{"id":17,"date":"2025-08-24T15:51:34","date_gmt":"2025-08-24T21:51:34","guid":{"rendered":"https:\/\/mrsspeechonline.com\/ai-slps-part-3-how-ai-works\/"},"modified":"2025-08-24T15:51:37","modified_gmt":"2025-08-24T21:51:37","slug":"ai-slps-part-3-how-ai-works","status":"publish","type":"post","link":"https:\/\/mrsspeechonline.com\/staging\/2315\/ai-slps-part-3-how-ai-works\/","title":{"rendered":"AI & SLPs Series, Part 3: How These Tools Actually Work"},"content":{"rendered":"\n

A Practical Guide to Demystifying Generative Artificial Intelligence<\/h2>\n\n\n\n

Welcome back, fellow SLPs! In our last two posts, we tackled the crucial topic of AI and Clinical Data<\/a>, as well as AI Truths & Myths<\/a>. We learned that responsible AI use starts with safeguarding our clients’ privacy and understanding how AI works in practice.<\/p>\n\n\n\n

We’ve debunked some common myths and explored the complex questions surrounding AI and data security. Now, let’s pull back the curtain and take a look at the mechanics room. In this post, we’ll demystify how AI tools actually work, providing a clear and simple explanation for SLPs.<\/p>\n\n\n\n

\"A<\/figure>
\n

Artificial intelligence can sometimes feel like magic, especially when it comes to generating coherent text for notes, therapy ideas, or reports. You type in a prompt, and seemingly out of nowhere, clinically relevant content appears. But behind this impressive capability lies a fascinating, albeit complex, system built on specific AI principles.<\/p>\n<\/div><\/div>\n\n\n\n

As SLPs, understanding the basics of how AI tools function is essential. It can not only alleviate some of the mystery, but also empower us to use these tools more effectively and with realistic expectations. So, let’s peek under the hood and demystify the mechanics of AI.<\/p>\n\n\n\n

The Foundation: Large Language Models (LLMs)<\/h3>\n\n\n\n

To truly understand how AI works, you must first understand its base. At the heart of most AI tools you encounter that generate text are Large Language Models (LLMs)<\/strong>. Think of these LLMs as incredibly advanced prediction machines for language.<\/p>\n\n\n\n

They’re not thinking, feeling beings; instead, they are sophisticated algorithms. They are often built on a “transformer” architecture, that have learned to predict what word or sequence of words is most likely to come next. This ability to predict is the foundation of how they generate paragraphs, articles, notes, and even initial drafts of reports.<\/p>\n\n\n\n

The Training Process: Learning from Text Patterns<\/h4>\n\n\n\n

The secret to this predictive power lies in the training process<\/strong>. LLMs are trained on truly massive amounts of text data \u2013 billions upon billions of words scraped from books, articles, websites, code, and just about any other digital text you can imagine. For models used in healthcare or clinical research, this could also include vast quantities of de-identified research papers, medical journals, and clinical guidelines.<\/p>\n\n\n\n

During this training, the AI doesn’t process this raw text directly. Instead, it goes through a crucial step called tokenization<\/strong>.<\/p>\n\n\n\n

What is Tokenization?<\/h5>\n\n\n\n

Tokenization<\/strong> is the process of breaking down the text into smaller, manageable units called tokens<\/strong>. These tokens can be whole words (“the,” “client”), parts of words (“-ing,” “re-“), or even individual characters or punctuation marks. The way text is tokenized can vary between different AI models, but the fundamental idea is to convert human language into a numerical format that the AI can process.<\/p>\n\n\n

\n
\"A
AI processes numerical tokens, not human thoughts.<\/figcaption><\/figure>\n<\/div>\n\n\n

Think of tokenization like the raw data stream in the movie The Matrix<\/em>. Instead of seeing words and sentences, the AI sees a constant flow of these individual numerical tokens. This numerical representation is the fundamental language that the AI understands and manipulates.<\/p>\n\n\n\n

Each unique token in the AI’s vast vocabulary is then assigned a numerical ID. Therefore, AI operates at this granular level of tokens and learns statistical relationships between them. It doesn’t typically “hold” or store entire books, articles, or clinical records as copies. Instead, it retains the learned patterns and probabilities of token sequences.<\/p>\n\n\n\n

How is Tokenization Used?<\/h5>\n\n\n\n

Once the text is tokenized and converted into these numerical representations, the AI analyzes this vast ocean of language, identifying patterns<\/strong>. It learns the statistical probability of certain tokens appearing together. This includes the common structures of sentences and paragraphs, and typical ways humans express different ideas in various contexts (e.g., academic writing, informal speech, or clinical documentation).<\/p>\n\n\n\n

For example, the AI learns that after the tokens “Client” followed by “demonstrated,” the token “progress” has a very high probability of appearing next in a clinical note. It learns the statistical relationships between diagnostic terms and common intervention strategies.<\/p>\n\n\n\n

It’s all about recognizing these statistical relationships between these numerical tokens. It’s crucial to remember that the AI doesn’t “understand” these relationships in the way a human clinician does. It doesn’t grasp the clinical meaning, the client’s emotional state, or the unique context of a therapy session. It simply recognizes that, based on the numerical sequences it has processed, certain token sequences are more probable than others.<\/p>\n\n\n\n

Mechanics of AI Text Generation: Prediction in Action<\/h4>\n\n\n\n
\"A<\/figure>
\n

So, how does this learned knowledge translate into generated text that can assist an SLP? Let’s discuss the AI principles at work. First, you give an AI a prompt<\/strong> \u2013 a request, a question, or even just a few keywords (e.g., “Draft a SOAP note for a 5-year-old with articulation goals for \/r\/”). Then, the AI first tokenizes your input.<\/p>\n<\/div><\/div>\n\n\n\n

After that, it processes its learned patterns and the numerical IDs of these tokens. It then predicts the most likely next token (represented by its numerical ID). This prediction is heavily influenced by the parameters<\/strong> of the LLM. Think of these parameters, often billions or even trillions, as the internal “knobs and dials”, like the Oz Wizard. The AI adjusts these during its training to learn the strength of the statistical relationships between tokens. The more parameters an AI has, generally the more nuanced and sophisticated its predictions can be.<\/p>\n\n\n\n

The Output of the AI Machine<\/h5>\n\n\n\n

The AI takes that newly predicted token. It compares the original prompt and the token it just produced. Then it predicts the next<\/em> most likely token, again relying on its vast network of parameters. This process continues, token by token, until the AI determines it has reached a stopping point. Finally, these numerical tokens are converted back into human-readable text.<\/p>\n\n\n\n

Think of it like a super-powered version of the autocomplete feature you might use when typing. However, it operates at the level of fundamental text units, and is guided by an incredibly complex web of learned relationships stored within its parameters. The “creativity” or “clinical relevance” we sometimes see in AI-generated text isn’t true clinical reasoning in the human sense. Rather, it is the result of the sheer vastness of the data it has been trained on. It has learned the intricate statistical relationships between these tokens. It is all governed by its numerous parameters.<\/p>\n\n\n\n

What AI Isn’t: Understanding the Clinical Limitations<\/h2>\n\n\n\n

Finally, it’s vital to reiterate what AI isn’t<\/strong>, especially from an SLP perspective. Even understanding the mechanics of AI, its human-like responses can often muddy the waters.<\/p>\n\n\n

\n
\"A
Knowing the limits of AI is essential.<\/figcaption><\/figure>\n<\/div>\n\n\n