AI: Straight Towards a Model Collapse?

The phenomenon of model collapse in the field of AI refers to a mechanism whereby AI learning models lose effectiveness because they rely on data that is itself generated by AI (synthetic data) and is riddled with errors. The risk is that models will go in circles with this synthetic and false data, subsequently producing results that are also false, but which feed into the data available on the internet and from which the models will continue to learn in the future. Various mechanisms may be at play in this vicious cycle, but in any case, without human detection, there is no way to break the catastrophic chain of events leading to this model collapse.

Artificial intelligence experts are divided on the likelihood of these models failing. One thing is certain: the amount of data and information produced by AI and circulating on the internet continues to grow with the adoption of generative AI technologies, thereby increasing the proportion of this data that could be used to train models. How are AI companies addressing this risk? How can we guard against it?