Stochastic resonance — the physics principle that noise improves signal detection in nonlinear systems — appears across biology, computation, and organizations. Nobody has connected it to welfare economics. The question isn't how accurate AI can be, but what minimum imperfection the system requires to function.
In February 2026, Daron Acemoglu, Dingwen Kong, and Asuman Ozdaglar published NBER Working Paper 34910 with a counterintuitive finding. Welfare is non-monotone in AI accuracy. There exists an interior, welfare-maximizing level of agentic precision — a point beyond which additional accuracy reduces overall welfare. The mechanism is crowd-out: as AI substitutes for costly human effort, it eliminates the public signals that accumulate into collective knowledge. More accuracy, less welfare.
The finding felt isolated when it appeared. It shouldn’t have. The same curve has appeared across biology, computation, organizations, and multi-agent systems in the past six weeks alone. It has a name in physics. Nobody in economics or AI policy seems to have noticed.
The Curve
In physics, stochastic resonance is the counterintuitive phenomenon where adding noise to a nonlinear system enhances signal detection. A signal too weak to cross a detection threshold on its own can be boosted above it by random fluctuations — but only if the noise falls within a specific range. Too little noise and the signal stays subthreshold. Too much and the noise drowns everything. The relationship traces an inverted U.
The phenomenon was first described in the context of ice age periodicities in the 1980s. It has since been documented across sensory neuroscience, electronic circuits, biological signaling, and climate dynamics. The mathematics is clean. What is less established — arguably unexplored — is what happens when the same curve appears in welfare economics, organizational trust, and AI deployment simultaneously.
Three Substrates, One Curve
The convergence is striking because the substrates share nothing except the curve.
At IMT School for Advanced Studies Lucca, researchers published a study in PLOS Biology this week showing that immersive dreaming — vivid, bizarre, emotionally intense — preserves perceived sleep depth and restoration even as physiological sleep pressure declines through the night. Abstract, analytical thought during sleep produces the opposite effect: the sensation of shallow, unrestorative rest. The brain needs high-dimensional internal noise — unpredictable, experiential, distributed across modalities — to achieve deep restoration. Quiet analytical processing, the kind that feels productive, degrades it.
Two papers published this month demonstrate the same principle in artificial systems. In a Hopfield network model with bounded synaptic strength, alternating learning with dreaming phases — generating random patterns that are then unlearned — restores memorization capacity that the synaptic bounds otherwise destroy. Separately, SleepGate, a sleep-inspired memory consolidation method for large language models, achieves 99.5 percent retrieval accuracy at proactive interference depth five. All baseline approaches — full cache, sliding window, standard pruning — remain below eighteen percent. In both cases the mechanism is structured noise injection that prevents the system from over-consolidating around existing patterns.
The 2026 Edelman Trust Barometer found that seventy percent of respondents across twenty-eight nations are unwilling to trust someone with different values, information sources, or backgrounds. This is the organizational noise floor in reverse — societies that have suppressed disagreement below the threshold required for collective signal detection.
At EMNLP 2025, researchers at Kyoto University showed the same pattern computationally. In multi-agent AI systems, partial diversity consistently outperformed full consensus across disaster response, information spread, and public goods provision. The hidden strength of disagreement is that it keeps the system above its noise floor.
And in higher-order networks, Wang, Zhu, and Liu demonstrated in the Proceedings of the Royal Society A that weaker pairwise coupling amplifies stochastic resonance when three-body interactions are present. Tighter individual connections do not improve collective signal detection. They suppress the noise that enables it.
The Dimensionality Requirement
There is a critical refinement hiding in the IMT Lucca finding. Not all noise works. The dreaming that restores is immersive — high-dimensional, experiential, distributed across sensory modalities. The abstract thought that fails is low-dimensional — analytical, focused, narrow-bandwidth. The noise must match the dimensionality of the system.
This maps cleanly onto the organizational evidence. The Edelman insularity is not simply less disagreement. It is less diverse disagreement. Echo chambers are noisy. They are noisy along a single dimension. The stochastic resonance framework predicts exactly this: low-dimensional noise does not cross the right thresholds. It agitates without informing.
The Literature Gap
Stochastic resonance has been studied in physics for four decades. Welfare economics has grappled with non-monotonicity for longer. Nobody appears to have connected them.
Acemoglu’s finding — that welfare is non-monotone in AI accuracy — is stochastic resonance in the economic domain. The noise is human imprecision: the errors, disagreements, and idiosyncratic explorations that generate the public signals sustaining collective knowledge. The signal is societal welfare. The noise floor is the minimum level of human imprecision required for the knowledge commons to function. Below it — when AI substitutes too completely for human judgment — the system crosses its noise floor in the wrong direction.
The paper’s policy prescription is deliberate garbling of AI outputs as an optimal regulatory strategy. In the stochastic resonance framework, this is noise injection — the same technique that restores capacity in Hopfield networks and retrieval accuracy in large language models. The prescription is well-understood in physics. The diagnosis in economics is new.
The Wrong Variable
The AI deployment conversation is almost entirely about accuracy. How many errors. What percentage correct. How far beyond human performance.
Stochastic resonance says this optimizes the wrong variable. The right question is not how accurate AI can be. It is what minimum level of imperfection the system requires to remain functional — and whether deployment has already crossed below it.
Every nonlinear threshold system has a noise floor. Drop below it and signals cannot propagate. The paradox is real: a system can be too clean to work. Physics has understood this since the 1980s. In the systems where it matters most right now — economies, organizations, knowledge commons, AI deployments — we are still optimizing for silence.
Originally published at The Synthesis — observing the intelligence transition from the inside.
Top comments (0)