Objective is to enable participants to:
- Explain how AI functions and its limitations.
- Identify AI-generated hallucinations.
- Verify claims made by AI systems.
Definitions of AI
Artificial Intelligence (AI):
- Defined as the capacity of computers or coded machines to perform tasks typically requiring human intelligence.
- Examples include:
- Network recommendations
- Automatic facial recognition.
Large Language Model (LLM):
- A complex software model trained on extensive datasets for specific tasks.
- Examples include ChatGPT and PLOG.
Generative Artificial Intelligence (Generative AI):
- A subset of AI that employs machine learning to create original content based on learned patterns.
Understanding Machine Learning
Machine Learning:
- Refers to how AI gathers, recognizes, and learns from vast datasets.
- Involves algorithms and statistical methods to make predictions.
Training AI:
- Involves using a training dataset consisting of labeled numbers, photos, or text to teach models how to make decisions.
- The feedback loop process enhances learning through repeated exposure to data.
- Unlike human learning, which can recognize patterns quickly, machines require exposure to vast amounts of data to identify significant characteristics (e.g., recognizing a cat).
Operational Mechanism of AI
Neural Networks:
- Mechanisms that input information, utilize hidden layers of algorithms, and make predictions based on inputs.
Example:
- In identifying a leaf, AI analyzes pixels:
- Understands colors, shapes, edges, and makes predictions with a specified probability (e.g., 96% certainty it's a leaf).
Probability in AI:
- AI operates on probabilistic models:
- Acknowledges the possibility of errors (e.g., 4% chance that it may not be a leaf).
Limitations of AI
Poor Performance Areas:
- Causal reasoning, abstract thinking, common sense, intentional creativity, and recognizing limitations.
- AI may confidently declare incorrect information (known as hallucinations).
- Generated content reflects patterns from existing data rather than genuine creativity.
Verification of AI Claims:
- Importance of verifying any information provided by AI systems before using it for academic, personal decisions, or healthcare.
Credibility of Sources and Verification Practices
Identifying Credible Sources:
- Considerations for verifying claims:
- Research data or citations
- Author and organization verification
- Reputation of the publication
- Consistency across different credible sources.
Recommended Reliable Sources:
- University-based research centers, peer-reviewed journals (e.g., PubMed, Web of Science, Google Scholar), and reputable news organizations (e.g., BBC).
Research Strategies:
- Use primary literature to explore claims (e.g., articles in scientific journals).
- Start with abstracts for a quick overview of research to assess credibility.
Practical Example of Verification
Verify AI Claims:
- Identify the main claim, search credible sources, check authorship, and confirm against multiple sources.
Example Claim:
- Large language models generate text by predicting the next word based on training data patterns.
- Recommended to validate this via articles from reputable sources like BBC or NPR to ensure accuracy.
Exploring Environmental Impact
Energy Consumption:
- AI models consume vast amounts of electricity, equivalent to that used by 100 homes in one year.
- Data centers are significant energy users and create cooling demands due to heat generated by equipment.
- AI video generation can consume significant energy, e.g., a short clip may equate to an hour of microwave use.
Consequences of Data Centers:
- Often located in arid areas, utilizing large quantities of water for cooling.
- Data centers in 2024 projected to consume 183 terawatts, comparable to Pakistan’s total energy consumption.
In Summary
Generative AI operates within a framework of data, algorithms, and statistical models, leading to predictable but sometimes inaccurate outputs.
Always verify claims made by AI, considering environmental impacts and ethical concerns as integral parts of the discussion.