Archana Arakkal, Synthesis

When generative AI algorithms from existing data sets and imagery are used to train the next generation of GenAI models, they go MAD. As a recent study from the Department of Computational Applied Mathematics and Operations Research at Rice and Stanford Universities explained, a lack of fresh data in each successive generation of an autophagous loop means that “future generative models are doomed to have their precision and recall progressively decreased”. In short, they are infected with a condition known as Model Autophagy Disorder, or MAD – an acronym coined by the university because of the similarity to mad cow disease.

It takes about five rounds of LLM training itself before MAD sets in. The data lacking the enrichment of external inputs moves even more rapidly towards algorithmic bias and bigotry. And this is just within the research realm. AI has also been showing signs of incompetence, chaos and insanity within the applications that are in use today. In 2023, attorney Steven Schwartz used ChatGPT as part of his research in a court case against Avianca, a Columbian airline. It turned out that the AI had invented at least six of the cases and the lawyer ended up having to pay a hefty fine. Then, in a slightly more comedic twist, the AI robot providing support to customers of DPD, a courier service, started writing poems and swearing at customers: