Order Is Not Control: Driven-Dissipative Response Laws Across Artificial and Biological Systems

A new study argues that imposing order on artificial intelligence systems is not the same as achieving genuine control, proposing instead a physics-inspired framework of driven-dissipative response laws to describe how interventions succeed or fail across both biological and machine-learning panels [1]. The preprint, authored by Lap-Hang Ho and posted to the arXiv repository, contends that many techniques in AI alignment, interpretability, and neural steering merely identify order-inducing objects without establishing true control [1]. Control, the paper asserts, requires a receiver-gated response law — a denominator-indexed operator that maps material state, action, bath, and receiver state onto response displacement, sinks, effort, and basin projection [1]. The work draws on evidence from mouse anterior lateral motor cortex, C. elegans, and zebrafish preparations to provide physical response-operator measurements while explicitly excluding conclusions about coordinate identity or the presence of a controller [1]. The framework treats interventions as local and conditional: a given drive can be admitted, saturated, sign-changing, leaky, or overdriven depending on the medium, bath, receiver state, action port, and comparator [1]. Control is formally assigned only when finite effort moves a target or outcome-readout class under the same denominator while damage, null responses, evasive outputs, invalid formats, overdrive, and unnecessary effort remain bounded [1]. This formulation echoes principles from deterministic dynamical systems, where small changes in initial conditions can produce widely divergent outcomes, a hallmark of chaos theory [3]. Large language model panels provided a testbed for the generated-output response laws. Across four material conditions, response vectors were predictable at 72.8–73.7% component-sign accuracy, rising to 84.3–84.8% when considering only nonzero components [1]. Held-out observers predicted system-effect families at 93.6% accuracy and target-oracle families at 91.7% accuracy [1]. The study also found that constitution-conditioned adapters reshape susceptibility, functioning as prepared media, while stochastic-operator panels separate measured opportunity from deployable action policies [1]. The paper’s mesoscopic account treats drives as acting through prepared media, baths, and receivers to produce admitted movement, impedance, sinks, or overdrive [1]. It explicitly leaves several questions outside its scope, including deployable pre-generation control, hidden or logit-level causal sufficiency, biological-to-LLM coordinate identity, and literal thermodynamic quantities [1]. The relationship between order and entropy has a long intellectual history, stretching from Erwin Schrödinger’s 1944 argument that life feeds on negative entropy to modern quantitative applications of Gibbs free energy in cellular metabolism [4]. The current work extends this lineage into the domain of artificial neural networks, which loosely model biological neurons and have become the basis of large language models through transformer architectures [5][9]. The preprint appeared on arXiv, an open-access repository that hosts scientific papers across physics, computer science, and related fields without peer review, and which now receives roughly 24,000 submissions per month [7].