Economic Collapse Scenarios

On May 6, 2010, at 2:32 in the afternoon, the American stock market began to disappear. Over the next thirty-six minutes the Dow Jones Industrial Average fell nearly a thousand points — close to nine percent — the steepest intraday drop in its history to that moment. Shares of Accenture, a company worth tens of billions of dollars, briefly traded for a single penny. Sotheby's, at the other extreme, spiked to $100,000 a share. Then, almost as fast as it had vanished, the market came back. By 3:07 most of the loss had reversed. No war had broken out. No bank had failed. The official post-mortem by the SEC and the Commodity Futures Trading Commission traced the event to a single large automated sell order that interacted with high-frequency trading algorithms in a way nobody had designed and nobody could stop in real time. The machines had sold to each other in a loop, faster than any human could intervene, and then — for reasons equally mechanical — stopped.

The 2010 Flash Crash lasted half an hour and left almost no lasting damage. It is worth beginning here precisely because it was small, brief, and reversible. It was a controlled demonstration, run by the market on itself, of a failure mode that has since grown more plausible, more consequential, and harder to contain. The question this chapter asks is what happens when that failure mode arrives not in a market of relatively simple, competing high-frequency bots, but in one dominated by large, sophisticated, and — crucially — similar artificial intelligence systems, at a moment when those same systems have been the object of one of the largest capital booms in economic history.

The Warning Nobody Acted On

In 2023, Gary Gensler, then chair of the U.S. Securities and Exchange Commission and a former professor who had taught a course on AI and finance at MIT, said something unusually blunt for a sitting regulator. He told the Financial Times that it was "nearly unavoidable" that AI would trigger a financial crisis within a decade unless regulators moved quickly, and that a future crash could well be traced to the technology. His specific worry was not rogue AI or malfunction. It was uniformity. If most institutions come to rely on the same handful of foundational models and the same data feeds, he argued, they will tend to act alike — and markets in which everyone acts alike are markets that move all at once.

Gensler's warning is the pivot for this chapter because it names the essential fact about this risk: it was foreseen, in public, by the person with the most authority to address it, and the machinery of prevention did not engage. To understand why, and to judge how much weight his warning deserves, we need to look at three things in turn — the financial conditions that make the market fragile, the mechanism by which AI could turn fragility into cascade, and the structural reasons the warning did not translate into action.

A Boom Built for a Bust

Every speculative bubble in history has shared three ingredients: overvaluation, leverage, and a story compelling enough to keep people buying past the point of prudence. The AI investment surge of the mid-2020s supplied all three at a scale that dwarfs most historical precedents. JP Morgan analysts estimated that something on the order of $6 trillion in capital would be needed between 2025 and 2030 for AI-related data centers, the power infrastructure to run them, and the supply chains to build them. Capital of that magnitude does not flow quietly. It floods, and it drags valuations with it.

The character of the boom bore the classic markings of speculative excess. Firms with tenuous connections to artificial intelligence commanded premiums for the association alone. Valuations detached from any demonstrated path to profit; analysts openly debated whether companies trading at a hundred times earnings could ever grow into their prices, and kept buying anyway. This is the signature trap of a bubble: staying in feels irrational, but so does leaving, because the person who exits a year early looks like a fool right up until the moment they look like a genius. Individual rationality and collective madness are, for a while, indistinguishable.

What matters for the collapse scenario is not whether this particular boom ends in a crash — booms end in many ways — but that a highly valued, highly concentrated, narratively fragile sector sits at the center of the market. When a large share of index value depends on a handful of firms whose worth rests on projections rather than earnings, the market has a soft spot. A disappointing product cycle, a stretch of slower-than-promised capability gains, a rise in the energy costs that AI infrastructure assumed would stay cheap, a wave of job-displacement anger translating into taxes and restrictions — any of these can start prices sliding. In an ordinary market, a slide is a correction. The next section is about what turns a correction into something worse.

The Cascade Mechanism

By the mid-2020s, the large majority of trading volume in major equity markets was already generated by algorithms rather than humans deciding trade by trade; estimates of the automated share commonly run from roughly seventy to eighty-five percent, depending on how one counts. These systems are fast, tireless, and, increasingly, intelligent in the specific sense that they learn patterns from vast quantities of market history rather than executing fixed rules. Taken one at a time, they are marvels of disciplined risk management. Taken together, they are something the market has never had before: a population of decision-makers that increasingly think alike.

That likeness is the whole problem, and it is worth being precise about why. AI trading systems at rival institutions are trained on overlapping historical data, built on similar architectures, and measured against similar benchmarks. Competition, paradoxically, drives them toward convergence rather than diversity, because every firm is solving the same optimization problem with the same tools against the same yardstick. Over time they come to interpret market conditions in similar ways and to reach for similar responses. This is a systemic risk produced not by any system failing but by every system succeeding at the same task in the same way — a risk that individual-firm risk management is structurally blind to, because each firm's model looks prudent precisely when measured in isolation.

When stress arrives, the convergence expresses itself as a sequence of reinforcing moves that unfolds faster than description.

graph TD
    A[Trigger: price shock in overvalued sector] --> B[Momentum algorithms detect downtrend, sell to cut losses]
    B --> C[Prices fall further]
    C --> D[Market-making AIs detect volatility, withdraw liquidity]
    D --> E[Bid-ask spreads widen, small orders move prices violently]
    E --> F[Cross-market AIs transmit shock to bonds, FX, commodities]
    F --> G[Diversification fails: everything falls together]
    G --> B

Momentum-tracking systems recognize a downward trend and sell to avoid loss, which deepens the trend. Market-making systems, whose job is to post the buy and sell quotes that keep trading orderly, detect the spike in volatility and do the rational thing for their own books: they pull back, widening spreads and thinning the order book exactly when depth is most needed. This is the same liquidity withdrawal that turned Accenture into a penny stock in 2010, now operating across a market where liquidity provision itself is automated. Systems trading across asset classes carry the shock from equities into bonds, currencies, and commodities, so that the diversification which was supposed to cushion losses instead synchronizes them. And because each step happens in milliseconds, the loop can run many times before a human supervisor has finished reading the first alert. The diversity that once made markets self-stabilizing — the fact that at any price, some participant saw a bargain and others saw a warning — has been quietly optimized away.

Why the Old Fire Brigade Cannot Reach the Fire

The financial crises of the modern era — 1987, 1998, 2008 — were ultimately contained by humans: central banks cutting rates and buying assets, regulators invoking emergency authority, institutions coordinating rescues over tense weekends. Every one of those responses assumed that decision-makers could recognize a crisis, agree on a response, and implement it before the damage became permanent. That assumption rests on a timescale. Central-bank tools work through credit conditions and credibility, over hours and days and weeks. An algorithmic cascade works in milliseconds.

The mismatch is not merely that human institutions are slow. It is that some interventions may be actively misread by the machines they are meant to reassure. When a central bank announces an emergency lending facility, an AI risk system scanning for danger signals can read the announcement itself as confirmation that things are dire, and tighten rather than loosen. When a central bank buys assets to prop up prices, systems capable of detecting the footprint of that buying can front-run it, taking profit while prices keep falling. Trading halts, meanwhile, are porous: a halt on one exchange does not stop trading in dark pools, foreign venues, or derivatives, and the cascade simply routes around the closed door.

The implication is uncomfortable for the entire architecture of financial crisis management. Reactive intervention — the model of the fire brigade that arrives after the alarm — presumes there is time between the alarm and the collapse. In a market moving at machine speed, that interval can shrink toward zero. Effective safeguards, if they exist, have to be built into the systems and the market structure before the stress event: pre-committed constraints, not emergency judgment calls. This is the single most important structural conclusion of the chapter, and it reframes the entire regulatory question from "how do we respond?" to "what do we require in advance?"

The Four Channels to the Real Economy

A crash confined to trading screens would be a spectacle, not a catastrophe. The reason financial cascades matter to people who own no stocks is that they leak into the real economy through four channels, each of which feeds the others.

The first and fastest is credit. When institutions take losses and grow uncertain about further losses, they stop lending to preserve capital. Businesses that rely on revolving credit for payroll and inventory can find themselves unable to function within weeks, even when they are fundamentally sound. As firms cut costs, they cut workers; laid-off workers spend less; falling demand cuts revenue further; and another round of layoffs begins. The second channel is housing: a severe shock can drive home prices down by a third or more within months, wiping out household equity, triggering payment spikes on adjustable-rate loans, and feeding losses back into the banks through their mortgage books — which tightens credit further, closing a loop between channels one and two. The third is pensions and savings: retirement funds heavily exposed to equities can lose half their value in a crash of this depth, and households that discover their savings have evaporated respond by saving more and spending less, prolonging the demand slump. The fourth is government fiscal capacity, which fails precisely when it is most needed: tax revenue collapses as profits and wages fall, while spending on unemployment and rescue surges, so the state's ability to sustain a long stimulus erodes at the exact moment a long stimulus is required.

The compounding is the point. Each channel makes the others worse. The credit freeze deepens the housing decline; the housing decline deepens the credit freeze; both destroy wealth that suppresses demand; the demand collapse drains the fiscal capacity that might have arrested any of the three. This is how a market event becomes a depression, and an AI-amplified event, being deeper and faster than its predecessors, drives the compounding harder.

The Unemployment That Does Not Come Back

Recessions produce two very different kinds of job loss, and the distinction governs how recovery works. Cyclical unemployment tracks the business cycle: demand falls, firms shed workers, demand returns, workers are rehired into roughly the jobs they left. Structural unemployment is different — the job itself disappears, and no amount of returning demand brings it back.

An AI-amplified crisis is dangerous because it converts the first kind into the second. Under severe cost pressure, a firm's automation calculus changes. Positions that survived the first wave of AI adoption did so not because they were impossible to automate but because the economics were marginal — the system worked well enough, but not so much better than a human that it justified the cost and disruption of switching. Crisis pressure moves that margin. Customer service, administrative support, routine analysis, entry-level professional tasks, and layers of middle management that were retained through inertia become targets when survival demands cost cuts. And here is the trap: the firm that replaces its support desk with an AI system during the downturn does not rebuild that desk when demand recovers. It expands the system. The crisis does not pause AI-driven displacement; it fast-forwards it, compressing into eighteen months a restructuring that might otherwise have taken a decade.

This is why the recovery arithmetic is so grim. Structural unemployment does not respond to demand stimulus the way cyclical unemployment does — you can hand consumers money, but the jobs are gone by design. Displaced workers need retraining, relocation, or transition into genuinely new roles, all of which are slower, costlier, and less complete than simply being rehired. Layered on top of the ordinary cyclical losses, this structural layer stretches recovery from years toward a decade or more, and makes it more socially corrosive, because the pain does not lift when the indices rebound.

The most cited number attached to this dynamic comes from the IMF. In a 2024 address, first deputy managing director Gita Gopinath framed AI as a potential "crisis amplifier" and warned that the next major downturn could see AI substitution threaten a large share of jobs in advanced economies — a share of roughly the same order as the IMF's earlier, widely reported finding that about sixty percent of jobs in advanced economies are exposed to AI, of which roughly half face negative effects. It is essential to be honest about what this figure is and is not. It is not a measurement; it is a projection built by combining occupational exposure estimates with an assumption that crisis conditions accelerate the automation of exposed roles. Its lower bound assumes firms behave in a downturn as they always have; its upper bound assumes crisis pressure sharply accelerates adoption. The figure's real message is directional, not decimal: economic stress and AI displacement are not independent risks that happen to coincide, but forces that multiply each other.

The Collapse of Trust

For years, AI in finance would have been sold as the antidote to human frailty — more rational, less emotional, better at processing complexity than any trader. Institutions adopted it precisely because it appeared to deliver steadier returns with tighter risk control. A crisis caused by the collective behavior of those same systems would demolish that story in an afternoon, in a way that no amount of cautionary academic literature ever could.

The consequences of that demolition are not merely reputational; they are functional. Public anger, already simmering over AI-driven job losses, would find vivid vindication and drive demands for emergency restrictions or outright bans on algorithmic trading, whether or not blanket bans are wise. Financial institutions would retreat from AI out of genuine risk reassessment and fear of legal exposure. But the deeper damage lands on the institutions charged with managing the aftermath. Regulators who acknowledged the risk and failed to prevent it — Gensler's warning, unheeded, would be replayed endlessly — face a credibility collapse. Central banks whose interventions proved ineffective lose some of the market credibility on which monetary policy actually depends, since that policy works largely through expectation and belief. Recovery from any crisis requires coordinated institutional action, and coordination requires trust. An AI-caused crisis attacks exactly the institutional credibility that the recovery would need to draw upon, which is why the trust collapse is not a side effect but a core part of the mechanism by which the damage persists.

Why Prevention Fails

If the risk was foreseen and the interventions are known, the hard question is why nothing is done. The answer lies in the shape of the costs and benefits.

Consider the four interventions most often proposed. Speed limits on algorithmic trading — minimum resting times for orders, or throttles on how fast systems can react — would blunt the cascade, but they impose an immediate competitive cost on every firm that adopts them and no firm wants to be first. Diversity mandates, requiring that trading systems be designed to behave differently from one another so they do not all sell at once, attack the root cause of correlation, but they run into the objection that prescribing how a firm designs its models exceeds regulatory authority and stifles innovation — and into the sheer technical difficulty of certifying that two proprietary systems are genuinely uncorrelated. Higher capital requirements for AI-trading institutions would build buffers against algorithmic losses, but they reduce profitability and push activity toward jurisdictions that don't impose them. International coordination, which alone could close that escape route, is undermined by the very competition it would have to overcome, as financial centers vie to attract the AI and trading business that stricter rules would repel.

The pattern beneath all four is the same, and it is a general law of systemic-risk governance. The costs of prevention are immediate, concentrated, and borne by identifiable actors in the present. The benefits are delayed, diffuse, probabilistic, and spread across society and the future. A political economy facing that asymmetry systematically under-produces precaution. We have watched this movie twice already: financial regulation before 2008, where the concentrated profits of the mortgage machine drowned out diffuse warnings about systemic fragility, and pandemic preparedness before 2020, where the certain cost of stockpiles and drills lost every budget fight to more urgent claims until the diffuse risk became a concrete catastrophe. AI financial risk fits the template exactly, which is both the strongest reason to expect inaction and the clearest guide to what would have to change.

This also frames the question of responsibility. No single actor in the scenario behaves irrationally or maliciously. Each firm optimizes sensibly; each regulator faces real constraints; each nation protects its competitive position. The catastrophe is emergent — individually rational behavior aggregating into a collectively disastrous outcome. That makes blame genuinely hard to assign, and points toward a conclusion the book returns to repeatedly: where harm arises from structure rather than intent, responsibility rests less on any single actor than on the design of the system that made rational actors produce an irrational result. The firms exploited a legal structure; the regulators failed to change it; but the deeper failure belongs to a political economy that reliably declines to pay concentrated present costs for diffuse future safety.

How Confident Should We Be?

Intellectual honesty requires separating what this scenario has going for it from what it is guessing at. The cascade mechanism is not speculation. We have seen liquidity vanish and algorithms feed on one another in the 2010 Flash Crash, and we saw something adjacent in March 2020, when pandemic panic triggered market-wide circuit breakers on multiple days and even the U.S. Treasury market — the deepest, safest market on earth — briefly seized as automated and forced selling overwhelmed dealers' capacity to absorb it. These episodes confirm the ingredients: automated selling can be self-reinforcing, liquidity can evaporate at speed, and shocks can jump across markets.

What the historical record does not yet show is the full scenario. The Flash Crash reversed within an hour; the 2020 seizure was arrested by an enormous and rapid central-bank intervention. In both cases the human fire brigade reached the fire in time — which is either reassuring or ominous, depending on whether you believe the interval between alarm and collapse will stay long enough for the next intervention to work. The specific claim that convergent, AI-trained systems will produce a cascade too fast and too correlated to contain is an extrapolation from these events, not an observation of them. It is a well-motivated extrapolation, grounded in a real and growing property of the market, but it belongs in the register of informed fear rather than established fact.

And there is a genuine gap in what we can measure. We do not actually know whether financial AI systems have grown more correlated over time, because there is no standard, published metric for the convergence of proprietary trading models, and no regulator systematically tracks one. The theoretical argument for rising correlation is strong — shared data, shared architectures, shared benchmarks all push the same way — but the empirical dial that would let us watch the risk climb, and would let a regulator know when to act, does not currently exist. Building that instrument, a real-time measure of trading-system homogeneity, may be the single most useful precautionary step available, precisely because it would convert a diffuse, invisible risk into a concrete, watchable number — the kind of thing that can, occasionally, overcome the political economy of inaction.

So: where on the pre-crisis arc are we? The preconditions Gensler named are demonstrably present — a large automated share of trading, a concentrated and richly valued AI sector at the market's core, convergent system design, and a regulatory framework built for human-speed crises. What is absent is the trigger, and the trigger is by nature unpredictable. The regulatory response, meanwhile, follows the discouraging pattern the 2008 and 2020 analogies would predict: acknowledgment in speeches and reports, studies commissioned, principles articulated, and very little binding, anticipatory rule-making that would bite before the next significant stress event. Recognition is running well ahead of action, which is exactly the condition under which foreseeable crises tend to arrive.

Summary

The risk was named in advance by the regulator best placed to see it. In 2023 the SEC chair called an AI-caused financial crisis "nearly unavoidable," locating the danger not in malfunction but in uniformity — too many institutions relying on the same models and acting alike.
Systemic risk here comes from homogeneity, not failure. AI trading systems trained on similar data, built on similar architectures, and judged by similar benchmarks converge in behavior. Each is prudent in isolation; together they can sell in unison. Risk management focused on individual systems is blind to this by construction.
The cascade outruns the fire brigade. Momentum selling, liquidity withdrawal by market-making AIs, and cross-market transmission compound in milliseconds, while central-bank tools and trading halts work in hours and days. This structural speed mismatch means safeguards must be built in before a crisis, not applied after — the chapter's central regulatory conclusion.
Four channels carry the crash into ordinary life. Credit freeze, housing collapse, pension and savings destruction, and shrinking government fiscal capacity each feed the others, turning a market event into a depression.
Crisis converts cyclical job loss into permanent structural loss. Cost pressure tips the automation calculus for roles that had marginally survived, and those jobs do not return when demand does — which is why the IMF frames AI as a "crisis amplifier" and warns the next downturn could threaten a large share of advanced-economy jobs. That figure is a projection, not a measurement, and depends on how sharply crisis accelerates adoption.
Prevention is feasible but structurally unlikely. Speed limits, diversity mandates, higher capital requirements, and international coordination each attack the risk and each impose immediate, concentrated costs for diffuse, probabilistic benefits — the same asymmetry that defeated financial regulation before 2008 and pandemic preparedness before 2020.
The evidence supports the ingredients, not yet the whole recipe. The 2010 Flash Crash and the March 2020 seizure confirm that automated selling can be self-reinforcing and liquidity can vanish, but in both cases human intervention arrived in time. The full cascade is a well-grounded extrapolation, held in the register of informed fear.
We cannot currently measure the thing that matters most. There is no standard metric for how correlated financial AI systems have become, and no regulator tracks one. Building that instrument may be the most valuable precautionary step available — because a visible number is the kind of thing that can occasionally overcome the politics of inaction.

Sources

Last updated: 2026-09-13

V2 (in progress) Previous: V1