This paper contrasts the approaches of deep learning, a type of machine learning in which neural nets find patterns in data, and generative AI, a derivative of deep learning in which large-scale models used to generate new data based on the patterns detectable in their training data. It is argued that they produce a different type of focus. Deep learning focuses on problems set for them by humans, connecting their output with reality in an analogue of the correspondence theory of truth. Generative AI focuses on producing output based on an exploration of its model, focusing therefore on the model itself rather than on the problem set by the query, in an analogue of the coherence theory of truth. The problem with coherence is that the model is detached from the world, while human input to correct this is minimised. Generative AI therefore risks inaccuracies such as bias, hallucinations, sycophancy, increasing the amount of noise relative to signal, misinformation, and internalising and side-stepping safety guardrails. The pursuit of artificial general intelligence (AGI) exacerbates these problems by increasing generative AIâs thirst for data, so the output of one generationâs systems is likely to be fed into the succeeding generations, polluting the data commons. We call this phenomenon pollucination, and show that it leads to a tragedy of the commons, where competition to reach AGI reduces the perceived importance of safety. We argue, however, that the competition to reach AGI is based on false premises and unfounded narratives, and that this tragedy of the commons is therefore not a collective action problem where the actors are governed by rational self-interest. The solution to the problem is not hard, and is to be found in virtue ethics and responsible AI, although we are pessimistic as to whether the state of the industry will enable it to be implemented.
O’Hara et al. (Thu,) studied this question.