Skip to main content
Log in

Framing the effects of machine learning on science

  • Published:
AI & SOCIETY Aims and scope Submit manuscript


Studies investigating the relationship between artificial intelligence (AI) and science tend to adopt a partial view. There is no broad and holistic view that synthesizes the channels through which this interaction occurs. Our goal is to systematically map the influence of the latest AI techniques (machine learning, ML and its sub-category, deep learning, DL) on science. We draw on the work of Nathan Rosenberg to develop a taxonomy of the effects of technology on science. The proposed framework comprises four categories of technology effects on science: intellectual, economic, experimental and instrumental. The application of the framework in the relationship between ML/DL and science allowed the identification of multiple triggers activated by the new techniques in the scientific field. Visualizing these different channels of influence allows us to identify two pressing, emerging issues. The first is the concentration of experimental effects in a few companies, which indicates a reinforcement effect between more data on the phenomenon (experimental effects) and more capacity to commercialize the technique (economic effects). The second is the diffusion of new techniques lacking in explanation (intellectual effect) throughout the fabric of science (instrumental effects). The value of this article is twofold. First, it provides a simple framework to assess the relations between technology and science. Second, it provides this broad and holistic view of the influence of new AI techniques on science. More specifically, the article details the channels through which this relationship occurs, the nature of these channels and the loci in which the potential effects on science unfolds.

This is a preview of subscription content, log in via an institution to check access.

Access this article

Price excludes VAT (USA)
Tax calculation will be finalised during checkout.

Instant access to the full article PDF.

Fig. 1

Source: author’s own

Fig. 2
Fig. 3

Source: Scopus

Similar content being viewed by others


  1. Even though we agree with Flexner (1939) that curiosity is as much an essential driver of scientific inquiry as to the prospect of use.

  2. Price probably wanted to allude to a technological innovation that escapes the dominant technological paradigm when he mentions unexpected technological innovation. He could not mobilize this concept because Giovanni Dosi (1984) would still spread it in the years to come.

  3. “Base of observations” and “instrumentation” refer to the initial trigger that emerges in the technological sphere and that will influence science: new data from the body of empirical knowledge or new technological artifacts for instrumentation; on the other hand, “influencing the agenda” refers to the result, already in the scientific sphere, of changes that have taken place in the technological sphere. There is some confusion between cause and effect.

  4. Also referred as an 'economy of research tools.' (Cockburn et al., 2018).

  5. "Opening up the set of problems that can be feasibly addressed, and radically altering scientific and technical communities' conceptual approaches and framing of problems." (Cockburn et al., 2018).

  6. According to Brooks (1994), “the more radical the invention, the more likely it is to stimulate wholly new areas of basic research or to rejuvenate older areas of research that were losing the interest of the most innovative scientists.”.

  7. “The re-construction problem is serious, because even with complete information about the operations of a system, an ex-post analysis of a specific decision may not be able to establish a linear causal connection which is easily comprehensible for human minds.” (Wischmeyer, 2020, p. 81).

  8. (TITLE-ABS-KEY("xAI" OR "XAI" OR "explainable artificial intelligence" OR "explainable AI" OR "explainable machine learning" OR "explainable deep learning" OR "explainable algorithms" OR "interpretable artificial intelligence" OR "interpretable AI" OR "interpretable machine learning" OR "interpretable deep learning" OR "interpretable algorithms" OR "opaque artificial intelligence" OR "opaque AI" OR "opaque machine learning" OR "opaque deep learning" OR "opaque algorithms" OR "responsible artificial intelligence" OR "responsible AI" OR "responsible machine learning" OR "responsible deep learning" OR "responsible algorithms" OR "transparent artificial intelligence" OR "transparent AI" OR "transparent machine learning" OR "transparent deep learning" OR "transparent algorithms") AND ( LIMIT-TO ( DOCTYPE,"cp") OR LIMIT-TO ( DOCTYPE,"ar") OR LIMIT-TO ( DOCTYPE,"re"))).

  9. “People can’t explain how they work, for most of the things they do. When you hire somebody, the decision is based on all sorts of things you can quantify, and then all sorts of gut feelings. People have no idea how they do that. If you ask them to explain their decision, you are forcing them to make up a story. Neural nets have a similar problem […] You should regulate them based on how they perform” Geoffrey Hinton interview in (Simonite, 2018).

  10. We understand AI science as Gazis (1979, p. 252) understands computer science/software science: “the search for knowledge, or the development of a methodology, that goes beyond satisfying the needs of a single application, but forms the basis for new applications."


  12. According to Russell and Norvig (2020) “Shared benchmark problem sets became the norm for demonstrating progress, including the UC Irvine repository for machine learning data sets, the International Planning Competition for planning algorithms, the LibriSpeech corpus for speech recognition, the MNIST data set for handwritten digit recognition, ImageNet and COCO for image object recognition, SQUAD for natural language question answering, the WMT competition for machine translation, and the International SAT Competitions for Boolean satisfiability solvers.”

  13. “Deep learning relies heavily on powerful hardware. Whereas a standard computer CPU can do 109 or 1010 operations per second. a deep learning algorithm running on specialized hardware (e.g., GPU, TPU, or FPGA) might consume between 1014 and 1017 operations per second, mostly in the form of a highly parallelized matrix and vector operations.” (Russell & Norvig, 2020).

  14. “Von Neumann architecture is seen by its critics as a major obstacle to good programming in general. In one area, however, the shortcomings of the conventional approach have a particular importance. This is the area of artificial intelligence.” (Peláez, 1990, p. 68).

  15. The development of NetWare (hardware for neural networks) was high on the agenda in the 1980s (van Raan & Tijssen, 1993). However, as already mentioned, all things related to neural networks were marginalized and would only return to the spotlight after AlexNet’s breakthrough in 2012.

  16. Intel researchers have released Loihi, the company's fifth generation of chips inspired by neuromorphic technologies (Davies et al., 2018).

  17. See Google’s Cloud TPU:


Download references


Preliminary versions of this article were presented at the 2021 SPRU PhD Forum and the XVIII International Schumpeter Society Conference. The authors are grateful for the excellent comments, criticisms and suggestions that were made on these occasions, especially from Professor Paul Nightingale (SPRU - University of Sussex), Dr. Simone Vannuccini (SPRU - University of Sussex) and Stephano Bianchini (BETA - Université de Strasbourg, France). We are also grateful for the careful reading of the document by Professor Altair Oliveira (Federal Institute of São Paulo) and Professor Koen Frenken (Copernicus Institute of Sustainable Development - Utrecht University). Any possible inaccuracies that may have persisted in the article are the sole responsibility of the authors.


The authors have no relevant non-financial interests to disclose. This article was funded by Coordenação de Aperfeiçoamento de Pessoal de Nível Superior, 88882.329792/2019-01, Victo Silva.

Author information

Authors and Affiliations


Corresponding author

Correspondence to Victo J. Silva.

Additional information

Publisher's Note

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Appendix 1

Appendix 1

The corpus consists of 32 documents distributed over 50 years of the author's scientific production (1963–2013). The criterion for defining the corpus was to include the principal works by Rosenberg (Albuquerque 2017; Mowery et al. 1994). In addition to these works, we included documents (preferring, when possible, peer-reviewed articles) dealing with the relationship between SandT, seeking to balance works from different phases of the author's scientific production. Our fundamental unit of analysis is the historical examples regarding the influence of technology on science, regardless of the nature of this influence at first. We consider as a codifiable unit of analysis specific historical examples that mention any effect of technology on science. An example is "the agricultural experiment stations conducted research aimed at improving the productivity of all types of agriculture enterprise and attempting to solve problems of plant and animal disease” (Rosenberg and Steinmueller 2013). Generic mentions of these same effects are not considered units of analysis; therefore, they were not codified. An example of generic mention is “The problems encountered by sophisticated industrial technologies and the anomalous observations or unexpected difficulties they produced as powerful stimuli to scientific research in the academic community” (Mowery and Rosenberg 1998). A unit of analysis can receive more than one code; however, we sought to minimize this multiple encoding. We searched and mapped units of analysis in the corpus until reaching a point of apparent saturation. Table

Table 4 Code application per categories (4) and sub-categories (8) of effects.

4 sums up the encoding results.

Starting from the initial analysis (Rosenberg 1982) and following terminological guidelines (Hodgson 2019), whenever possible we adopted the terms and definitions already proposed by Rosenberg. We propose new categories and sub-categories when we find units (historical cases) that do not fit into any sufficiently defined ex-ante category; one that stands out is the intellectual effect, absent from the corpus in the meaning we adopt here.

Rights and permissions

Reprints and permissions

About this article

Check for updates. Verify currency and authenticity via CrossMark

Cite this article

Silva, V.J., Bonacelli, M.B.M. & Pacheco, C.A. Framing the effects of machine learning on science. AI & Soc 39, 749–765 (2024).

Download citation

  • Received:

  • Accepted:

  • Published:

  • Issue Date:

  • DOI: