From Inference to Prediction: How Machine Learning is Reconfiguring Science (1990-2025)

Diego Kozlowski; Malena Mendez Isla; Vincent Lariviere

arxiv: 2606.20995 · v1 · pith:E6D655DBnew · submitted 2026-06-19 · 💻 cs.CY

From Inference to Prediction: How Machine Learning is Reconfiguring Science (1990-2025)

Malena Mendez Isla , Vincent Lariviere , Diego Kozlowski This is my paper

Pith reviewed 2026-06-26 13:12 UTC · model grok-4.3

classification 💻 cs.CY

keywords machine learningscientific practiceinferencepredictionepistemic opacityhealth sciencessocial sciencesdeep learning

0 comments

The pith

Machine learning displaces inference-oriented methods with predictive architectures in health and social sciences across two distinct waves since 2015.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper maps the spread of machine learning across 4.9 million scientific publications from 1990 to 2025 using a taxonomy of techniques and semantic analysis. It claims that predictive approaches are steadily replacing inferential ones in fields that once emphasized interpretability, most clearly in health sciences and social sciences. The change occurs in two phases, the first powered by deep learning and the second by systems supplied through external companies. A sympathetic reader would care because the shift alters both what science can measure and the standards by which its claims are accepted or challenged. The result is greater analytical reach accompanied by new forms of opacity that researchers cannot fully inspect or report.

Core claim

A hierarchical taxonomy of 255 ML techniques and embedding-based semantic mapping applied to OpenAlex publications reveals a core-periphery structure in which physical sciences anchor the methodological core while health sciences show the largest growth. Predictive techniques cluster in computer science while inferential approaches remain distributed across applied domains. In health and social sciences the paper documents a displacement of inference by prediction that unfolds in two waves: the first (2015-2021) driven by deep learning architectures that lower predictive error yet increase epistemic opacity, and the second (post-2022) organized around a small set of architectures delivered b

What carries the argument

The hierarchical taxonomy of 255 ML techniques together with embedding-based semantic mapping, which distinguishes inferential from predictive approaches and tracks changes in validation regimes across disciplines.

If this is right

Analytical capacity expands in health and social sciences as predictive architectures spread.
The first wave reduces predictive error while introducing epistemic opacity through deep learning.
The second wave adds opacity over inaccessible data and processes supplied by external companies.
Validation regimes that once differed across domains are being reorganized around predictive performance.
Scientific knowledge is now produced and evaluated under conditions that include components researchers cannot inspect or report.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

Reproducibility standards may need revision when key pipeline elements lie outside researcher control.
Physical sciences, positioned at the methodological core, may show slower adoption of the same displacement pattern.
Funding and data-sharing policies could be adjusted to address the growing role of external company architectures.
Similar semantic-mapping methods could test whether the two-wave pattern appears in additional disciplines beyond those examined.

Load-bearing premise

The taxonomy of ML techniques and the semantic mapping method correctly separate inferential from predictive approaches and detect genuine changes in how disciplines validate results.

What would settle it

A re-analysis of the same publication corpus that finds no net replacement of inferential by predictive techniques in health and social sciences after 2015, or that shows no measurable difference in opacity sources between the 2015-2021 and post-2022 periods.

Figures

Figures reproduced from arXiv: 2606.20995 by Diego Kozlowski, Malena Mendez Isla, Vincent Lariviere.

**Figure 1.** Figure 1: (A) Absolute temporal evolution of the corpus. (B) Relative share of ML publications in OpenAlex by domain and in total over time. (C) Relative share of ML publications in OpenAlex by field (left X axis) and Percentage of publications in ML corpus per field (right X axis). Data for 2002 is influenced by Crossref reindexing effects and the inclusion of backlogged 1990s proceedings. These metadata inconsiste… view at source ↗

**Figure 2.** Figure 2: maps topics and fields according to the content of their papers’ titles and abstracts (see Methods section). The spatial proximity between topics and fields represent their semantic similarity, while the absolute orientation in the axis is arbitrary in UMAP. Topics and fields are coloured according to the domain. The domains of physical sciences (blue) and health sciences (pink) are well defined with conto… view at source ↗

**Figure 3.** Figure 3: ML Semantic Space: Methodological Overlay and Epistemic Objectives (1990- 2025). This map visualizes the distribution of all 255 machine learning techniques across the scientific landscape. Each triangle represents the semantic barycenter of a technique, colorcoded by its epistemic objective: inference, prediction, or other. The top 10 most frequent techniques are explicitly highlighted with labels. Furth… view at source ↗

**Figure 5.** Figure 5: Evolution of ML and DL Techniques. (A) Temporal trends for the top 10 machine learning techniques (1990–2025). (B) Temporal trends for the top 10 deep learning architectures (2012–2024). (C) Proportion of publications in the corpus for the top 20 techniques (absolute counts are indicated in parentheses). Percentages in panels A and B are computed using document-level fractional counting, where each publica… view at source ↗

**Figure 6.** Figure 6: Temporal evolution of the top 10 ML techniques across scientific domains (1990– 2024). The vertical axis represents the yearly share of each technique within a domain’s total output. Percentages are computed using fractional counting to account for multi-label cooccurrence and to reflect relative keyword prominence within articles. From Inference to Prediction [PITH_FULL_IMAGE:figures/full_fig_p018_6.png] view at source ↗

**Figure 7.** Figure 7: Relative Growth Rate of and Machine Learning Techniques across Scientific Domains (1990–2024). The RGR compares the year-over-year growth of a keyword’s share within our AI corpus against the baseline year-over-year growth of the entire corresponding scientific domain in the OpenAlex database. The dashed horizontal line at RGR = 1 denotes growth parity. When a line is above this threshold (RGR > 1), it ind… view at source ↗

read the original abstract

Machine learning (ML) has reshaped scientific practice across disciplines, yet its epistemic consequences remain poorly understood. This paper analyzes how its broad diffusion reconfigures the conditions under which scientific claims are produced and evaluated. Using a hierarchical taxonomy of 255 ML techniques and embedding-based semantic mapping, we analyze 4.9 million scientific publications from OpenAlex (1990-2025). We reconstruct the semantic space of ML research and show a core-periphery structure, with physical sciences forming the methodological core and health sciences representing the primary growth area. We identify distinct methodological profiles across domains: predictive techniques concentrate in computer sciences while inferential approaches remain distributed across applied fields, reflecting historically differentiated validation regimes. We observe the displacement of inference-oriented techniques by predictive architectures in domains that have traditionally prioritized interpretability-most notably health sciences and social sciences. This displacement unfolds in two qualitatively distinct waves. The first (2015-2021) was driven by deep learning architectures that reduced predictive error while introducing epistemic opacity. The second (post 2022) is organized around a small number of architectures delivered through external companies, introducing a further layer of opacity over data and processes that researchers cannot access or report. This transformation expands the analytical capacity of science, and also reorganizes the conditions under which scientific knowledge can be produced and evaluated.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

Large-scale corpus shows two waves of predictive ML displacing inference in health and social sciences, but the 255-technique taxonomy has no reported validation.

read the letter

The main takeaway is a bibliometric mapping of 4.9 million papers that splits ML adoption into two waves of displacement from inference to prediction, first via deep learning 2015-2021 and then via external company architectures after 2022, concentrated in health and social sciences. The work also flags a core-periphery pattern with physical sciences at the center.

It does the scale part cleanly and the core-periphery observation is a useful descriptive cut. Distinguishing predictive techniques that cluster in computer science from more distributed inferential ones across applied fields lines up with known differences in how fields validate claims.

The soft spot is the taxonomy. The abstract gives no numbers on how the 255 techniques were labeled inferential versus predictive, no inter-rater checks, and no sensitivity tests on the embedding step. Without those, the two-wave story could shift if the category boundaries move. The stress-test note on this point holds up from the abstract alone.

This is for people working in science studies or quantitative STS who want data on how ML changes what counts as evidence in applied domains. A reader already tracking epistemic opacity in medicine or policy-facing social science could pull the patterns even if the causal story needs tightening.

Send it to peer review. The corpus size and the question are worth referee time, but the methods section will have to show the taxonomy was built and tested in a way that survives scrutiny.

Referee Report

2 major / 1 minor

Summary. The paper claims that analysis of 4.9 million OpenAlex publications (1990-2025) via a hierarchical taxonomy of 255 ML techniques and embedding-based semantic mapping reveals a core-periphery structure in ML research, with predictive techniques concentrating in computer science and inference-oriented ones distributed across applied fields. It identifies displacement of inference by predictive architectures in health and social sciences unfolding in two waves (2015-2021 deep learning; post-2022 external company architectures), expanding capacity while increasing epistemic opacity over data and processes.

Significance. If the taxonomy and mapping are robust, the work supplies a large-scale empirical reconstruction of ML diffusion and its effects on validation regimes across disciplines. The 4.9M-paper corpus and identification of temporally distinct waves constitute a data-driven contribution to understanding how ML reorganizes scientific knowledge production. The embedding approach to semantic mapping is a methodological strength for tracing technique profiles.

major comments (2)

[Methods] Methods (taxonomy and labeling): The hierarchical taxonomy of 255 ML techniques is used to partition inferential from predictive approaches and to identify the two displacement waves; however, no inter-rater reliability statistics, external validation against expert labels, ablation on boundary definitions, or sensitivity tests on the embedding mapping are reported. This directly undermines the link between observed corpus patterns and the claimed epistemic regime shifts.
[Results] Results (wave identification): The claim that the post-2022 wave is 'organized around a small number of architectures delivered through external companies' and introduces a further layer of opacity requires explicit quantification of company-architecture prevalence and a demonstration that the embedding distances preserve epistemic (rather than merely technical) distinctions; without these, the qualitative distinction between the two waves rests on untested classification choices.

minor comments (1)

[Abstract] Abstract: The phrase 'reconstruct the semantic space of ML research' is used without a forward reference to the specific embedding method or dimensionality reduction technique employed.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive and detailed comments, which identify key areas where additional rigor can strengthen the manuscript. We address each major comment below, indicating planned revisions where appropriate.

read point-by-point responses

Referee: [Methods] Methods (taxonomy and labeling): The hierarchical taxonomy of 255 ML techniques is used to partition inferential from predictive approaches and to identify the two displacement waves; however, no inter-rater reliability statistics, external validation against expert labels, ablation on boundary definitions, or sensitivity tests on the embedding mapping are reported. This directly undermines the link between observed corpus patterns and the claimed epistemic regime shifts.

Authors: We acknowledge that the manuscript does not include inter-rater reliability statistics or external expert validation for the taxonomy. The taxonomy was derived from a systematic synthesis of existing ML classifications in the literature. In the revised manuscript we will add sensitivity tests on the embedding mapping parameters and ablation analyses on the inferential/predictive boundary definitions. A full external inter-rater study lies beyond the scope of the current revision and will be noted as a limitation. revision: partial
Referee: [Results] Results (wave identification): The claim that the post-2022 wave is 'organized around a small number of architectures delivered through external companies' and introduces a further layer of opacity requires explicit quantification of company-architecture prevalence and a demonstration that the embedding distances preserve epistemic (rather than merely technical) distinctions; without these, the qualitative distinction between the two waves rests on untested classification choices.

Authors: We agree that explicit quantification is required. The revised results section will report quantitative prevalence statistics for company-associated architectures in the post-2022 period, derived from affiliation and technique co-occurrence data in the corpus. We will also add analysis correlating embedding distances with paper-level indicators of validation practices to demonstrate that the mappings capture epistemic rather than purely technical distinctions. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical corpus analysis with independent classification step

full rationale

The paper conducts a large-scale empirical mapping of 4.9M publications using a pre-defined hierarchical taxonomy of 255 ML techniques plus embedding-based semantic analysis. No equations, fitted parameters, or first-principles derivations are presented whose outputs reduce to the authors' own inputs by construction. The distinction between inferential and predictive techniques is a methodological labeling choice applied to the corpus; the reported waves (2015-2021 deep learning, post-2022 external architectures) are observational trends in that labeled data rather than quantities forced by self-definition or self-citation chains. The analysis is self-contained against external benchmarks (OpenAlex corpus) and does not rely on load-bearing self-citations or uniqueness theorems imported from the authors' prior work.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The central claim rests on the untested premise that the chosen taxonomy and embeddings faithfully separate inference from prediction and that OpenAlex coverage is representative of epistemic practice.

axioms (2)

domain assumption OpenAlex database provides a representative sample of scientific publications from 1990-2025
All quantitative claims depend on this data source being unbiased in coverage and metadata quality.
domain assumption The 255-technique taxonomy and embedding mapping distinguish inferential from predictive validation regimes
This mapping is required to interpret the observed displacement as an epistemic shift rather than a surface trend in terminology.

pith-pipeline@v0.9.1-grok · 5775 in / 1376 out tokens · 26290 ms · 2026-06-26T13:12:40.397197+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

11 extracted references · 5 canonical work pages

[1]

(2018) The locus of legitimate interpretation in Big Data sciences: Lessons for computational social science from -omic biology and high-energy physics

Bartlett A, Lewis J, Reyes-Galindo L, et al. (2018) The locus of legitimate interpretation in Big Data sciences: Lessons for computational social science from -omic biology and high-energy physics. Big Data & Society 5(1): 2053951718768831. Benz P, Pradier C, Kozlowski D, et al. (2025) Mapping the unseen in practice: comparing latent Dirichlet allocation ...

2018
[2]

Journal of Machine Learning Research 3: 993–1022

23 Blei D, Ng A and Jordan M (2003) Latent Dirichlet Allocation. Journal of Machine Learning Research 3: 993–1022. Borgman CL and Brand A (2024) The Future of Data in Research Publishing: From Nice to Have to Need to Have? Harvard Data Science Review (Special Issue 4). Epub ahead of print 2 April

2003
[3]

Borgohain DJ, Bhardwaj RK and Verma MK (2022) Mapping the literature on the application of artificial intelligence in libraries (AAIL): a scientometric analysis

DOI: 10.1162/99608f92.b73aae77. Borgohain DJ, Bhardwaj RK and Verma MK (2022) Mapping the literature on the application of artificial intelligence in libraries (AAIL): a scientometric analysis. Library Hi Tech 42(1): 149–179. Breiman L (2001) Statistical Modeling: The Two Cultures. Statist. Sci. (16(3)). Epub ahead of print

work page doi:10.1162/99608f92.b73aae77 2022
[4]

Brown TB, Mann B, Ryder N, et al

DOI: 10.1214/ss/1009213726. Brown TB, Mann B, Ryder N, et al. (2020) Language Models are Few-Shot Learners. arXiv:2005.14165. arXiv. Available at: http://arxiv.org/abs/2005.14165 (accessed 2 March 2026). Burrell J (2016) How the machine ‘thinks’: Understanding opacity in machine learning algorithms. Big Data & Society 3(1): 2053951715622512. Bzdok D, Altm...

work page doi:10.1214/ss/1009213726 2020
[5]

Scientometrics 130(9): 5093–5114

Ding L, Lawson C and Shapira P (2025) Rise of Generative Artificial Intelligence in Science. Scientometrics 130(9): 5093–5114. Ding L, Lawson C and Shapira P (2026) Tracking AI’s Scientific Anatomy: A Novel Framework for Analyzing the Use and Diffusion of AI in Science. Epub ahead of print

2025
[6]

(2024) Oil & Water? Diffusion of AI Within and Across Scientific Fields

Duede E, Dolan W, Bauer A, et al. (2024) Oil & Water? Diffusion of AI Within and Across Scientific Fields. arXiv:2405.15828. arXiv. Available at: http://arxiv.org/abs/2405.15828 (accessed 25 February 2026). Dwivedi R and Elluri L (2024) Exploring Generative Artificial Intelligence Research: A Bibliometric Analysis Approach. IEEE Access 12: 119884–119902. ...

arXiv 2024
[7]

Hastie T, Tibshirani R and Friedman J (2017) The Elements of Stadistical Learning: Data Mining, Inference, and Prediction

DOI: 10.1038/s41586-025-09922-y. Hastie T, Tibshirani R and Friedman J (2017) The Elements of Stadistical Learning: Data Mining, Inference, and Prediction. Springer International Publishing. Available at: https://www.sas.upenn.edu/~fdiebold/NoHesitations/BookAdvanced.pdf. He K, Zhang X, Ren S, et al. (2015) Deep Residual Learning for Image Recognition. In...

work page doi:10.1038/s41586-025-09922-y 2017
[8]

Iliadis A and Russo F (2016) Critical data studies: An introduction

DOI: 10.1007/s00146-025-02835-4. Iliadis A and Russo F (2016) Critical data studies: An introduction. Big Data & Society 3(2): 2053951716674238. Ioffe S and Szegedy C (2015) Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift. Epub ahead of print

work page doi:10.1007/s00146-025-02835-4 2016
[9]

(2025) Application of artificial intelligence in academic libraries: a bibliometric analysis and knowledge mapping

Islam MN, Ahmad S, Aqil M, et al. (2025) Application of artificial intelligence in academic libraries: a bibliometric analysis and knowledge mapping. Discover Artificial Intelligence 5(1):

2025
[10]

(2023) An Introduction to Statistical Learning with Applications in R

James G, Witten D, Hastie T, et al. (2023) An Introduction to Statistical Learning with Applications in R. Second Edition. Springer. Available at: https://www.tandfonline.com/doi/full/10.1080/24754269.2021.1980261 (accessed 6 December 2025). Jordan MI and Mitchell TM (2015) Machine learning: Trends, perspectives, and prospects. Science Vol. 349(Issue 6245...

work page doi:10.1080/24754269.2021.1980261 2023
[11]

Expert Systems with Applications 186: 115728

Su M, Peng H and Li S (2021) A visualized bibliometric analysis of mapping research trends of machine learning in engineering (MLE). Expert Systems with Applications 186: 115728. Van Der Vlist F, Helmond A and Ferrari F (2024) Big AI: Cloud infrastructure dependence and the industrialisation of artificial intelligence. Big Data & Society 11(1): 2053951724...

Pith/arXiv arXiv 2021

[1] [1]

(2018) The locus of legitimate interpretation in Big Data sciences: Lessons for computational social science from -omic biology and high-energy physics

Bartlett A, Lewis J, Reyes-Galindo L, et al. (2018) The locus of legitimate interpretation in Big Data sciences: Lessons for computational social science from -omic biology and high-energy physics. Big Data & Society 5(1): 2053951718768831. Benz P, Pradier C, Kozlowski D, et al. (2025) Mapping the unseen in practice: comparing latent Dirichlet allocation ...

2018

[2] [2]

Journal of Machine Learning Research 3: 993–1022

23 Blei D, Ng A and Jordan M (2003) Latent Dirichlet Allocation. Journal of Machine Learning Research 3: 993–1022. Borgman CL and Brand A (2024) The Future of Data in Research Publishing: From Nice to Have to Need to Have? Harvard Data Science Review (Special Issue 4). Epub ahead of print 2 April

2003

[3] [3]

Borgohain DJ, Bhardwaj RK and Verma MK (2022) Mapping the literature on the application of artificial intelligence in libraries (AAIL): a scientometric analysis

DOI: 10.1162/99608f92.b73aae77. Borgohain DJ, Bhardwaj RK and Verma MK (2022) Mapping the literature on the application of artificial intelligence in libraries (AAIL): a scientometric analysis. Library Hi Tech 42(1): 149–179. Breiman L (2001) Statistical Modeling: The Two Cultures. Statist. Sci. (16(3)). Epub ahead of print

work page doi:10.1162/99608f92.b73aae77 2022

[4] [4]

Brown TB, Mann B, Ryder N, et al

DOI: 10.1214/ss/1009213726. Brown TB, Mann B, Ryder N, et al. (2020) Language Models are Few-Shot Learners. arXiv:2005.14165. arXiv. Available at: http://arxiv.org/abs/2005.14165 (accessed 2 March 2026). Burrell J (2016) How the machine ‘thinks’: Understanding opacity in machine learning algorithms. Big Data & Society 3(1): 2053951715622512. Bzdok D, Altm...

work page doi:10.1214/ss/1009213726 2020

[5] [5]

Scientometrics 130(9): 5093–5114

Ding L, Lawson C and Shapira P (2025) Rise of Generative Artificial Intelligence in Science. Scientometrics 130(9): 5093–5114. Ding L, Lawson C and Shapira P (2026) Tracking AI’s Scientific Anatomy: A Novel Framework for Analyzing the Use and Diffusion of AI in Science. Epub ahead of print

2025

[6] [6]

(2024) Oil & Water? Diffusion of AI Within and Across Scientific Fields

Duede E, Dolan W, Bauer A, et al. (2024) Oil & Water? Diffusion of AI Within and Across Scientific Fields. arXiv:2405.15828. arXiv. Available at: http://arxiv.org/abs/2405.15828 (accessed 25 February 2026). Dwivedi R and Elluri L (2024) Exploring Generative Artificial Intelligence Research: A Bibliometric Analysis Approach. IEEE Access 12: 119884–119902. ...

arXiv 2024

[7] [7]

Hastie T, Tibshirani R and Friedman J (2017) The Elements of Stadistical Learning: Data Mining, Inference, and Prediction

DOI: 10.1038/s41586-025-09922-y. Hastie T, Tibshirani R and Friedman J (2017) The Elements of Stadistical Learning: Data Mining, Inference, and Prediction. Springer International Publishing. Available at: https://www.sas.upenn.edu/~fdiebold/NoHesitations/BookAdvanced.pdf. He K, Zhang X, Ren S, et al. (2015) Deep Residual Learning for Image Recognition. In...

work page doi:10.1038/s41586-025-09922-y 2017

[8] [8]

Iliadis A and Russo F (2016) Critical data studies: An introduction

DOI: 10.1007/s00146-025-02835-4. Iliadis A and Russo F (2016) Critical data studies: An introduction. Big Data & Society 3(2): 2053951716674238. Ioffe S and Szegedy C (2015) Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift. Epub ahead of print

work page doi:10.1007/s00146-025-02835-4 2016

[9] [9]

(2025) Application of artificial intelligence in academic libraries: a bibliometric analysis and knowledge mapping

Islam MN, Ahmad S, Aqil M, et al. (2025) Application of artificial intelligence in academic libraries: a bibliometric analysis and knowledge mapping. Discover Artificial Intelligence 5(1):

2025

[10] [10]

(2023) An Introduction to Statistical Learning with Applications in R

James G, Witten D, Hastie T, et al. (2023) An Introduction to Statistical Learning with Applications in R. Second Edition. Springer. Available at: https://www.tandfonline.com/doi/full/10.1080/24754269.2021.1980261 (accessed 6 December 2025). Jordan MI and Mitchell TM (2015) Machine learning: Trends, perspectives, and prospects. Science Vol. 349(Issue 6245...

work page doi:10.1080/24754269.2021.1980261 2023

[11] [11]

Expert Systems with Applications 186: 115728

Su M, Peng H and Li S (2021) A visualized bibliometric analysis of mapping research trends of machine learning in engineering (MLE). Expert Systems with Applications 186: 115728. Van Der Vlist F, Helmond A and Ferrari F (2024) Big AI: Cloud infrastructure dependence and the industrialisation of artificial intelligence. Big Data & Society 11(1): 2053951724...

Pith/arXiv arXiv 2021