Coming to Terms with the AI Monoculture

The leading AI systems are few in number, because the cost of training one is so enormous that only a handful of companies can afford it. Moreover, their basic training data and architecture are shared. This is why we see is a handful of leading models that are extremely similar to one another. This similarity is not just a matter of all of them being equally good. The models make correlated errors. When two models get a question wrong, they tend to give the same wrong answer. They reveal similar vulnerabilities, so an attack devised against one often works on others. And their creative output often follow the same clichés. Ask for a story and you repeatedly get stories about a lighthouse keepers or a clockmakers. Why is this a worry? Outside mathematics and coding, there is not generally a single right answer for systems to converge on, so the convergence reflects shared biases rather than truth. Humans, flawed and limited as we are, are numerous, and we can have very different upbringings. The proposition that AI should replace human intellectual labour is therefore also a proposition for homogeneity: many different minds exchanged for a few similar ones. The reason for the monoculture, moreover, is deeply embedded in the technology: the data-hungry process of learning. These systems acquire new skills only via vast quantities of data, which limits the ability to differentiate models by their training.