"A Nova Eletricidade": Aplicações, Riscos e Tendências da IA Moderna

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. MoreiraEscola de Computação PPGC/UFRGS 50 Anos 2023 · DOI · DBLP

Abstract

Artificial Intelligence (AI) has transitioned from a niche academic pursuit to a general-purpose technology, often compared to electricity in its potential to restructure modern society. This paper, synthesized from research by the Federal University of Rio Grande do Sul (UFRGS), examines the evolution of AI from symbolic logic to deep learning. It identifies a critical gap between intuitive pattern matching and formal reasoning, advocating for a "Third Wave" of Neuro-symbolic AI to address the inherent risks of opacity, bias, and unreliability in current systems.

Introduction: The New Electricity

In recent years, Andrew Ng’s analogy that "AI is the new electricity" has moved from a bold prediction to a tangible reality. Much like the electrical grid transformed every industry in the 20th century, modern AI—specifically deep learning—is currently rewriting the rules of healthcare, finance, and urban infrastructure. However, this rapid expansion brings significant challenges. As we move from classical, rule-based systems to massive neural architectures, the industry faces a "trust gap" that necessitates a fundamental shift in how we design and deploy intelligent systems.

The Cognitive Gap: System 1 vs. System 2

To understand the limitations of modern AI, the researchers utilize Daniel Kahneman’s framework from Thinking, Fast and Slow.

  • System 1 (Fast Thinking): This represents the current state of Deep Learning. It is instinctive, fast, and excels at pattern recognition (e.g., identifying a face or generating a sentence). However, it is often opaque and lacks a logical foundation.
  • System 2 (Slow Thinking): This represents deliberate, logical reasoning. While early AI (Symbolic AI) attempted to replicate this through formal logic, it was too brittle for the complexities of the real world.

The core problem addressed by the UFRGS review is that while current AI is a "System 1" powerhouse, it lacks the "System 2" capabilities required for high-stakes decision-making. This deficiency leads to "hallucinations"—where models generate plausible but false information—and a lack of explainability in critical sectors.

The Evolution of AI: Three Waves

The paper categorizes the history and future of AI into three distinct methodological waves:

  1. The First Wave (Symbolic AI): Focused on knowledge representation and formal logic (e.g., Prolog). These systems are transparent and easy to audit but fail when faced with noisy, real-world data.
  2. The Second Wave (Connectionist AI/Deep Learning): Inspired by biological neural networks, this wave includes Convolutional Neural Networks (CNNs) for vision and Transformers for language. These models handle "messy" data with ease but operate as "black boxes."
  3. The Third Wave (Neuro-symbolic AI): The authors advocate for an integration of the previous two. Neuro-symbolic AI aims to combine the pattern-recognition strengths of neural networks with the logical rigor of symbolic systems, creating models that are both powerful and "explicable by design."

Data and Foundation Models

The current AI explosion is fueled by massive datasets. The researchers highlight the role of LAION-5B (billions of image-text pairs) and Common Crawl as the backbones of Foundation Models. In specialized fields, the shift toward domain-specific data—such as medical imaging (CT scans) and industrial sensor logs—is enabling AI to move beyond general conversation into high-precision technical applications.

Transformative Results Across Sectors

The impact of this "new electricity" is already measurable across several domains:

  • Healthcare: AI models are now matching specialists in dermatology and radiology. Notably, some systems can predict lung cancer risk up to three years in advance by identifying subtle patterns in imaging that escape the human eye.
  • Finance: The banking sector is projected to see over $1 trillion in savings by 2030. AI is being used to automate complex risk assessments and detect fraudulent patterns in real-time.
  • Urban Mobility: Through Reinforcement Learning (RL), cities are optimizing traffic flow and light timing, leading to significant reductions in congestion and carbon emissions.
  • Industry 4.0: The use of "Digital Twins" and computer vision allows for predictive maintenance, where AI identifies potential equipment failures before they occur, saving millions in downtime.

Risks, Ethics, and Limitations

Despite these successes, the transition to an AI-driven society is fraught with risks. The researchers identify three primary concerns:

  1. Bias and Fairness: Models often mirror the prejudices found in their training data, leading to discriminatory outcomes in facial recognition and hiring algorithms.
  2. Environmental Impact: The computational cost of training Large Language Models (LLMs) like GPT-4 is immense, raising concerns about the carbon footprint of the AI industry.
  3. The Explainability Crisis: As models grow in complexity, auditing their decision-making process becomes nearly impossible. This is particularly dangerous in legal or medical contexts where a "correct" answer is insufficient without a logical justification.

Conclusion: Toward a Robust Future

The UFRGS researchers conclude that while the "Second Wave" of AI has provided us with incredible tools, we cannot rely on scale alone to solve the problems of logic and ethics. The path forward lies in Neuro-symbolic AI and Reinforcement Learning from Human Feedback (RLHF). By embedding human values and logical constraints directly into neural architectures, we can ensure that the "new electricity" powers a future that is not only efficient but also transparent and just.

Summary updated

BibTeX
@inproceedings{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 and Anderson Rocha Tavares and André M. Pereira and Jacob Scharcanski},
  title = {{"A Nova Eletricidade": Aplicações, Riscos e Tendências da IA Moderna}},
  booktitle = {Escola de Computação PPGC/UFRGS 50 Anos},
  year = {2023},
  doi = {10.5753/sbc.13058.5.8},
}
Figure