Skip to content

AI and Machine Learning Foundations

How AI Uses Machine Learning

Artificial Intelligence (AI) has been enhancing user experiences for years, powering features like streaming recommendations and GPS routing. While AI tools may seem inherently smart, they are not self-taught; they are powered by Machine Learning (ML).

Machine Learning (ML) is a subset of AI focused on developing computer programs that can analyze data to make decisions or predictions. ML enables AI tools to process information quickly and efficiently without being explicitly programmed for every scenario.

Training Sets and Bias

AI developers build ML programs using training sets—large collections of data used to teach the AI what to expect and how to respond. For instance, to train a model to identify ripe apples, it is fed thousands of images of ripe and unripe apples until it learns to identify the relevant features.

However, the quality and relevance of training data are critical. Bias within training data can inadvertently cause an AI tool to produce inaccurate or skewed outputs. If the apple-sorting AI was trained only on red apples, it would perform poorly on green apples.

Approaches to Machine Learning

There are three common ML approaches used to develop AI tools:

  1. Supervised Learning: Trains tools using a massive dataset that has been labeled by humans. This is used when there is a specific, known output in mind (e.g., recognizing "cats" from labeled images of cats).
  2. Unsupervised Learning: Trains tools on unlabeled datasets to identify patterns and structures. The ML model clusters similar data on its own (e.g., grouping photos of animals with whiskers and pointy ears without knowing the label "cat").
  3. Reinforcement Learning: Trains tools through a trial-and-error process guided by feedback. The model receives positive feedback for successful outputs, helping to continuously refine its performance over time.

Note: The "learning" in these approaches primarily occurs during the tool's development and training phase, not actively in real-time as end-users interact with it.

Rule-Based AI

Unlike ML, rule-based AI operates using a set of hard-coded rules created by human developers. It does not learn from new data but follows its specific instructions precisely (e.g., "If message contains 'tracking', send tracking link"). While less flexible than ML, it is highly reliable for predictable tasks.

Foundations of Generative AI

Generative AI is a type of AI capable of creating new content such as text, images, or code. A defining characteristic of generative AI tools is that they can be operated using natural language—the way people talk or write when communicating with each other.

How Generative AI Works

  1. Input: You provide information or data to the AI tool. A prompt is text input that provides instructions to the AI model on how to generate output. Prompts can be edited and refined as many times as needed.
  2. Processing: The AI processes the data based on its underlying ML models.
  3. Output: The tool generates a response in the form of text, images, audio, or video.

Conversational AI

A conversational AI tool (like Google's Gemini) is a type of generative AI that processes text requests and generates text responses. These tools can boost productivity by drafting emails, brainstorming ideas, answering questions, and providing creative inspiration.

Capabilities and Limitations of AI

AI tools can significantly enhance your work by generating content, analyzing information quickly, and simplifying day-to-day tasks. However, understanding their limitations is just as important.

Limitations and Hallucinations

  • Requires Human Training: AI cannot learn entirely independently; it relies on people to continually update and refine its training data.
  • Bias: Shortcomings in training data can reflect or amplify societal biases, leading to unfair outcomes.
  • Knowledge cutoff: The concept that an AI model is trained at a specific point in time, so it doesn't have any knowledge of events or information after that date.
  • Hallucinations: AI outputs can sometimes contain inaccuracies or fabricate facts entirely. These range from minor nonsensical errors to significant distortions that could impact business decisions (e.g., misinterpreting seasonal sales trends).

The Importance of Human Oversight

Given these limitations, human oversight over AI-generated outputs is crucial. Humans are needed to handle nuanced or sensitive issues and ensure that AI outputs are accurate, ethical, and aligned with core values. Effective management of AI requires teamwork and an inclusive approach to shape a future where AI benefits everyone.