Murray Webster

One of the things that we at Fractal AI are most concerned with is public education about artificial intelligence, the risks and rewards associated with its use and the effects it has on our social, political and economic lives. I recently co-authored a paper for our journal on early-childhood AI education and the failures of the UK Government to adapt quickly to emerging trends.

We are, therefore, always keeping an eye out for situations in which the public’s received wisdom about AI doesn’t match what is actually occurring. We take the position that the general public is generally capable of understanding difficult ideas, which means that a failure of understanding is a failure of education.

Looking back on the way in which AI was talked about a couple of years ago, it’s interesting to see with hindsight just how concerned – or, depending on your disposition, delighted – many were at the potential for the collapse in the capabilities of AI models.

The potential cause of such a collapse, it was widely thought, could be the result of these models being recursively trained on the data that they had output. As the theory went, models were capable of producing so much more text, so much more quickly, than people that the internet, the ground from which AI businesses scrape data to train their models, would eventually become saturated with AI-generated data. 

In this scenario, such data would come to constitute, to within a rounding error, all of the internet. AI models would have no choice but to train on data they themselves had produced. Many potential metaphors are possible here, but feeding oneself on one’s own output hardly appeals from a nutritional standpoint, steadfastly disregarding other important factors.

Much like photocopying a document over and over again, or repeatedly taking screenshots of an image, the quality of data generated by models, even those sitting at the frontier, would degrade over time. This was by no means a fringe theory: it was discussed in Nature, along with many other prestigious publications. The mathematics of the situations was laid out in detail and verified.

We are, therefore, required to ask: what happened? Or, more appropriately, what didn’t? No matter how well supported these theories were, they haven’t come to pass. There are a number of overlapping explanations from a technical perspective, but I want us to consider science communication for a moment. Many of the supposed prophecies put forward by AI researchers about model collapse were anything but; they were conditional statements predicated upon specific events occurring. And yet, that is not the picture that the public received. What was a nuanced and careful thesis was mushed into a bland paste, all to make it easier for us to consume.

AI labs took the possibility of model collapse seriously and worked hard to mitigate it. Their work here has produced enormous dividends. Why, then, does the myth of its inevitability still persist? There is a gap between what labs produce, what academics write and what the public understands. While such a problem of science communication is hardly unique to AI, few fields are developing quite as fast or having quite such an impact. The scale and speed of AI development behoves us to argue for a better standard of AI education for the general public.

At Fractal AI, we are already working with a network of universities and the Department for Education to make this a reality. We want to make sure that early-career researchers can be mentored by their more experienced seniors, and educate their peers in turn. The development of AI and its potential benefits are too great to allow the public to be misled about what it can do. If you’d like to join is, we’d be delighted to hear from you.

Interested in starting your own professional registration journey?

Find out more about our registration accreditation framework and the training we offer.