The Dark Consequences of Artificial Intelligence Data Self-Cannibalism

2023-07-19

Finally, the AI model that generates content will begin to train on the data it generates, leading to data self-cannibalism. Artificial intelligence is devouring itself. The internet has now become a landfill for artificial intelligence, with models trained on the internet feeding on each other. This is known as data self-cannibalism. In an article for The New Yorker, renowned science fiction writer Ted Chiang warns of the dangers of AI copies reproducing copies in various digital duplications. He compares this rapidly evolving dilemma to the JPEG effect, where the quality of each copy deteriorates, revealing ugly artificial traces resembling mosaics. As the boundaries of AI replication become blurred, it is worth considering what happens when AI-generated content spreads on the internet and AI models start training on them instead of primarily on human-generated content. Researchers from the UK and Canada have recently discovered a phenomenon called "model collapse" in generative AI models. This degradation occurs when the model learns from data generated by other models, causing it to gradually lose an accurate representation of the real data distribution. Importantly, it is considered inevitable even under near-ideal learning conditions. Ross Anderson, co-author of the research paper on "model collapse" and a professor of security engineering at the University of Cambridge, suggests that the internet is at risk of being flooded with irrelevant content, much like plastic waste in the ocean. This content deluge could hinder the training of new AI models through web scraping, benefiting companies that have accumulated data or control large-scale human-machine interfaces. According to a recent report by media research organization NewsGuard, a concerning trend has emerged: websites filled with low-quality content generated by AI to attract advertisers. The report reveals that over 140 well-known brands unknowingly paid for ads displayed on websites with AI-written content. This growing influx of low-quality AI-generated content poses a threat to AI companies responsible for these models. As training datasets become increasingly saturated with AI-generated content, concerns arise about the diminishing utility of language models. Ilia Shumailov, a researcher from the University of Oxford's Applied and Theoretical Machine Learning Group, warns that this trend could have more serious implications, such as discrimination based on gender, race, or other sensitive attributes. Due to their lack of inclusivity, current models have been blacklisted by AI ethicists. In 2021, a group of researchers raised concerns about biases in language models towards white males. Anders Søgaard, the first author of the study and a professor in the Department of Computer Science at UCPH, explained that these models exhibit systemic biases. Surprisingly, they align most closely with the language used by white males under 40 with lower levels of education, while being least aligned with the language used by young non-white males. This finding underscores the urgent need to address and correct biases in language models to ensure fairness and inclusivity for all. Similarly, Shumailov states, "To prevent model collapse, we need to ensure fair representation of minority groups in future datasets." While some companies are striving to develop more inclusive AI, such as Meta's recently released open-source dataset called "Casual Conversations v2," which aims to serve a wide range of use cases, on the other hand, since the dismissal of renowned AI ethicist Timnit Gebru, followed by the departure of other members of the Google team, Google has been labeled a "white tech organization." Despite their many flaws, language models also have some advantages. For example, Shumailov and his team initially named it "model collapse," the dementia effect model, but decided to rename it after opposition from a colleague. He wrote, "We couldn't come up with an alternative name until we asked Bard, who provided five names, and we chose 'The Curse of Recursion.'" Currently, language models are becoming part of every company's strategy. Companies across industries are learning how to unleash the full potential of advanced chatbots based on language models like GPT-4. While it is still too early to judge, companies are adapting and finding the best use cases for themselves amidst the chaos of generative AI. Source: https://analyticsindiamag.com/the-dark-consequence-of-ais-data-cannibalism/

RECENT AI NEWS

RECENT AI TOOLS