Learn
Home Foundations Glossary Research
Do
Prompts Workflows Tasks
Adapt
Domains Settings Patterns
Verify
Antipatterns Case Studies Policies Resources

Grounding

Clinical Definition

Connecting AI output to specific source material rather than letting the model generate from its training data alone. Think of the difference between a student citing their textbook and a student confidently making something up. When output is grounded, you can trace claims back to specific provided text.

Technical Definition

A technique for constraining model output to information contained in provided reference documents, typically implemented through retrieval-augmented generation (RAG) or explicit citation requirements. Grounding reduces hallucination by anchoring generation to verifiable source material rather than parametric memory alone.

Also known as: grounded output, source attribution, anchored generation

Why SLPs Need to Know This

Ungrounded AI output is the default. When you ask a model to “write a treatment plan for a 4-year-old with CAS,” it generates text based on statistical patterns, not a specific reference. Grounded output ties the response to documents you provide: your assessment data, a specific protocol, or a peer-reviewed source. This is the difference between useful and dangerous.

Practical Guide

  1. Provide your source material explicitly. Paste in your assessment notes, the specific protocol, or the relevant guideline rather than asking the model to work from memory
  2. Ask for citations. Instruct the model to quote or reference specific sections of what you provided
  3. Verify the grounding. Check that the model’s output actually reflects your source material and hasn’t drifted into generated content
  4. Watch for blending. Models will mix grounded content with ungrounded filler, often seamlessly

The Clinical Analogy

Think of the difference between evidence-based practice and clinical folklore. A grounded response is the clinician who says “According to the Strand 2020 checklist, this client meets 8 of 10 criteria for CAS.” An ungrounded response is the clinician who says “In my experience, this is probably CAS,” except the model has no experience. It only has statistical patterns.

  • RAG (Retrieval-Augmented Generation): the primary technical method for grounding model output
  • Hallucination: what happens when output is not grounded
  • Evidence-Based Practice: the clinical framework that demands grounded reasoning

SLP/IO Assistant

Powered by Claude · No PHI accepted
AI assistant for clinical workflow support. Never enter student names, DOBs, or identifiable information.
Hi! I'm the SLP/IO assistant, an opinionated AI grounded in clinical practice. I can help with goal wording, note structure, ethical reflection, and navigating LLMs responsibly. What are you working on?