Artificial intelligence gathers the closest answers and determines a reliable report  by quickly understanding ambiguities and questions. But the passage of time and the review of the results have revealed new realities.

Perhaps in the early days of using artificial intelligence, we were delighted with the answers it received. How pleasant and desirable it would be telling us which of the watermelons in the in the fruit table is sweet and juicy, without saying “please” at the beginning of question, or to want to learn about Jamasp’s opinions in the history of philosophy and compare it with the Plato’s.

A reference that understands our ambiguities and questions and gives a suitable answer in the shortest possible time with just an imperceptible prompt. With the support of these pleasant experiences, some people have imagined the future in the hands of artificial intelligence and have imagined a redundant role for humans and their natural intelligence.

Not considered aspects of AI-generated answers

But let’s look a little further. The passage of time and a focus on some of the newer results have revealed new realities: it seems that AI answers become less attractive the more they are used, polished, and published on the Internet.

Let’s take a look at one of the most important AI algorithms:
AI collects the closest answers by searching data centers and Internet pages. Something similar to Google’s special selection, which is of course also subject to the supervision of some content producers and copyright laws.

The early AI answers were mainly tips written by humans and had a soul. But the second and third generation answers no longer had a human behind them. And it can be guessed that the previous answers of artificial intelligence will continue to be published, republished and reproduced.

Answers that have been simplified before and the human aspects of writing and the gentleness of writing in them are becoming less and less. Just like watery series! Lacking enough coincidence and charm in the scenario.

Many of the creative aspects in the answers that are likely to be more doubtful and that are the potential source of new opinions and perspectives are no longer seen. Because they are less certain and have less weight in comparing sources. These types of answers are left out of the authority of artificial intelligence and are marginalized. Artificial intelligence does not accept the risk of these types of answers and leaves them aside. To the extent that one can no longer enjoy and enjoy the answers of artificial intelligence.

Research results on artificial intelligence published in the Nature

This experience has not gone unnoticed by researchers from the Universities of Oxford and Cambridge. In an article in the Nature, they showed that as generative AI changes, like large language models, model collapse is occurring, which will ultimately change the online text and image ecosystem.

The findings of this research show that the unconditional use of model-generated content in learning causes irreversible defects in the resulting models in which the original content distribution sequences disappear. That is, supplementary explanations and criticisms are ignored by the artificial intelligence, and deficiencies and limitations are ignored.

IBM also explained this collapse as follows:
Model collapse refers to the decline in the performance of generative AI models that are trained on content generated by artificial intelligence.

The Technical and Software Aspects of Machine Learning That Reveal Shortcomings

Let’s see how this process works from a technical perspective:
It is common for AI companies to crawl (or scrape) the Internet to train the next generation of models. In this way, a web crawler, or “spider,” like a person with a lot of time, roams the Internet and observes various links to index and search for content by following links and exploring.

This means that they first “crawl” the web environment or a specific website to find the addresses of Internet pages (URLs). They then transfer them to their scraper to extract the desired data from them.

By repeating this activity, which is inherently very conservative, what remains is the average or something that is approved by the majority and has provoked less criticism and opposition. Very safe, expected and tasteless content.

In this way, subsequent generations are educated on the same content devoid of creativity and criticism, and in return they lose their opportunities to pay attention to criticism, contradictory evidence and side issues, and the next generation will follow the same path more quickly.

Perhaps one of the priorities of AI designers should be to pay attention to newer human findings, less-considered aspects and critical perspectives on various topics. At the same time, users’ attention to these shortcomings of AI can keep them away from the harm and risk of not paying attention to more complex aspects, violations of common beliefs and less-known ones.

RRC HRM Consultancy in Iran