How to define artificial intelligence is a difficult problem because scientists have never had agreement about the definition. As a result it is difficult to assess whether particular technology is artificial intelligence technology or not. The implications are arguably far reaching in terms of the law and regulation because it is difficult to specify what technology is subject to AI law and AI regulation without defining AI. Shortly before Christmas the European Commission high level expert group on AI published a 7 page document with the following definition of AI and an explanation of the definition. There is also a disclaimer saying that the definition is an oversimplification. The definition is interesting for a number of reasons.
First, there is no requirement for the technology to learn, just a statement that "AI systems can also be designed to learn". So presumably that means that technology such as constraint satisfaction solvers (which use heuristics to search vast seach spaces but which do not learn), and traditional expert systems technology (which uses collections of rules and rule based reasoning but no learning), is potentially within the definition if it meets the other criteria of the definition. Second there is a requirement for "perceiving in their environment" where the environment can be "the physical world or digital world".
When I think of the term "perception" it makes me think of the task of interpreting raw sensor data about the environment such as sounds, tastes, smells, light and pressure/touch. In the digital world it makes me think of passively receiving event data such as click logs of a search engine, social media messages, and other digital world events and then interpreting the data. In the case of a chat bot there is perception of the physical world since sound is input and interpreted as speech. In the case of a traditional expert system, connected to a speech recognition system, there would arguably be perception of the physical world. In the case of a constraint satisfaction system receiving sofa delivery requests from a call centre agent computer in order to compute shortest path delivery routes, will the sofa delivery requests count as incoming data which is "interpreted" to give "perception" of digital world events? There are arguments both for and against.
The incoming delivery requests are events in the digital world and they are interpreted and pre-processed by the constraint satisfaction system as being data to include in the shortest path calculation.
However, I am not sure that really counts as perception?
Third, there is a requirement for "reasoning on the knowledge" and "deciding the best actions to take". The reasoning has to be rational as explained in the 7 page document, since the definition says "deciding the best action(s) to take". The word "best" could be tricky here since there is always debate as to what "best" is even when there is a defined goal. Also, it is likely that the goal may change over time. When a neural network makes a decision it is often a good working decision in practice, but may not always be the best, and sometimes can be erroneous. In the case of a chat bot, it arguably reasons and decides a course of action to take since it decides what reply to give to the user. In the case of the constrain satisfaction system it carries out a search of a huge search space (a form of reasoning) and it computes an answer which is a selection from a huge number of choices (i.e. a decision).
"Artificial intelligence (AI) refers to systems designed by humans that, given a complex goal, act in the physical or digital world by perceiving their environment, interpreting the collected structured or unstructured data, reasoning on the knowledge derived from this data and deciding the best action(s) to take (according to pre-defined parameters) to achieve the given goal. AI systems can also be designed to learn to adapt their behaviour by analysing how the environment is affected by their previous actions.
As a scientific discipline, AI includes several approaches and techniques, such as machine learning (of which deep learning and reinforcement learning are specific examples), machine reasoning (which includes planning, scheduling, knowledge representation and reasoning, search, and optimization), and robotics (which includes control, perception, sensors and actuators, as well as the integration of all other techniques into cyber-physical systems).
The content above was originally posted on CMS DigitalBytes - CMS lawyers sharing comment and commentary on all things tech.