What is the general content, and what features are discussed?
Is there anything in the text that you did not understand, but thought it may be relevant for law?
Is there anything in the text that you did not understand, but thought it would probably be of little legal interest, so can be left aside?
Does a specific issue in the chapters pose legal issues – here you may have to be creative – for instance, a description of how an AI system can be built to work accurately can raise legal issues if it malperforms: was it built according to the best of our knowledge, or was it below acceptable standards?
How realistic do you think the depictions in public are? Can current systems perform tasks depicted, will they soon be able to, or are the depictions and claims of their abilities unrealistic?
Introduction
Chapters 3 and 5 of Boadu's book Machine_Learning_in_Aluminium_Reduction focus on neural network theory and the practical training of neural networks for industrial problem-solving.96 Chapter 3 explains neural networks by drawing inspiration from biological neurons and describing how artificial neurons, synaptic weights, transfer functions, biases, and network architectures operate.97 It discusses feedforward and recurrent networks, as well as supervised and unsupervised learning, and learning paradigms such as the Perceptron, Adaline (Widrow-Hoff), and Backpropagation networks.98 Chapter 5 applies these concepts to estimating alumina concentration in aluminium reduction cells through neural network training, experimentation, error minimisation, function approximation, and generalisation.99 This paper examines the legal significance of artificial neural networks and whether public perception of AI reflects technological reality.
Challenging Concept and Its Legal Relevance
One concept found difficult to understand was generalisation.100 The ability of a neural network to learn from limited data and accurately respond to unseen inputs is technically complex, yet highly relevant to law. Modern AI systems frequently make predictions based on learned patterns rather than explicit rules. This raises important legal questions concerning accountability, transparency, and reliability. Where an AI system generalises incorrectly and causes harm, courts may need to determine whether the training data, system architecture, or deployment process satisfied the applicable standard of care.
Concepts of Limited Legal Relevance
Some technical aspects appear less significant from a legal perspective. Mathematical derivations of transfer functions, learning rates, gradient descent calculations, and error minimisation formulas are primarily engineering concerns.101 While important for system design, these calculations are unlikely to be central to legal disputes unless expert evidence is required to establish causation, system failure, or defective design.
Legal Implications of Neural Network Design
The chapters raise significant legal issues relating to the design and deployment of AI systems. A recurring theme is that neural network performance depends heavily on architecture, training quality, and error management.102 It is demonstrated that some network designs failed to learn accurately, while others achieved improved results only after extensive testing and modification.103 This has implications for negligence, product liability, and regulatory compliance. If an AI developer deploys a system despite being aware of known limitations, inadequate testing, poor-quality training data, or excessive error rates, liability may arise if foreseeable harm results. The discussion of local minima, overfitting, underfitting, and generalisation errors demonstrates that AI systems are inherently imperfect, making questions of reasonable care and industry standards particularly important.104
Public Perception and the Reality of AI
Public depictions of AI are often exaggerated. The chapters demonstrate that neural networks require extensive training, careful parameter selection, and continual evaluation before they can perform reliably.105 Current AI systems, including Generative AI (GAI) and Large Language Models (LLMs), are powerful prediction and pattern-recognition tools, but they remain dependent on training data, computational architectures, and statistical learning processes. Artificial Superintelligence (ASI) remains a theoretical concept. The chapters, therefore, suggest that AI is highly useful but far from infallible. Claims that current systems possess human-level understanding, unlimited reasoning, or autonomous superintelligence should consequently be treated with caution.106
Conclusion
Chapters 3 and 5 provide a practical understanding of neural networks and machine learning. While the technology offers significant benefits, AI systems remain susceptible to error, uncertainty, and design limitations. As AI becomes increasingly integrated into decision-making processes, issues of accountability, liability, transparency, and regulatory oversight will become increasingly important.