Synthetic Data Is a Dangerous Teacher

In April 2022, when Dall-E, a text-to-image visio-linguistic model, was released, it purportedly attracted over a million users within the first three months. This was followed by ChatGPT, in January 2023, which apparently reached 100 million monthly active users just two months after launch. Both mark notable moments in the development of generative AI, which in turn has brought forth an explosion of AI-generated content into the web. The bad news is that, in 2024, this means we will also see an explosion of fabricated, nonsensical information, mis- and disinformation, and the exacerbation of social negative stereotypes encoded in these AI models.

The AI revolution wasn’t spurred by any recent theoretical breakthrough—indeed, most of the foundational work underlying artificial neural networks has been around for decades—but by the “availability” of massive data sets. Ideally, an AI model captures a given phenomena—be it human language, cognition, or the visual world—in a way that is representative of the real phenomena as closely as possible.

For example, for a large language model (LLM) to generate humanlike text, it is important the model is fed huge volumes of data that somehow represents human language, interaction, and communication. The belief is that the larger the data set, the better it captures human affairs, in all their inherent beauty, ugliness, and even cruelty. We are in an era that is marked by an obsession to scale up models, data sets, and GPUs. Current LLMs, for instance, have now entered an era of trillion-parameter machine-learning models, which means that they require billion-sized data sets. Where can we find it? On the web.

This web-sourced data is assumed to capture “ground truth” for human communication and interaction, a proxy from which language can be modeled on. Although various researchers have now shown that online data sets are often of poor quality, tend to exacerbate negative stereotypes, and contain problematic content such as racial slurs and hateful speech, often towards marginalized groups, this hasn’t stopped the big AI companies from using such data in the race to scale up.

With generative AI, this problem is about to get a lot worse. Rather than representing the social world from input data in an objective way, these models encode and amplify social stereotypes. Indeed, recent work shows that generative models encode and reproduce racist and discriminatory attitudes toward historically marginalized identities, cultures, and languages.

It is difficult, if not impossible—even with state-of-the-art detection tools—to know for sure how much text, image, audio, and video data is being generated currently and at what pace. Stanford University researchers Hans Hanley and Zakir Durumeric estimate a 68 percent increase in the number of synthetic articles posted to Reddit and a 131 percent increase in misinformation news articles between January 1, 2022, and March 31, 2023. Boomy, an online music generator company, claims to have generated 14.5 million songs (or 14 percent of recorded music) so far. In 2021, Nvidia predicted that, by 2030, there will be more synthetic data than real data in AI models. One thing is for sure: The web is being deluged by synthetically generated data.

The worrying thing is that these vast quantities of generative AI outputs will, in turn, be used as training material for future generative AI models. As a result, in 2024, a very significant part of the training material for generative models will be synthetic data produced from generative models. Soon, we will be trapped in a recursive loop where we will be training AI models using only synthetic data produced by AI models. Most of this will be contaminated with stereotypes that will continue to amplify historical and societal inequities. Unfortunately, this will also be the data that we will use to train generative models applied to high-stake sectors including medicine, therapy, education, and law. We have yet to grapple with the disastrous consequences of this. By 2024, the generative AI explosion of content that we find so fascinating now will instead become a massive toxic dump that will come back to bite us.