Reflexivity (Soros)

The protagonist here is not an academic but a speculator who made billions of dollars on the idea. George Soros calls his investment philosophy reflexivity, and the core sentence is: your view of the market is necessarily wrong, but once you act on your wrong view you really do change the market, and the changed market then feeds your next wrong view. Ideas and reality chase each other, with no end point. This framework is stronger than the self-fulfilling prophecy you learned in R02, and it has an unexpected ending: long criticised as unfalsifiable, it acquired a mathematical form in machine learning after 2020.

The framework stacks two principles. First, fallibility: a person's overall judgment about the world can never be entirely correct; you can hold individual facts, but the moment you form a theory or a global view, your perspective is necessarily biased, or inconsistent, or both. Second, reflexivity: these imperfect views change, through your actions, the very objects they describe. Together Soros calls them the "human uncertainty principle."

The engine is a distinction he draws inside "thinking." The cognitive function: reality shapes ideas, you observe the market and form a judgment. The manipulative function: ideas shape reality, you buy and sell on that judgment, and the buying and selling change the thing you were trying to assess. Both directions run at once, each feeding the other, and in Soros's words the relation "is circular or recursive." This is stronger than the self-fulfilling prophecy: the prophecy in R02 is a single pass, predict, act, come true, loop closed; reflexivity is continuous two-way coupling, with no foothold anywhere in the loop that is uncontaminated by belief.

Everything turns on the sign of the feedback. Negative feedback is ideas and reality converging: you overestimated, reality corrects you, you revise, the system tends toward equilibrium, and that is the world textbook economics assumes. Positive feedback is ideas and reality pushing each other away: you are bullish, you buy and push prices up, the rise confirms your judgment, you become more bullish. Self-reinforcing, but eventually self-defeating, because an expansion detached from reality cannot go on forever. Soros's bubble formula: "Every bubble has two components: an underlying trend that prevails in reality and a misconception relating to that trend," the two reinforcing each other. The full sequence is: a trend starts, perception lags behind it, the two accelerate each other, then comes the "moment of truth" when reality can no longer support expectations, followed by a plateau, a reversal of belief, and a crash. From this he directly rejects the efficient market hypothesis and rational expectations: markets do not tend toward equilibrium, they are endogenously unstable, and equilibrium is only one of two phases.

The idea is not an isolated one. Merton's 1948 paper on the self-fulfilling prophecy also hides its frequently dropped mirror image: the self-defeating prophecy, where a prediction that might have come true triggers, by being believed, the actions that prevent it, and so falsifies itself. A warning prompts precautions, the crisis does not happen, and the warning "looks" wrong. The philosopher Popper independently named the same structure the Oedipus effect: the oracle itself brought about its own fulfilment. Line them up: Soros's positive feedback corresponds to the amplifying direction of the self-fulfilling prophecy; negative feedback corresponds to the cancelling direction of the self-defeating prophecy. In R07 (how theory shapes markets) these two directions are called Barnesian performativity and counterperformativity. The same feedback structure was discovered independently in sociology (1948), philosophy of science (1936 to 1957) and finance (1987). Soros's distinctive contribution is fitting both axes into one dynamic model: the same bubble inflates under positive feedback and then collapses under negative feedback, whereas academics usually handle amplification and cancellation separately.

The heaviest criticism reflexivity takes is "insightful but unfalsifiable": in 2013 the Journal of Economic Methodology devoted a whole special issue to it, with 18 economists and philosophers of science weighing in, and many noted that it offers no criterion for telling a reflexive market from a non-reflexive one. That criticism got a partial answer after 2020, from machine learning. Perdomo and colleagues introduced performative prediction: deploying a predictive model changes the distribution of the data, so the distribution itself becomes a function of your decision. This creates a fork: a model's "optimum" should be relative to some distribution, but the distribution changes with the model, so change the model and there is no single "optimum" left to speak of. The framework therefore gives two kinds of "solution." One is a stable point: the model is optimal on the distribution it itself induces, and repeated retraining converges here. The other is an optimal point: the true optimum once "I will change the world" is counted in. The headline theorem is that stable is not optimal, and a stable point can be arbitrarily far from the optimal one. In plain terms: a self-confirming equilibrium can be an arbitrarily bad equilibrium, and "retraining converged" does not mean "the outcome is good." Better still, the convergence condition was written down: when the sensitivity of the data to the model, that is the strength of reactivity, is below a threshold set by the shape of the loss function, repeated retraining converges steadily; above the threshold, retraining oscillates and diverges. Convergence looks exactly like negative feedback: ideas draw closer to reality and settle down. Divergence looks exactly like positive feedback: ideas push reality further and further away and never settle. That is Soros's equilibrium phase and far-from-equilibrium phase, with a dividing line you can actually write down for the first time. Two later additions: in 2021 someone gave the stable point an accurate nickname, the echo chamber, since a naive optimiser only ever sees the distribution it induces; and in 2023 Piliouras and Yu proved that when multiple learners are reflexive at once, the dynamics can drop straight from convergence into chaos.

One open question: Soros can only narrate bubbles after the fact, and performative prediction supplies an ex ante convergence threshold, but the two languages have not been joined; and real-world distributions depend not only on the current model but on history, and convergence theory for this kind of reflexivity with memory is far from complete. Turning "is this AI narrative in the equilibrium phase or the far-from-equilibrium phase" into an estimable quantity is the most worthwhile work on this line.

The one-line takeaway: when your judgment can change the thing being judged, right and wrong stop being between you and the world, and become between you and the world you caused.