The European Commission recently issued a formula for identifying Artificial Intelligence Systems:

Machine-based system

Designed to operate with varying levels of autonomy

• Some degree of independence of actions from human involvement
• Some inference capacity
• BUT broad: Includes a system that requires manually provided inputs to generate an output by itself

That may exhibit adaptiveness after deployment

• Self-learning capabilities, allowing the behavior of the system to change while in use
• BUT only “may”

And that, for explicit or implicit objectives

• Implicit objectives are deduced from the behavior or underlying assumptions of the system
• Objectives may be different from the intended purpose (Purpose is externally oriented; objectives are internal).

Infers, from the input it receives, how to generate outputs

• Includes:
  • Deriving outputs through AI techniques enabling inferencing e.g.: machine learning approaches, and logic- and knowledge-based approaches
  • Machine learning approaches
  • Supervised learning: Learn from labeled data e.g email spam detection
  • Unsupervised learning: Learn from unlabeled labeled data. e.g. AI systems used for drug discovery by pharmaceutical companies
  • Self-supervised learning: learn from unlabeled data in a supervised fashion, using the data itself to create labels. e.g. learn to predict the next token in a sentence
  • Reinforcement learning: Learn from data collected from own experience through a ‘reward’ function. e.g. personalised content recommendations in search engines
  • Deep learning: Utilize layered architectures (neural networks) for representation learning.
• Logic- and knowledge-based approaches: E.g. early generation expert systems intended for medical diagnosis
• Excludes:
  • Automatically execute based on rules defined solely by natural persons
  • E.g. satellite telecommunication system to optimize bandwidth allocation and resource management
  • Basic data processing
  • Systems based on classical heuristics: E.g. a chess program assessing board positions
  • Simple prediction systems: E.g. using the average temperature of last week for predicting tomorrow’s temperature

Such as predictions, content, recommendations, or decisions

• Predictions: E.g.: AI systems deployed in self-driving cars are designed to make real-time predictions in an extremely complex and dynamic environment
• Content: E.g. text, images, videos, music and other forms of output.
• Recommendations: E.g. candidate to hire in a recruitment system
• Decisions: Conclusions or choices made by a system

That can influence physical or virtual environments

• The influence of an AI system may be both to tangible, physical objects (e.g. robot arm) and to virtual environments, including digital spaces, data flows, and software ecosystems.