"A Nova Eletricidade: Aplicações, Riscos e Tendências da IA Moderna - "The New Electricity": Applications, Risks, and Trends in Current AI

Ana L. C. Bazzan, Anderson R. Tavares, André G. Pereira, Cláudio R. Jung, Jacob Scharcanski, Joel Luis Carbonera, Luís C. Lamb, Mariana Recamonde Mendoza, Thiago L. T. da Silveira, Viviane P. MoreiraarXiv 2023 · DOI · arXiv

Abstract

Artificial Intelligence has transitioned from an academic pursuit to a foundational societal force, often described as the "new electricity." This paper provides a comprehensive synthesis of the modern AI landscape, exploring the evolution from traditional symbolic logic to modern deep learning. By analyzing applications in healthcare and finance alongside the risks of bias and "black-box" decision-making, we argue for the necessity of neuro-symbolic AI—a "third wave" that combines the perceptual power of neural networks with the logical transparency of symbolic systems.

Introduction: AI as Infrastructure

In the 19th century, electricity redefined the "plumbing" of industry and daily life. Today, Artificial Intelligence is undergoing a similar transformation into a "general-purpose technology." Its value no longer resides solely in standalone applications but in its ability to reshape urban mobility, global finance, and medical diagnostics. However, as AI moves from the lab to the core of our infrastructure, a significant gap has emerged between the rapid deployment of these models and our ability to manage their inherent risks.

The Dual Paradigm: Thinking Fast and Slow

To understand the current state of AI, the authors employ Daniel Kahneman’s framework of "System 1" and "System 2" thinking to categorize AI methodologies:

  • System 1 AI (Learning-based): Represented by modern Deep Learning, these models are fast, instinctive, and capable of incredible pattern recognition. However, they are often "black boxes"—prone to errors that are difficult to trace and lacking a rigorous understanding of the world.
  • System 2 AI (Symbolic): This traditional approach focuses on knowledge representation and formal logic (e.g., Prolog). While deliberate, logical, and transparent, symbolic AI is often too brittle to handle the messy complexity of real-world data.

Modern successes, such as Large Language Models (LLMs), operate primarily as "System 1" entities. They function as "stochastic parrots," generating plausible-sounding text without a true semantic grasp of the underlying concepts, leading to the well-documented phenomenon of "hallucination."

Methodological Evolution and the Scale of Data

The field has evolved through several distinct methodologies, moving from the logic-based A* search of Symbolic AI to the Connectionist revolution. This shift saw the rise of Convolutional Neural Networks (CNNs) for vision and Transformers for Natural Language Processing (NLP). Furthermore, Reinforcement Learning (RL) has moved from game-playing agents (DQN) to Reinforcement Learning with Human Feedback (RLHF), which is essential for aligning modern chatbots with human expectations.

This evolution is fueled by massive datasets. For instance, Vision-Language Models are now trained on sets like LAION-5B, containing 5.85 billion image-text pairs. However, the use of uncurated web-scale corpora introduces significant risks, as these models frequently replicate and amplify societal biases (racism, sexism) found in the source data.

Sectoral Impact and Real-World Performance

The paper highlights significant milestones across various industries:

  • Finance: AI integration is projected to save U.S. banks over $1 trillion by 2030 through automation and enhanced risk assessment.
  • Healthcare: While AI has reached superhuman maturity in fields like radiology and dermatology, a "disparity" remains between research and clinical utility. As of late 2022, only 521 AI-based medical devices had received FDA approval, despite thousands of published studies.
  • Computer Vision: While performance on benchmark datasets is high, models often suffer from "domain shift." A model trained on clear, daytime driving data may fail significantly when faced with nighttime or rainy conditions, highlighting a lack of robust generalization.

The Challenges of Modern AI

The transition to "Broad AI" is hindered by several critical factors:

  • Explainability and Bias: We lack the tools to explain high-stakes decisions, such as why a model denied a loan or flagged a medical image. Furthermore, facial recognition error rates show a staggering disparity: 0.8% for light-skinned men versus 34.7% for dark-skinned women.
  • The Academic-Industry Divide: The cost of entry has become a barrier to innovation. Training a model with 1.5 billion parameters can cost upwards of $1.6 billion, effectively pricing out academic institutions and concentrating power within a few tech giants.
  • Environmental Impact: The massive energy requirements for training and maintaining LLMs are flagged as a growing concern for the sustainability of the field.

Conclusion: The Neuro-symbolic Future

The authors conclude that the "Holy Grail" of AI research is the integration of System 1 and System 2. This "Neuro-symbolic AI" aims to embed logical constraints directly into neural architectures. By combining the perceptual strengths of deep learning with the formal reasoning of symbolic logic, the next generation of AI can become more robust, explainable, and fundamentally aligned with human values.

Summary updated

BibTeX
@misc{bazzan2023nova,
  author = {Ana L. C. Bazzan and Anderson R. Tavares and André G. Pereira and Cláudio R. Jung and Jacob Scharcanski and Joel Luis Carbonera and Luís C. Lamb and Mariana Recamonde Mendoza and Thiago L. T. da Silveira and Viviane P. Moreira},
  title = {{"A Nova Eletricidade: Aplicações, Riscos e Tendências da IA Moderna - "The New Electricity": Applications, Risks, and Trends in Current AI}},
  year = {2023},
  eprint = {2310.18324},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2310.18324},
}
Figure