Transparency and Information (Article 13)

Learning Objectives

By the end of this chapter, you will be able to:

Design AI systems with appropriate transparency for deployer interpretation
Create comprehensive instructions for use meeting Article 13 requirements
Communicate AI capabilities, limitations, and risks effectively
Calibrate transparency level to intended purpose and user competence
Integrate transparency with human oversight requirements

Article 13: Transparency Requirements

Intended Purpose

Clear description of what the AI does

Capabilities & Limitations

What it can and cannot do

Performance Metrics

Accuracy, error rates, known biases

Human Oversight Needs

Required level of human control

Known Risks

Identified risks and mitigations

Data Requirements

Training and input data specifications

Instructions for use must be comprehensive, understandable, and accessible to deployers

Article 13 establishes the transparency foundation for trustworthy AI. High-risk AI systems must be designed so deployers can understand, interpret, and appropriately use the system's outputs. Without transparency, human oversight becomes impossible.

The Transparency Imperative

Why Transparency Matters

Purpose	Benefit
Interpretability	Deployers understand what AI outputs mean
Appropriate use	Deployers use AI within intended parameters
Risk awareness	Deployers know limitations and failure modes
Human oversight	Enables meaningful human control
Accountability	Clear understanding of AI behaviour

Article 13 Requirements

Core Transparency Standard

High-risk AI systems must be designed and developed to ensure operation is sufficiently transparent to enable deployers to:

Interpret the system's output appropriately
Use it appropriately for its intended purpose

Transparency Calibration

The level of transparency must be appropriate to:

Factor	Consideration
Intended purpose	Higher stakes = more transparency needed
User competence	Technical vs. non-technical users
State of the art	What transparency is technically feasible
Deployment context	Operational environment constraints

Instructions for Use (Article 13(3))

Mandatory Content

Providers must supply deployers with instructions including:

Element	Required Information
(a) Provider identity and contact	Name, address, contact details; authorised representative if applicable
(b) Characteristics, capabilities and limitations	Includes: (i) intended purpose, (ii) accuracy/robustness metrics, (iii) foreseeable misuse and risk circumstances, (iv) known limitations, (v) group-specific performance, (vi) input specifications, (vii) interpretability of outputs
(c) Pre-determined changes	Modifications since initial market placement
(d) Human oversight measures	Technical capabilities and measures for human oversight
(e) Computational/hardware resources and expected lifetime	Required resources, maintenance and care instructions
(f) Log collection mechanisms	Description of logging capabilities per Article 12

Writing Effective Instructions

Best Practices:

Principle	Implementation
Audience-appropriate	Match user technical level
Actionable	Tell users what to DO
Complete	Cover all required elements
Clear	Avoid jargon where possible
Accessible	Easy to find and navigate
Updated	Reflect current system state

Common Instruction Failures

Failure	Problem
Generic templates	Don't reflect specific system
Missing limitations	Deployer doesn't know boundaries
Unclear metrics	Performance claims not verifiable
Buried warnings	Risk information not prominent
Outdated content	Doesn't match current system

Communicating AI Capabilities and Limitations

Performance Disclosure

Metric Type	How to Communicate
Accuracy	"System achieves X% accuracy on [specific task]"
Precision/Recall	"X% of positive predictions correct; Y% of actual positives detected"
Error rates	"False positive rate: X%; False negative rate: Y%"
Confidence ranges	"Outputs include confidence scores from 0-100%"

Limitation Disclosure

Clearly communicate:

Input limitations: What data types/quality the system can process
Population limitations: Groups where performance may differ
Environmental limitations: Conditions affecting performance
Temporal limitations: When data currency affects outputs
Failure modes: How the system fails and warning signs

Risk Communication

Risk Type	Disclosure Approach
Safety risks	Prominent warnings with specific scenarios
Fundamental rights risks	Clear statement of potential impacts
Bias risks	Groups that may be disproportionately affected
Misuse risks	Explicit "do not use for" statements

Expert Insight

Think of instructions like medical package inserts—comprehensive information for professional users who need to understand benefits, risks, contraindications, and proper use.

Transparency for Different User Types

Technical Users

Provide:

Detailed performance metrics
Model architecture information
Training data characteristics
Validation methodology
API specifications

Business Users