This is the definition that the High-Level Expert Group on AI (AI HLEG) proposes for the purpose of its deliverables. It accompanies the draft AI Ethics Guidelines wich are now subject to a Stakeholder Consultation.
I think there are (at least) two competing views on what AI is:
1. Marketing view (both in science and business), which wants to have AI label on nearly each product, which is legitimate. My understanding is that the guidelines do not have ambition to regulate usage of AI term.
2. The ethics guidelines view, which should focus on the technologies and approahces that bear significant or new risks compared to traditional computer systems. Under traditional coumputer system I understand a system, which is designed, developed and tested by well undestood techniques and approaches to provide
- reasonable predictability of its bahaviour,
- defined reliability in its operation
- etc.
so the developers, owners and users can UNDERSTAND and share responsibility for its actions.
We probably want to keep the guidelines acceptable for business and limit regulation (and business promises) to the risky areas, so I propose to focus the guidelines mainly on those AI techniques and approaches, which are risky (in line with the bold goal in the executive summary) and which are not "traditional" in this sense. This might help with some disputes around the guidelines.
Two more examples, which fall within the proposed definiton of AI, but might not be considered AI applications:
1. PID controller (e.g. a thermostat) is sensing environment (temperature), makes decisions, even considering the past responses, which mimics learning and memory, and acts by affecting an actuator (heating system) to achieve specific goal (preset temperature).
2. Optical Character Recognition - there are opinions that OCR is not considcered AI anymore
You might need to pull something on Big Data into this. Data Mining is the process of getting software to analyse a dataset to pull correlations out of the data that can then be used to program AIs to utilise them to predict future data (for example, training a neural net AI).
In the reasoning and decision making it is possible, using an object based knowledge structure, to construct a network of AI cells that distributes knowledge throughout the network, so that reasoning decisions become simple zen like reactions. This is a somewhat different technique from the 'reasoning' approach described here. It lends itself to make emergent networks of AI nodes where nodes can easily be added and removed and the localised reasoning cost is, more or less, independent of the total number of reasoning nodes.
I think chapter 2 needs to distinguish between trained AI and AI that actively learns as it processes. There are some significant differences in the behaviour and vulnerabilities of both types. Training is a way of discovery a working algorithm, which is quite different from Ai that actually learns by expanding (or pruning) it's knowledge structure as it runs.
There is also AI that can be taught as it runs, buy feeding it an updated knowledge structure or an edit to its existing knowledge structure while it is running. Again, this is a diferent and distinct process from AI that learns as it runs.
Goal directed AI can accept multiple, sometimes conflicting goals from multiple sources, and use various techniques to identify the most important ones for each distinct subset of entities that it manages. This can mean that understanding the AIs current goals for a particular subset of elements isn't straight forwards. The AI can also change it's localizede goals in response to input events - for example selecting a backup plan if the event indicates the primary plan is no longer achievable.
For chapter 1, large and complex AI applications can exhibit chaotic behaviour, especially if they are built using an emergent framework. Given enough data you can understand exactly why it took any particular action, but, unless you know all the relevant state data you can't predict how it will react to any given piece of input. The whole process is entirely determanistic, but I'm not sure 'rational' is how many of the users would describe it.
Because of blurry border between AI and non-AI, it might be helpful to include in the definition a section defining, what is not considered AI. E.g. is usage of trained neural network (with fixed coefficients) an artificial intelligence, or is it an algorithm created by an AI approach?
Yes. Expert systems, which are somewhat out of fashion, especially blur the line between simple linear programs and AI. Classifiers can even be implemented as linear programs and a lot of Big Data driven AI (so-called machine learning) is done with statistical analysis and simple linear forecasting.
Hi Anders,
As the purpose of this document is to define "AI" merely for the purpose of the AI HLEG's deliverables, no consultation similar to the one on the draft Ethics Guidelines is foreseen. Feedback is, however, always useful, so we would certainly wecome your suggestions for improvement - please feel free to post these here below - and we will ensure that the AI HLEG drafters of the definition will take this into consideration.
Is it possible to provide feedback about this definition of AI, similar to the stakeholder consultation form for the ethics guidelines document? Great that we try to find a common definition of AI, but believe some formulations in suggest definition could be improved.
