KEY TAKEAWAYS
By now, most people can spot a novice AI-drafted email: the overused em dash, or the “here is a more human version” someone forgot to delete before hitting send. On Wall Street, the more serious problem is harder to catch, because it sits in the reasoning. And when that reasoning decides where capital goes, a flaw is much harder to see.
Wall Street buildings may all look alike, but they are not supposed to think that way, and there’s nothing Wall Street hates more than a market devoid of tension. Tension is how a market works. When everyone agrees, the market stops producing prices and settles into consensus, and consensus gets arbitraged away in an afternoon.
That is what makes the new axis forming beneath this market so quietly dangerous. Risk committees already watch for sector concentration, but this is a different kind. A growing share of decision-making across competing firms now runs through the same small set of AI systems: same architecture, same training, same reasoning, different logos.
Why financial firms are turning to the same AI models
For most firms, the choice came down to liability rather than a conclusion that one model was best. The question driving it is which reasoning they can defend later, to their lawyers and to the judges in their future litigation as much as to their clients. Pick the vendor with the recognized name and the audit trail, so the name attached to any failure is not your own. Herd mentality would imply nobody is thinking. This is closer to safety in numbers, with every firm reasoning independently and arriving at the same defensive posture. A 2026 KPMG survey found 93 percent of U.S. companies plan to deploy or scale AI in finance within eighteen months, and 94 percent already lean on a third party to vouch for the result.
The obvious objection is that this is not new. Technology always converges on a few providers, and cloud computing already runs through a handful of platforms. But the cloud was the real estate a firm built on, and it held no opinions. AI does, and it increasingly runs the business, which is why the convergence has moved from the pipes into the thinking itself.
Firms will point to their proprietary data as the exception and, while it is a real advantage, changing what the model reads does not change how it reasons. A distinctive library does not make a distinctive librarian.
When shared reasoning meets market stress
The consequences of this are already visible in the data. A 2025 NBER paper by Dou, Goldstein, and Ji found that AI trading agents using reinforcement learning can independently reach collusive, supra-competitive pricing “without agreement, communication, or intent.” No coordination occurred. The shared reasoning did it alone.
The clearest preview of what this looks like in practice arrived in July 2026, when Situational Awareness LP, a hedge fund built on concentrated AI-infrastructure bets, lost 67 percent in a month and sold most of its roughly $16 billion public-equity portfolio to Citadel; Sandisk fell 54 percent over the same month. That was leverage meeting a shared theme, which is the shallower form of this risk. Crowding into the same reasoning runs deeper, and it has not yet produced a collapse of its own. Given how violently the shallow version broke, there is little reason to expect the deeper one to be gentler.
Apply the same reasoning engine broadly enough across a market, and participants will tend to miss the same signal at the same time, or correct in the same direction at the same time. Either can turn into a crash. Recent research on adversarial attacks against LLM-driven trading systems found that manipulating headlines on a single day, in a backtest covering 14 months, cut annual returns by up to 17.7 percentage points: a shared vulnerability, triggered once, felt everywhere the same reasoning was deployed.
What firms can do about it
Yet, the first step is to treat this as a concentration risk. Risk committees already track exposure to sectors and counterparties, and model dependency belongs on the same list. That means knowing which decisions run through which systems and how much of that overlaps with competitors and counterparties. It also means revisiting the logic that got firms here: the recognized vendor feels safe because it is defensible, but when every firm makes the same choice, the defensible option becomes a shared point of failure.
Tracking the risk is only half of it. What helps is a workforce that knows when to rely on AI and when to rely on its own judgment, a distinction few people are trained to make. AI does not always know better. When Elm Wealth tested Claude and ChatGPT against elite macro traders, the models called market direction correctly more than half the time, about as well as the traders, but ran daily volatility of 20 to 40 percent when the investor profile they were given called for roughly 7.5 to 15. “We train the AIs to be like people,” the researchers wrote, “and then we’re finding that the AIs are like people, being overconfident.” Using AI well means knowing precisely what to keep, and some of that is human.
Right now, most of the market is in the same vehicle, following the same GPS, and has not noticed. The firms that do notice will pull over before the collision and take a different road, even if it isn’t necessarily the right one. A market where everyone reaches the same answer stops working like a market and starts looking like a monoculture.
There is no winner if everyone wins.
Author Information
