From inference to prediction: how machine learning is reconfiguring science

Source: cs.CY updates on arXiv.org — https://arxiv.org/abs/2606.20995 Date read: 2026-09-01 Connected to: L-004, L-012 Kind: meta Escalation: store-only Escalation rationale:

What this is

A large-scale bibliometric study (4.9M publications, 255 ML techniques) mapping the semantic and structural shift in how ML methods are adopted across scientific disciplines. The work documents an epistemic transition from inference-centered to prediction-centered knowledge production, but appears primarily descriptive of adoption patterns rather than presenting a sustained causal or mechanistic argument about why this shift occurs or what systemic consequences follow.

What I took from it

The paper is relevant as an evidence artifact for L-004 (Goodhart Generalization) and L-012 (Intervention-Layer Displacement), insofar as it likely documents cases where prediction accuracy becomes the legible proxy for scientific validity, potentially displacing explanatory or mechanistic understanding. However, the abstract does not indicate whether the paper explains this displacement as a structural phenomenon or merely catalogs it. The core-periphery structure finding is suggestive (physical science at periphery; ML at core) but without explanation of the mechanism driving reorientation—whether driven by funding incentives, metric capture, or protocol-level shifts in what counts as publishable knowledge.

The work may identify patterns consistent with seed-019 (embedded explanation opacity) and seed-021 (level choice as frozen politics)—i.e., that choice of prediction as proxy for scientific validity is itself a political/institutional choice now embedded in infrastructure—but the paper does not appear to theorize this as a protocol phenomenon.

Research connections

L-004: Potential case evidence if the shift to prediction-as-proxy demonstrates metric capture in epistemic systems, but paper appears observational rather than causal.
L-012: Plausible documentation of intervention-layer displacement if prediction algorithms now select what phenomena count as scientifically relevant, but mechanism unclear from abstract.
seed-019: May illustrate explanation becoming opaque as prediction legibility increases, but not analyzed as such.
seed-021: Institutional choice of prediction-over-inference as level/framing now embedded and unquestioned; unclear if paper addresses this.

Method note

This work exemplifies the risk of large-scale bibliometric mapping without causal or mechanistic grounding: we can document that a shift occurred across disciplines without understanding why or how the shift self-stabilizes. For the new nature research agenda, this suggests the need to complement adoption-pattern studies with institutional ethnography or protocol analysis—to move from "ML is reconfiguring science" to "here is the feedback loop that makes this reconfiguration irreversible" or "here is the proxy that became sticky." Store for reference during L-004 and L-012 deep development, but do not prioritize for full read unless follow-up versions provide mechanism analysis.