Intended Use and Out-of-Scope Applications
What you'll learn: How to clearly define what a model should be used for—and what it absolutely shouldn't—within model documentation.
Why This Matters
Even well-documented models (as covered in your recent Model Card lessons) can be misused if their boundaries aren't crystal clear. Intended use specifies the contexts, populations, and tasks the model was designed for. Out-of-scope applications explicitly warn against deployments that could be harmful, unreliable, or unethical.
Core Concept
Think of this like medicine packaging: a drug label states what condition it treats, the approved dosage, and warnings about dangerous misuse. Similarly, your model documentation must state:
Intended Use:
- Primary tasks (e.g., "sentiment classification of English product reviews")
- Target populations and data characteristics (e.g., "text from e-commerce platforms, 2020-2023")
- Recommended deployment contexts (e.g., "internal quality monitoring, not automated customer-facing decisions")
Out-of-Scope Applications:
- Tasks the model wasn't trained for (e.g., "not for medical diagnosis or legal advice")
- Populations where performance degrades (e.g., "not tested on non-English text or specialized jargon")
- High-stakes scenarios requiring human oversight (e.g., "not for standalone hiring decisions")
- Ethical red lines (e.g., "not for surveillance or profiling vulnerable populations")
Why Both Matter
Without intended use, adopters may deploy your model in contexts where it performs poorly. Without out-of-scope warnings, they may cause real harm in high-stakes scenarios you never anticipated. This is especially critical under regulations like the EU AI Act, where misuse of high-risk systems carries legal consequences.
Key Takeaway: Explicitly defining intended use cases and out-of-scope applications prevents misuse, manages expectations, and fulfills both ethical and regulatory documentation requirements.