The Role of Culture in the Intelligence of AI
Mercedes Bunz · transcript Verlag eBooks · 2023
Artificial Intelligence (AI) clearly suffers from its name, which easily leads to misunderstandings.The name suggests that its intelligence is like human intelligence, only 'artificial'; it would have been far better and more precise to call it 'machine intelligence'.Neural networks, the technology that is in part the foundation of the current boom and facilitated deep learning approaches, adds to this, as it is a term that further confuses the discussion.The term suggests that machine learning systems are built on 'neurons' that operate just like those in the brain.If you ask experts in the field, however, they will quickly explain that biological neurons function very differently.They are much more complex than the mathematical functions we find in the nodes of a human neural network; this is the case internally (they seem to transmit signals that are chemical, but can also be electromagnetic or operate on ion channels), as well as externally with respect to their architecture.So why do we hold on to these misleading descriptions?Unfortunately, pretending that the technology is 'inspired' by human biology seemed to be an easy way to persuade everyone to believe that the technology will work at some point.Of course, it makes sense to a human that something inorganic is becoming intelligent when it imitates 'the master'.Only that machine intelligence functions very differently.Machine learning (ML) models are composed of multiple processing layers that analyse and learn representations of data with multiple levels of abstraction (LeCun/Benigo/ Hinton 2015).The model starts on the lowest layer by looking at, for example, pixel formations, while subsequent layers configure more complex features, that is, forms typical for the dataset from those pixel formations.This bottom-up approach starting with the smallest entity enables the models to discover 'intricate structures in large datasets' (LeCun/Benigo/Hinton 2015, 436).The intelligence we find here is therefore rather particular: the model learns stochastically by adding up very small elements to calculate 'a bigger picture' or (in the case of large language models) to calculate the meaning of a sentence from analysing the context of thousands of tokens (entities similar to words), taking note of which other tokens are in nearby vectors.This means that ML models operate on a very different level than human intelligence.They have very particular abilities,