This glossary provides definitions and explanations for artificial intelligence (AI) terms used in the context of evaluating AI-enhanced features offered by vendors.

Accuracy / Accuracy Rates
A measure of how often the AI model makes a correct prediction overall. (e.g., % of emails correctly identified as spam or not spam).

AI Act (EU)
Upcoming European Union regulation intended to harmonize rules for Artificial Intelligence, classifying AI systems based on risk and setting requirements for development, deployment, and use.

AI Model
The specific algorithm or computational system, trained on data, that performs an AI task like prediction, classification, or generation.

API (Application Programming Interface)
A set of definitions and protocols that allows different software applications to communicate and exchange data with each other, crucial for integrating AI into existing systems.

Audit (Third-Party)
An independent assessment of a system, process, or compliance status conducted by an external, objective organization.

Bias (AI)
Systematic errors or unfair preferences in an AI system’s outputs, often originating from skewed training data or flawed model assumptions, leading to potentially discriminatory outcomes.

Bias Mitigation
Techniques and processes used during AI development and deployment to identify, measure, and reduce unfair bias in the system’s behavior.

Carbon Footprint
The total amount of greenhouse gases (primarily carbon dioxide) generated by actions, processes, or technologies; relevant here to the energy consumption of AI systems.

Compliance
Adhering to specific laws, regulations, standards, or contractual obligations (e.g., GDPR compliance, compliance with industry standards).

Data Ownership
Legal rights and control over data, determining who can access, use, modify, share, or delete data generated or processed by a system.

Data Protection Impact Assessment (DPIA)
A risk assessment process required under GDPR for data processing activities considered high risk to individuals’ privacy rights.

Data Subject Rights
Rights granted to individuals under data protection laws like GDPR, such as the right to access, correct, delete, or transfer their personal data.

Datasets
Organized collections of data used for training AI models (teaching them patterns) and testing their performance.

Due Diligence
The reasonable steps a person or organization should take to satisfy a legal requirement or standard of care, often used in the context of vetting vendors or investments.

Encryption
The process of converting data into a secure code to prevent unauthorized access. Only authorized users with the key can decode it.

Explainability / Interpretable Results
The extent to which the reasoning behind an AI model’s decision or prediction can be understood by humans. Sometimes referred to as XAI (Explainable AI).

F1 Score
A performance metric for classification models that balances Precision and Recall into a single score, useful when dealing with uneven class distributions.

Fairness (in AI)
An ethical principle ensuring that AI systems do not produce discriminatory or unjust outcomes for individuals or groups based on sensitive attributes like race, gender, etc.

Future Roadmap
A company’s plan outlining future development, features, and strategic direction for a product or service.

GDPR (General Data Protection Regulation)
The primary data protection and privacy regulation in the European Union, setting strict rules for handling personal data.

HIPAA (Health Insurance Portability and Accountability Act)
A US federal law setting standards for protecting sensitive patient health information.

Indemnification
 A contractual clause where one party agrees to cover the losses or damages incurred by another party under specific circumstances (e.g., if sued due to the first party’s product failure).

Integration
The process of combining different software systems or components (like AI features) so they work together seamlessly within existing workflows.

Liability
Legal responsibility for damages or harm caused to others.

Licensing
The legal agreement that grants permission to use software or technology under specific terms and conditions (cost, duration, usage limits, etc.).

Performance Metrics
Specific, quantifiable measures used to evaluate how well an AI system performs its intended task (e.g., accuracy, precision, recall, F1 score, response time).

Personal Data
Under GDPR, any information relating to an identified or identifiable natural person.

Precision
A performance metric indicating the proportion of positive identifications made by the AI that were actually correct. (e.g., Of all emails flagged as spam, what percentage actually were spam?). Helps measure false positives.

PSD2 (Payment Services Directive 2)
An EU directive regulating payment services and providers, relevant for AI used in financial technology.

Recall
A performance metric indicating the proportion of actual positive instances that the AI correctly identified. (e.g., Of all the actual spam emails received, what percentage did the filter correctly flag?). Helps measure false negatives.

Risk Management
The process of identifying, assessing, and controlling threats or potential negative impacts to an organization’s capital and earnings (here, specifically related to AI adoption).

Sandbox Environment
An isolated, controlled computing environment used for testing software or running programs without affecting the live operational system.

Scalability
The ability of a system or application to handle a growing amount of work or users efficiently, or its potential to be enlarged to accommodate that growth.

Sensitive Data
Personal data requiring special protection due to its potential for misuse (e.g., health information, racial origin, religious beliefs). Often subject to stricter rules under laws like GDPR.

Transparency (in AI)
The principle that sufficient information about an AI system (its purpose, data, logic, performance) should be made available to allow for understanding, scrutiny, and trust.

Validation
The process of confirming that a system or model meets specified requirements and performs accurately for its intended use, often through testing.