Artificial Intelligence (AI): How AI Is Trained

How AI Is Trained

Generative AI tools, such as LLMs, create responses by predicting text based on the information or data from their “training.” Training in this context means inputting specific information or data into a generative AI program from which answers will be mined and presented to the user. These data often include publicly available content from websites and user interactions, which can impact the accuracy and relevance of the response. The data sets that commercial LLMs use are not specifically curated for the Communication Sciences and Disorders (CSD) field. This is particularly relevant for new codes, legislation, or evolving standards in health care delivery.

Clinicians should be aware of two primary limitations.

Accuracy and Relevance

Fabricated information: Generative AI models may generate “hallucinations”—plausible sounding statements, images, or other content that contain false information or disinformation.

Inaccurate Information: The information used to generate a response may include misinformation, disinformation, or factually inaccurate content.

Domain gaps: Rare disorders or specialized clinical protocols may be underrepresented in training data, leading to superficial or generic outputs that do not consider the individual client, patient or student.

Outdated information: Early LLMs were trained on data that might be outdated. LLMs do not “forget” old information and may generate misleading content risking the exclusion of updated guidelines, policies, and innovations.

Representation in Training Data

Under-representation of marginalized populations: Training sets may not reflect the full range of dialects, languages, needs, and cultural contexts encountered in practice.

Potential for biased language: Without careful review, outputs may perpetuate stereotypes or insensitive phrasing.

Strategies for enhancing the reliability of generative AI responses include using tools that

Allow the user to provide credible and relevant sources for the LLM to use to generate a response

Facilitate retrieval-augmented generation— RAG—where the generative AI tool can reference a trusted, verified, source of information to inform responses

Perform web searches and provide publicly available web links to referenced material in generated responses, including those from ASHA or the American Medical Association.

Because AI-generated content can be incorrect or misleading, it is your responsibility—as the user—to verify any cited references as well as evaluate and revise the content to ensure that it is accurate, it is appropriate, and it aligns with your own clinical knowledge.

Prompts

A prompt is the text entered into an AI tool to generate the information that one is seeking. The prompt includes context, instructions, questions, and relevant information that the AI uses to produce responses or output. Results could include text, images, or other media, depending on the platform you use.

Designing or engineering a sufficient and appropriate prompt to solicit useful information is key. The way a question is phrased significantly impacts the quality, tone, and accuracy of a generative AI tool’s response. Specific, clear, and context-rich prompts tend to lead to more relevant, useful outputs. It is best practice to ask AI to “explain its reasoning” or “cite sources” to assess the reliability of a response. However, you should always check citations for accuracy.

ChatGPT created this chart comparing two examples of prompts—one poorly designed and one improved—to illustrate the principles of effective prompt engineering:

Prompt Type	Prompt Example	Analysis
Poor Prompt (General)	"Make an email."	This prompt is overly general, lacks context, and doesn't specify the purpose, audience, tone, or details required to generate a useful response.
Improved Prompt (Detailed)	"Compose a professional email to the parents of my pediatric speech therapy client reminding them of their next session on Thursday at 2 PM. Mention bringing their AAC device and any documentation they have about the recent audiology appointment. Use a friendly, supportive tone and ask them to confirm attendance."	This prompt clearly defines the task, recipient, client details, purpose, timing, required materials, and desired tone, ensuring an accurate, comprehensive, and directly applicable response.