What is AI, what is it not, how we use it in physics and how it impacts… you
Claire David 1 · 2025
Read What is AI, what is it not, how we use it in physics and how it impacts… you by Claire David 1 (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2504-01827.txt.
Introduction
Claire David asks what AI is and is not, and how physics uses it, in a 2025 essay. Machine learning has shaped particle physics for over three decades, from event selection to simulation.
David separates machine learning from classic programming, where rules are written by hand rather than learned from data. The essay surveys uses in high energy physics, highlighting simulation-based inference, uncertainty-aware models, and fast learning for anomaly spotting. It then turns to harms AI systems can cause in wider society, urging responsible engagement. Physicists, as users and citizens, must adapt their practice as tools evolve. Staying critical, David argues, keeps the field at the front of innovation rather than swept along by it.
Catalog record
| CREATOR | Claire David 1 |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 19 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2504-01827
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
What is AI, what is it not, how we use it in physics and how it impacts… you Claire David 1 Abstract Artificial Intelligence (AI) and Machine Learning (ML) have been prevalent in particle physics for over three decades, shaping many aspects of High Energy Physics (HEP) analyses. As AI’s influence grows, it is essential for physicists — as both researchers and informed citizens — to critically examine its foundations, misconceptions, and impact. This paper explores AI definitions, examines how ML differs from traditional programming, and provides a brief review of AI/ML applications in HEP, highlighting promising trends such as Simulation-Based Inference, uncertainty-aware machine learning, and Fast ML for anomaly detection. Beyond physics, it also addresses the broader societal harms of AI systems, underscoring the need for responsible engagement. Finally, it stresses the importance of adapting research practices to an evolving AI landscape, ensuring that physicists not only benefit from the latest tools but also remain at the forefront of innovation. Conference Proceedings for BCVSPIN 2024: Particle Physics and Cosmology in the Himalayas Kathmandu, Nepal, December 9-13, 2024 1 Introduction Machine learning (ML) techniques have been used in particle physics for the past three decades. However, with the technological leap that occurred during the past decade in generative Artificial Intelligence (AI), there is now a growing need to ask ourselves critical relevant questions and take a stance on present and future AI deployments, both within and outside the field of particle physics. After exploring the definitions of the key terms of AI and explaining the drastic conceptual changes brought by AI in Section 2 (“What is AI?”), Section 3 will consider the common misconceptions about the field (“AI: What it is not”). A succinct review of the usage of AI and ML in particle physics is presented in the section 4. Section 5 steps outside the realm of physics to examine the individual and societal impacts of AI tools in our daily lives. The document concludes with some reflections on how we might navigate the ongoing AI revolution. 2 What is AI? What does the somewhat provocative term ‘artificial intelligence’ mean? What is it and what is its scope? Is the terminology borrowed from human behaviour truly appropriate? This section challenges the reader with deceptively simple questions, presents the paradigm shift AI has introduced through the data-driven approach of machine learning, and demystifies the core mathematics behind most AI algorithms. 2.1 A few basic questions Defining the key terms of artificial intelligence is challenging because it presupposes that there is agreement about the meaning of intelligence. What is intelligence? Intelligence is a complex and widely debated concept, with a pluralistic nature involving abstraction, logic, emotional processing, memory, and more. Its definition varies across disciplines, from psychology to neuroscience and philosophy, depending on the perspective taken. In [ 1 ] , Shane Legg and Marcus Hutter attempt to provide a universal definition of intelligence. However, the very first sentence of their article reads: “ A fundamental problem in artificial intelligence is that nobody really knows what intelligence is. ” They eventually settle with an informal working definition: “Intelligence measures an agent’s ability to achieve goals in a wide range of environments.” One common critique of this definition is that “a system may appear to be intelligent without really understanding anything” [ 2 ] . This leads us to ask: “What is understanding?” And, as it appears in machine learning, we must also ask: “What is learning, and how do the two differ?” Understanding implies deeper insights and a stronger grasp of knowledge, whereas learning is about the acquisition of knowledge. The key takeaway here is to question whether it is legitimate to borrow terms that are rooted in human behaviour when these terms themselves lack clear and widely accepted definitions. 2.2 The paradigm shift from explicit programming What machine learning brings to the computer world is a drastic paradigm shift. Before the advent of ML algorithms, programming referred to sets of instructions prepared by a human programmer. These could be dynamic and complex with different outcomes depending on inputs fulfilling or not conditional statements ( if , while , for etc). Yet these outcomes are limited by design. Explicit programming is ruled-based. By contrast, supervised machine learning uses the ‘answers’ (that is, the targets) and other data, to derive implicit rules from the data in order to make predictions (a proxy for answer). In the context of unsupervised learning, implicit rules are derived from data without the aid of targets (see Figure 1 ). Such a paradigm shift brings the focus to the data — a fortunate turn for physics, as data ultimately holds the answers. Figure 1: The fundamental change from explicit programming (top in blue) versus machine learning (upper green, supervised machine learning, lower green, unsupervised machine learning). Illustration from the author. 2.3 The core mathematics of AI/ML As complex and sophisticated as modern AI can be, the mathematics at the core of the majority of those algorithms are simple: it’s all about fitting a function to the data. Deep down, it is a minimization problem. Mathematically, given a dataset of features \bm{x} and targets \bm{y} , D={(x_{i},\>y_{i})},i=1,\cdots n\;, (1) fitting the data — that is training — is choosing the hypothesised function h * that minimises the empirical risk \hat{R} — that is averaged loss {L} , or cost — defined as: \hat{R}(h)=\frac{1}{n}\sum_{i}^{n}L(h(x_{i}),\,y_{i})\;. (2) The hypothesised function needs to be sufficiently flexible (which is usually the case with a deep neural network). It is important to note that the optimal h * only depends on the form of the loss function and the distribution of the data [ 3 ] . The most advanced algorithms may involve functions of immense complexity, yet at their core, the underlying mathematical operations are governed by the general Equation 2 . 3 AI: What it is not The field of AI is interdisciplinary, built on mathematics, statistics, and computer science, while also drawing concepts from physics. However, it is not “glorified statistics” as some may argue on the internet. The goals differ: statistics are geared more toward an “autopsy” of the data, studying correlations and making inferences, whereas AI is largely about predictions. Most AI deployed as tools for physics do not attempt physics modelling. For a theory in physics to be complete, it must be both descriptive and predictive. Machine learning accomplishes the latter but does not achieve the former at present. AI is not “the solution” for our research but rather a tool for a given, precise task. However, with the advent of Large Language Models (LLMs), researchers at all levels now use them for tasks like writing papers, coding, or even as self-teaching tools. But caution is warranted because these agents are still prone to hallucinations, a colloquialism for nonsensical outputs. 4 AI in Physics 4.1 Brief overview of AI in HEP AI/ML is now ubiquitous in High Energy Physics (HEP) and the applications are wide-ranging and far too numerous to review here. Instead we highlight a few key trends. Figure 3 provides an overview focused solely on the ATLAS Experiment and only covers the past few years: ML techniques are deployed throughout each step of an analysis workflow.