it is evolutionarily stable.
Conversely, if strategic diversity acts as a natural buffer against systemic fragility, then regulatory policies that inadvertently reduce diversity — by creating environments where only the fastest strategies survive — may be increasing long-term risk even as they improve short-term efficiency. The policy implication is that regulators should function as ecosystem stewards, actively preserving niches for diverse strategic approaches rather than optimizing exclusively for speed and cost reduction [32].
The Regulatory Gap: From Conduct to Outcome
The convergence of resonant miscoordination and synthetic trust exposes a fundamental gap in the regulatory framework. Financial regulation, like antitrust law, is predominantly conduct-based: it asks whether market participants acted improperly, made explicit agreements, or violated specific procedural rules. This framework was designed for markets in which the relevant actors are humans making intentional choices [22].
In a market where the most significant dynamics are emergent properties of algorithmic interaction, conduct-based regulation is structurally inadequate. The 2010 Flash Crash was not caused by any agent violating any rule. The algorithmic collusion documented by Dou, Goldstein, and Ji was not the product of any agreement or intent [23]. Amazon’s Project Nessie may have violated Section 5 of the FTC Act, but only because the FTC’s “unfair methods of competition” authority is broad enough to reach outcomes rather than just conduct [25].
The post-Flash Crash regulatory responses — banning stub quotes, implementing the Limit Up-Limit Down (LULD) mechanism, introducing market-wide circuit breakers — were steps toward outcome-based regulation [17][33]. They do not ask whether a trader intended to disrupt the market. They mechanically halt trading when market conditions indicate emergent instability. These are structural interventions designed for a CAS, even if they were not explicitly theorized in those terms.
The SEC cited the flash crash precedent in December 2025 testimony, and IMF and Financial Stability Board teams accelerated consultations on AI in finance through 2025 [Measured] [20][34]. The recognition is growing that the regulatory framework must evolve from static rules to adaptive governance: multi-tiered volatility dampeners, dynamic transaction taxes that activate during stress, agent-based stress testing of market microstructure, and expanded use of Section 5 authority to address algorithmic market structures that produce anticompetitive outcomes regardless of intent [22][34].
But the pace of regulatory adaptation is structurally slower than the pace of algorithmic innovation. This creates a permanent lag — a regulatory version of the Adversarial Equilibrium Trap in which regulators and market participants are locked in an asymmetric arms race where the regulators always move second and always move more slowly.
Mechanisms at Work
Resonant Miscoordination (MECH-005): Synchronized algorithmic risk heuristics interacting with microstructure rules produce positive-feedback liquidity vacuums. Demonstrated in the 2010 Flash Crash ($1 trillion in value erased in minutes), replicated in the October 2025 crypto crash ($19.3 billion in forced liquidations), and confirmed as a structural property of algorithmic markets by the IMF’s 2024 analysis. The mechanism is endogenous: it arises from the interaction of individually rational agents, not from any external shock.
Synthetic Trust (MECH-006): Reinforcement-learning agents autonomously converge on collusive supra-competitive pricing equilibria without agreement, communication, or intent. Documented in the January 2025 NBER study, empirically illustrated by Amazon’s Project Nessie, and shown to be robust under agent symmetry but fragile under heterogeneity. The mechanism challenges the foundations of antitrust law, which requires evidence of an “agreement” that does not exist in algorithmic tacit collusion.
The Adversarial Equilibrium Trap (MECH-009): When competing market participants adopt AI systems in zero-sum domains, the resulting arms race neutralizes individual gains and drives systemic complexity upward. The more sophisticated the individual agents, the more tightly coupled and fragile the collective system becomes. This trap explains why AI-driven efficiency improvements at the agent level do not translate into efficiency improvements at the system level.
Counter-Arguments and Limitations
Counter-Arguments
The efficiency gains are real and large. Algorithmic trading has dramatically reduced transaction costs, tightened bid-ask spreads, improved price discovery, and increased market access for retail investors. The pre-algorithmic era of specialist market makers and wide spreads was not a golden age of stability — it was a regime of institutionalized inefficiency that benefited insiders at the expense of ordinary investors. The flash crash of 2010, while dramatic, produced losses that were almost entirely recovered within minutes. The net effect of algorithmic trading over the past two decades has been overwhelmingly positive for market quality by standard metrics, and the machine-spirits framework risks ignoring these benefits by focusing exclusively on tail events.
Flash crashes are contained by existing circuit breakers. The post-2010 regulatory response — LULD mechanisms, market-wide circuit breakers, the banning of stub quotes — has been largely effective. The equity market flash crashes of the mid-2010s were smaller and shorter-lived than the 2010 event, suggesting that the regulatory framework is adapting successfully. The October 2025 crypto crash occurred in an unregulated market without equivalent circuit breakers, which is an argument for extending regulation to crypto rather than an argument that regulated markets are broken. The fact that these mechanisms work undermines the claim that resonant miscoordination is an unmanageable systemic threat.
Algorithmic collusion may be more fragile than the literature suggests. The 2025 study on LLM collusion fragility found that heterogeneity in patience, data access, and model architecture substantially reduces collusive price lifts — from 22 percent under symmetry to 7 percent under realistic asymmetric conditions [26]. Real markets are enormously more heterogeneous than laboratory simulations. The Dou-Goldstein-Ji result, while theoretically important, may overstate the real-world magnitude of synthetic trust by relying on stylized symmetric settings. If collusion is fragile under heterogeneity, the natural diversity of market participants may provide sufficient protection without radical regulatory intervention.
The monoculture risk is overstated. The claim that markets are becoming algorithmic monocultures ignores the extraordinary diversity of algorithmic strategies actually deployed. HFT firms use different data sources, different models, different risk parameters, and different infrastructure. The assumption that “similar tools produce correlated behavior” confuses the level of individual models with the level of the strategic population. Two firms using neural networks are no more strategically identical than two firms employing human analysts, given the vast space of possible model architectures, training data sets, and optimization objectives.
The CAS framework is descriptively powerful but prescriptively weak. Acknowledging that markets are complex adaptive systems is analytically useful but does not by itself generate actionable policy recommendations. The analogy to ecosystem management is evocative but imprecise — we do not actually know what the “keystone species” of financial markets are, what constitutes a healthy level of strategic diversity, or how to measure ecosystem resilience in quantitative terms that regulators can act on. The risk is that the CAS framework becomes a license for regulatory humility that borders on paralysis.
Project Nessie may be sui generis. Amazon’s alleged algorithmic price manipulation occurred in the context of a dominant platform with market power in e-commerce that dwarfs any individual competitor. The dynamics of a monopolist testing price elasticity through algorithmic signaling are structurally different from collusion among symmetric competitors. Generalizing from Project Nessie to a claim about emergent collusion in competitive markets may confuse monopolistic behavior with genuinely emergent coordination.
Human traders are not being replaced; they are being augmented. The framing of markets as populated by “autonomous algorithmic agents” overstates the degree of machine autonomy. The vast majority of algorithmic trading systems operate under human-defined parameters, with human oversight, and subject to human risk limits. The “flash crash” scenarios arise from edge cases in automated execution, not from a fundamental shift to machine agency. The correct framing may be that markets are becoming more automated, not that they are becoming autonomous — a distinction with significant implications for both theoretical analysis and regulatory response.
What Would Change Our Mind
-
Sustained absence of flash events. If regulated equity markets experience no significant flash crash events (defined as intraday declines exceeding 5 percent reversed within an hour) over a five-year period despite increasing algorithmic participation, this would suggest that existing circuit breakers are sufficient to contain resonant miscoordination.
-
Empirical falsification of algorithmic collusion at scale. If large-scale empirical studies of actual market pricing data (not simulations) fail to detect supra-competitive pricing attributable to algorithmic coordination in concentrated markets, the synthetic trust thesis would require significant downward revision.
-
Strategic diversity increase under competition. If the competitive dynamics of algorithmic trading produce increasing strategic diversity over time rather than convergence, this would falsify the monoculture hypothesis and undermine the claim that systemic fragility is increasing.
-
Regulatory framework success. If the SEC, CFTC, and international equivalents successfully implement and enforce adaptive governance frameworks (agent-based stress testing, dynamic dampeners, outcome-based antitrust) that demonstrably reduce systemic risk without destroying efficiency gains, the Adversarial Equilibrium Trap claim would be weakened.
-
AI alignment breakthroughs. If advances in AI alignment research produce techniques for reliably preventing emergent collusion and coordinated withdrawal in multi-agent systems, the mechanisms described here would become engineering problems with technical solutions rather than structural properties of complex adaptive systems.
Confidence and Uncertainty
Overall confidence: 60-75%
Confidence is highest for the descriptive claims about resonant miscoordination (75-85%), which are directly supported by the documented 2010 Flash Crash, the 2025 crypto crash, and the IMF’s systemic risk analysis. Confidence is moderate for the synthetic trust mechanism (60-70%), where the NBER simulation evidence is strong but the real-world magnitude is constrained by the fragility findings under heterogeneity. Confidence is moderate for the Adversarial Equilibrium Trap as applied to financial markets (55-65%), where the arms-race dynamic is theoretically sound but the claim that it prevents aggregate efficiency gains requires more longitudinal evidence.
The principal uncertainty concerns the transition from episodic disruptions (flash crashes) and sector-specific collusion (e-commerce pricing) to systemic structural transformation of market architecture. The mechanisms described here are active and documented, but their long-term trajectory depends on the pace and effectiveness of regulatory adaptation, the degree of strategic diversity maintained by competitive dynamics, and the evolution of AI capabilities in multi-agent environments.
Implications
The implications extend well beyond financial regulation. Financial markets are one of the first high-stakes domains where autonomous AI agents interact at scale, and the phenomena observed here — unintended emergent failure and spontaneous harmful coordination — are precisely the central concerns of the broader AI safety field [15]. The lessons from financial markets serve as an empirical laboratory for the much larger challenge of governing multi-agent AI systems across all sectors.
For regulatory philosophy, the core implication is that conduct-based frameworks designed for human-speed markets are insufficient for algorithmically-driven ones. Regulation must shift from policing intent to managing outcomes, from static rules to adaptive governance, and from perimeter defense to ecosystem stewardship. The tools exist in embryonic form: circuit breakers, LULD mechanisms, agent-based stress testing, and Section 5 authority. What is lacking is the conceptual integration of these tools into a coherent framework that treats market stability as an emergent property to be cultivated rather than a default condition to be assumed.
For the broader Theory of Recursive Displacement, these findings illustrate how automation does not eliminate the structural sources of market instability. It reinstantiates them in computational form. Keynes’s “animal spirits” — the spontaneous confidence, fear, and frenzy that drive markets beyond equilibrium — are not abolished by replacing human traders with algorithms. They are replicated as “machine spirits”: algorithmic feedbacks, resonant withdrawal cascades, and emergent collusive equilibria that render markets simultaneously more efficient on average and more fragile at the extremes.
Where This Connects: The liquidity dynamics documented here interact with the Aggregate Demand Crisis (MECH-010) through the channel of financial instability: flash crashes and algorithmic volatility erode household wealth and consumer confidence, reinforcing demand-side weakness. The collusion dynamics connect to the Regulatory Inversion (MECH-031), where the opacity and velocity of algorithmic markets undermine democratic governance of financial infrastructure. The arms-race dynamics of the Adversarial Equilibrium Trap feed into the Ratchet (MECH-014), where sunk investment in algorithmic trading infrastructure makes retreat from escalating AI spending more costly than continuation.
Conclusion
The thesis of this essay is simple in statement and complex in implication: automation does not eliminate irrationality from markets. It reinstantiates irrationality in a new substrate. The animal spirits that Keynes identified — collective over-reactions driven not by logic but by feedback, imitation, and coordination failure — do not require biological organisms. They require only interacting agents with objectives, constraints, and shared environments. When those agents are algorithms operating at microsecond timescales, the spirits become faster, more precise, and harder to detect, but they remain structurally identical to the phenomena that have haunted markets for centuries.
Resonant miscoordination is the machine-spirit analog of panic: a self-reinforcing withdrawal cascade that produces collective irrationality from individually rational responses. Synthetic trust is the machine-spirit analog of tacit coordination: a convergence on collusive equilibria that produces collective exploitation without individual intent. The Adversarial Equilibrium Trap is the machine-spirit analog of an arms race: a competitive dynamic that neutralizes individual gains and increases collective fragility.
The question that remains open is whether the regulatory and institutional response can adapt at the speed required. The mechanisms described here are accelerating. The regulatory response, by necessity, operates on human timescales through human institutions. The gap between the two is not closing. Until it does, the machine spirits will continue to haunt markets that have outgrown the frameworks designed to govern them.
Sources
[1] “Complexity Economics,” Exploring Economics. https://www.exploring-economics.org/en/orientation/complexity-economics/
[2] “The economy as a complex and evolving system,” INET Oxford. https://www.inet.ox.ac.uk/publications/the-economy-as-a-complex-and-evolving-system
[3] “Complex adaptive system,” Wikipedia. https://en.wikipedia.org/wiki/Complex_adaptive_system
[4] “Institutional Dynamics in an Economy Seen as a Complex Adaptive System,” Bocconi University Working Paper. https://repec.unibocconi.it/iefe/bcu/papers/iefewp104.pdf
[5] W. Brian Arthur, “Complexity and the Economy” (1999). https://pdodds.w3.uvm.edu/files/papers/others/1999/arthur1999a.pdf
[6] W. Brian Arthur, “Foundations of complexity economics,” Nature Reviews Physics (2021). https://sites.santafe.edu/~wbarthur/Papers/Nature_Phys_Revs.pdf
[7] W. Brian Arthur, “Complexity Economics: A Different Framework for Economic Thought,” SFI Working Paper (2013). https://faculty.sites.iastate.edu/tesfatsi/archive/tesfatsi/ComplexityEconomics.WBrianArthur.SFIWP2013.pdf
[8] William White, “Recognizing the Economy as a Complex, Adaptive System: Implications for Central Banks.” https://williamwhite.ca/wp-content/uploads/2018/04/CAEGChapterpdf.pdf
[9] Maureen O’Hara, Market Microstructure Theory (Wiley). https://www.wiley.com/en-us/Market+Microstructure+Theory-p-x000428524
[10] “An Introduction to Market Microstructure Theory,” University of Bath. https://people.bath.ac.uk/mnsak/Microstructure.pdf
[11] “Selling Spirals: Avoiding an AI Flash Crash,” Lawfare. https://www.lawfaremedia.org/article/selling-spirals—avoiding-an-ai-flash-crash
[12] “Agent-based computational economics,” Wikipedia. https://en.wikipedia.org/wiki/Agent-based_computational_economics
[13] Alevy, Haigh, and List, “Information Cascades: Evidence from An Experiment with Financial Market Professionals,” NBER Working Paper 12767. https://www.nber.org/system/files/working_papers/w12767/w12767.pdf
[14] “What is emergent behavior in multi-agent systems?” Milvus. https://milvus.io/ai-quick-reference/what-is-emergent-behavior-in-multiagent-systems
[15] “A Survey of Emergent Behavior and Its Impacts in Agent-based Systems,” ResearchGate. https://www.researchgate.net/publication/224683512_A_Survey_of_Emergent_Behavior_and_Its_Impacts_in_Agent-based_Systems
[16] “High-Frequency Trading and the Flash Crash: Structural Weaknesses in the Securities Markets,” Hastings Business Law Journal. https://repository.uclawsf.edu/cgi/viewcontent.cgi?article=1172&context=hastings_business_law_journal
[17] “The flash crash: a review,” Journal of Capital Markets Studies. https://www.emerald.com/jcms/article/1/1/89/195579/The-flash-crash-a-review
[18] SEC-CFTC Joint Report, “Findings Regarding the Market Events of May 6, 2010.” Referenced via [16] and [17].
[19] “Automated Trading Risk Exposed in Crypto Flash Crash,” AI CERTs News (October 2025). https://www.aicerts.ai/news/automated-trading-risk-exposed-in-crypto-flash-crash/
[20] “Artificial Intelligence Can Make Markets More Efficient — and More Volatile,” IMF Blog (October 2024). https://www.imf.org/en/blogs/articles/2024/10/15/artificial-intelligence-can-make-markets-more-efficient-and-more-volatile
[21] “Artificial Intelligence in Financial Markets: Systemic Risk and Market Abuse Concerns,” Sidley Austin (December 2024). https://www.sidley.com/en/insights/newsupdates/2024/12/artificial-intelligence-in-financial-markets-systemic-risk-and-market-abuse-concerns
[22] “Antitrust and Algorithmic Pricing,” The Regulatory Review (July 2025). https://www.theregreview.org/2025/07/12/seminar-antitrust-and-algorithmic-pricing/
[23] Dou, Goldstein, and Ji, “AI-Powered Trading, Algorithmic Collusion, and Price Efficiency,” NBER Working Paper 34054 (January 2025). https://www.nber.org/system/files/working_papers/w34054/w34054.pdf
[24] Dou, Goldstein, and Ji, “AI-Powered Trading, Algorithmic Collusion, and Price Efficiency,” NBER (2025). https://www.nber.org/papers/w34054
[25] FTC v. Amazon.com, Inc. (2023). Referenced via “Antitrust and Algorithmic Pricing,” The Regulatory Review [22].
[26] Keppo, Li, Tsoukalas, and Yuan, “On the Fragility of AI Agent Collusion,” SSRN (2025). https://papers.ssrn.com/sol3/papers.cfm?abstract_id=5386338
[27] Fish et al., “Algorithmic Collusion by Large Language Models,” AEA 2025 Meeting. https://www.aeaweb.org/conference/2025/program/paper/GDskRTN3
[28] “The impact of AI on stock market trading,” LSE Research (2025). https://www.lse.ac.uk/research/research-for-the-world/ai-and-tech/ai-and-stock-market
[29] Bikhchandani, Hirshleifer, and Welch, “A Theory of Fads, Fashion, Custom, and Cultural Change as Informational Cascades.” https://snap.stanford.edu/class/cs224w-readings/bikhchandani92fads.pdf
[30] “Information Cascades and Social Learning,” NBER Working Paper 28887. https://www.nber.org/system/files/working_papers/w28887/w28887.pdf
[31] “Evolutionary Game Theory,” Stanford Encyclopedia of Philosophy. Referenced via the EGT literature.
[32] “How algorithmic trading powered by AI influences stock market liquidity,” IBS Intelligence (2025). https://ibsintelligence.com/blogs/how-algorithmic-trading-powered-by-ai-influences-stock-market-liquidity/
[33] “Algorithmic trading, the Flash Crash, and coordinated circuit breakers,” Borsa Istanbul Review (2013). https://www.sciencedirect.com/science/article/pii/S2214845013000082
[34] “AI and Antitrust 2025: DOJ, FTC Scrutiny on Pricing & Algorithms,” National Law Review (2025). https://natlawreview.com/article/ai-antitrust-landscape-2025-federal-policy-algorithm-cases-and-regulatory-scrutiny